Google DeepMind’s Co-Scientist, formalised in a Nature paper in May 2026, compresses the early stage of scientific discovery, generating and refining a research hypothesis, from weeks to days. It works as a multi-agent system built on Gemini: separate AI agents continuously generate hypotheses, critique each other’s reasoning, and refine what survives, in a loop that used to depend on a human researcher’s slower cycle of reading, thinking and drafting. It isn’t a demo. Now rolling out experimentally to individual researchers through Gemini for Science, it has already produced hypotheses that survived contact with a wet lab: a novel drug-repurposing candidate for acute myeloid leukaemia, and a new epigenetic target for liver fibrosis, both independently confirmed rather than merely plausible-sounding. This is a real capability, working, now.

In the same literature this capability is meant to accelerate, something else is happening. Fabricated references in biomedical papers have risen twelvefold in two years. Maxim Topaz, an associate professor at Columbia University’s School of Nursing and Data Science Institute, led an audit of 2.5 million papers — published as a letter in The Lancet in May 2026 — that found roughly one in every 277 PubMed-indexed papers now contains some fabricated references, up from one in 2,828 three years earlier. Review articles, the papers other researchers lean on most heavily to orient themselves in a field, showed a fabrication rate 57% higher than other paper types. It’s already sitting in the permanent scientific record, not a future risk to be managed.

One documented case, precisely

Most of that twelvefold rise is inference from pattern, not confession. But at least one instance is on the record in detail. A 2026 retraction from the Journal of Medical Ethics was explicitly attributed to an author who had used generative AI to identify and understand referenced sources, without independently verifying them before submission. The citations looked right. Nobody checked whether they were.

That’s the mechanism in miniature: a tool that can write a fluent, well-formatted reference list faster than any human, attached to a verification step that still assumes a human wrote it carefully. Multiply that gap across an entire literature moving at AI speed, and the twelvefold rise stops being surprising. The higher fabrication rate in review articles matters more than the raw number suggests, because a review article is exactly the paper other researchers reach for first when they’re trying to orient themselves in an unfamiliar field — it’s the summary everyone trusts precisely because checking every citation in it individually is what the review was supposed to save them from doing.

The system nobody rebuilt

Peer review and citation-checking were designed for a slower, lower-volume era of publishing — one submission at a time, one careful human reader at a time. AI can now generate a polished hypothesis, a literature review, and a citation list faster than that process was ever built to check — and the researchers behind the audit say plainly that the checking side hasn’t caught up to the volume the generating side now produces. There is still no accepted, mandatory standard for flagging an AI-assisted finding as provisional before it enters the permanent scientific record.

My Opinion

I don’t read this as a publishing-etiquette problem, the kind a stricter house style or a sterner editor’s note could fix. Verification is infrastructure, in exactly the same sense that the wet lab confirming Co-Scientist’s drug-repurposing candidate is infrastructure. Nobody would fund a hypothesis-generation system this fast and then leave the checking to whoever happens to notice. Yet that’s precisely what’s happening on the citation side, because the money and the attention are on the generation half of the pipeline, not the verification half. Until that changes, the speed is the risk, not the safeguard.

That gap has consequences beyond the literature itself. A funding committee weighing a grant application, a systematic review synthesising a field, or a clinical trial recruiter screening candidates has no reliable way to know whether the citation or hypothesis in front of them survived independent checking, like Co-Scientist’s confirmed drug-repurposing candidate, or only survived generation, like the reference list behind the Journal of Medical Ethics retraction.

Even Google is candid about this. The company acknowledges Co-Scientist’s own limitations in literature review and factual accuracy — meaning the most rigorously validated AI-assisted discovery tool currently available still depends on external checking that the surrounding system isn’t built to provide at the speed the tool now works at. The researchers behind the audit have recommended that citation-checking move earlier, built into the submission workflow before peer review begins rather than caught, if at all, after publication — a change that depends on publishers adopting it, which none has yet done at scale.

On the page, a hypothesis nobody has independently checked looks identical to one that’s been rigorously verified. Right now, nothing in the publishing system forces that distinction to be visible before the finding starts shaping what gets funded and which patients get recruited into which trials. How many findings currently doing that work have never actually been checked — and would anyone know?

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

For most people with idiopathic pulmonary fibrosis, the question is not whether their lungs will fail but how quickly. There is no cure. The available treatments slow the damage; they do not stop it. Median survival after diagnosis is three to five years.

In June, at the BIO International Convention in San Diego, Insilico Medicine announced a collaboration worth up to $2.5 billion in milestone payments with SK Biopharmaceuticals for neuroimmune disorders — a year after publishing Phase IIa trial results for its lead IPF drug in Nature Medicine. The molecule at the centre of this is Rentosertib, a TNIK inhibitor designed to slow the fibrotic damage that kills IPF patients. It is the first drug candidate where both the target and the compound were identified entirely by AI. The platform synthesised and tested 79 candidate molecules; Rentosertib was the 55th. From target identification to clinical candidate: 18 months.

The question is what to do with that.

What 18 months actually means

Traditional drug discovery takes years before a compound reaches a trial. Identifying the right biological target is itself a years-long process — hypothesis, experiment, elimination. Then designing a compound that binds to it, screening variants, testing in the laboratory, a Phase I trial to establish safety. By the time a drug enters efficacy testing, four to six years have typically passed.

Insilico collapsed that phase. The AI generated and assessed virtual molecules computationally, identified the best candidates, and selected one for synthesis without the iterative cycle that makes traditional chemistry slow. The biology remained; the chemistry accelerated.

Rentosertib’s Phase IIa trial enrolled 71 patients across three dose arms, randomised, double-blind, placebo-controlled. Patients on the highest dose showed a mean improvement in lung function of 98.4 millilitres over the study period; the placebo group declined by 20.3 millilitres. For a disease measured in how fast function falls, that gap matters. The results are not definitive; no Phase IIa trial is. But they were published in Nature Medicine in June 2025, and they were positive.

Discovery has changed. The question is whether the rest of the process noticed.

The infrastructure was not built for this

After Phase IIa comes Phase III. Larger trials, typically thousands of patients, with full efficacy and safety data required for regulatory submission. Insilico opened Rentosertib’s own Phase III trial in July, enrolling 320 patients across 47 sites in China. The median time from Phase IIa completion to approval still runs several years, even with expedited review designations. Fast Track status accelerates the conversation with regulators, not the data requirement.

The regulatory requirement is appropriate: evidence from 71 patients is not enough to approve a drug for general use. Phase IIa is promise; Phase III is proof.

But the clinical trial system was not designed around a world in which a promising molecule can be found in 18 months. Recruiting patients, running trials across multiple sites, generating data, submitting for review: the queue was sized for a pipeline that moved at human speed. AI has shortened the front of that process without changing anything that follows it.

For someone diagnosed with IPF today, the arrival of AI drug discovery does not change their timeline. Rentosertib may eventually be approved, or it may not; Phase III trials fail more often than they succeed. The drug was discovered faster, but the path from discovery to patient has not changed. The bottleneck moved from the laboratory to the clinic.

What gets built next

The collaboration announced in June is, among other things, an industry bet that AI will produce candidates at a scale and pace the traditional pipeline never achieved. If that bet is correct — and the discovery-side evidence is accumulating — then the clinical trial system will face pressure it has not faced before: not a shortage of candidates but a surplus of them, each requiring the same years of Phase III data.

Some adaptation is already under way. Adaptive trial designs, accelerated patient matching, synthetic control arms built on existing patient data, earlier integration of real-world evidence from approved therapies — these approaches exist and are being explored. None of them is available to the patient who needs a drug now, and none changes the gap between AI-accelerated discovery and human-speed delivery.

The FDA and the EMA published joint guiding principles for AI in drug development in January, and the MHRA has spent the same period expanding its regulatory sandbox for AI tools. None of it touches the actual constraint. The guidance governs how AI models get validated for use inside a regulatory submission, not how many Phase III trials the system can run at once, or how fast a surplus of candidates gets through it.

My opinion

I think this is the pattern, not the exception. We keep running into the same asymmetry: what a technology like AI can now do at pace and at scale, set against governance and legal procedures that were never built to move at that speed. Some of that caution is earned — Phase III exists because it should be hard to approve a drug on the strength of 71 patients. But most of the controls now standing in the way were built for a more analogue world, and they have not moved into this one. Until they do, every acceleration in discovery just relocates the wait.

For IPF specifically, the constraint is concrete. Median survival after diagnosis is three to five years. The drug that might help was discovered in 18 months. The distance between those two facts is not a scientific problem. It is a system problem — and the system in question is not the one that found the molecule.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Google pays Reddit roughly $60 million a year for access to its forums. OpenAI pays Reddit around $70 million for the same thing, and separately signed a five-year, $250 million deal with News Corp. None of the people who wrote the posts, comments and articles now changing hands for that money were party to any of these deals. They wrote for free, on platforms that were free to join, and the platforms sold what they wrote.

That should be the fact that stops you, before any argument gets built on top of it. Live-access licensing deals like these have grown from two in 2023 to a projected thirty-four in 2026 — a market that didn’t exist three years ago, already moving hundreds of millions of dollars a year, for something the AI industry used to scrape for nothing.

Why the market suddenly needs “real” data

The reason is a 2024 finding, published in Nature, that AI models degrade when they’re trained on their own output — not goodwill toward journalism or forum culture. Ilia Shumailov, a computer scientist then at Oxford, and his co-authors showed that models fed a diet of AI-generated text suffer what they called model collapse: an early drift away from the true distribution of language, followed by the permanent loss of rare, low-frequency information — the unusual phrasing, the minority view, the edge case that only shows up once in a million real sentences. Train a model on text that was itself written by a model, repeatedly, and each generation grows narrower and worse.

The supply of text that hasn’t been through that loop is shrinking fast. By April 2025, an estimated 74.2% of new webpages already contained AI-generated text. Epoch AI has projected that the internet’s stock of fresh, verifiably human-written material could be exhausted sometime between 2026 and 2032. Put those two facts together and the licensing boom stops looking like opportunism and starts looking like the market pricing in a resource it only just discovered was finite.

What buying real data gets you

For a lab that can afford it, a licensing deal buys something specific: a continuously refreshed pipeline of text a real, currently living human being wrote, with a paper trail behind it. That’s a genuine practical advantage: slower degradation, a defensible answer to “what did you train this on,” and a supply of language that keeps including the rare, awkward, human sentence that a model trained on its own past output would quietly lose.

Regulation has started reaching toward this problem, but not from the direction that matters here. The EU AI Act’s labelling mandate, which takes effect in August 2026, will require AI-generated images, video and audio to carry machine-readable labels, with fines of up to €15 million for non-compliance. Text is explicitly exempted — the European Commission’s own reasoning is that reliable detection of AI-written text doesn’t exist yet. So the one piece of law built to tag AI output has nothing to say about the market this article is describing: hundreds of millions of dollars a year, changing hands over exactly the kind of text the mandate can’t touch.

What it leaves out, and for whom

The money flows to the platform, not the person. Reddit and News Corp get paid because they own the pipe the content flows through; the individual who posted the comment or wrote the article does not see a share of either deal, and has no legal or technical route to ask for one. Nothing built so far reaches past the platform to the creator.

Smaller and open-source developers are shut out from a different direction. A verified, licensed data pipeline costs tens or hundreds of millions of dollars to assemble. A developer who can’t pay that price is left training on the same open web that’s now more than 74% synthetic, with no way to check what’s clean and what isn’t. The gap this opens is concrete: a well-resourced lab’s model gets steadily better data, a smaller developer’s model gets steadily worse data, and neither the developer nor their users are told which side of that split they’re on.

Government hasn’t closed the gap either. The UK’s March 2026 Report on Copyright and Artificial Intelligence chose to keep the existing rules in place, favouring voluntary licensing and transparency over reform, even though 88% of the more than 10,000 people who responded through the consultation’s Citizen Space platform backed the option that would have required licences for AI training — the strongest protection on the table. The market has already answered the question of what this content is worth — hundreds of millions of dollars a year and rising. Nobody in government has yet answered the question of who, if anyone, is owed a share of it.

My Opinion

I write a lot both here and in my published books, so I read this as someone whose words could end up in the same pipeline. The question of paying for what someone else created isn’t new — publishers have negotiated royalties for a century. What’s new is the scale: this isn’t licensing a book or a photograph, it’s the wholesale absorption of everything a person has ever written, folded into a model with no accounting for whose sentences did the work. Some form of monetary recognition has to follow the creator, not stop at the platform that happened to host them.

The AI industry has just told everyone, in real money, exactly what their unpaid words are worth. It has not told anyone how, or whether, they will ever see any of it.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

The ISS and what comes next

In a factory in Long Beach, California, a team of engineers is building a space station. Not designing one. Not modelling one. Building one — machining aluminium panels, rolling and welding the pressure vessel that will contain breathable air, testing solar arrays and life-support systems. The station is called Haven-1. The company building it is called Vast. And if the schedule holds, a SpaceX Falcon 9 rocket will carry it to low Earth orbit sometime in early 2027.

Haven-1 is small. At 45 cubic metres of habitable volume — roughly the interior of a small tour bus — it is far from the football-field sprawl of the International Space Station, which took over a decade to assemble and cost upwards of $150 billion to build. It will host four astronauts at a time for missions of up to thirty days. Its operational life is three years. It does not try to be everything. It tries to prove that a private company can build and operate a human habitat in orbit — safely, on a realistic schedule, and without the resources of a national space agency.

That proof matters because the era it is trying to inaugurate has a deadline. NASA plans to deorbit the International Space Station around 2030. After twenty-five years as humanity’s permanent outpost in orbit — the most ambitious collaborative engineering project in history, shared by fifteen nations — the ISS will be guided out of the sky and into the Pacific Ocean. What replaces it, if anything does, will not be built by governments. It will be built by companies.

Whether that is exciting or alarming depends on what you think space is for.

The International Space Station was conceived as a symbol as much as a laboratory. Built in the years after the Cold War, assembling modules launched by American Space Shuttles and Russian Soyuz rockets, it expressed something about what cooperation between nations could achieve. The science it enabled was real and significant — advancing our understanding of human physiology in microgravity, producing pharmaceutical research that could not have been conducted on Earth, hosting thousands of experiments across biology, materials science, Earth observation and physics. But it was also, always, enormously expensive.

NASA spends roughly $3 billion per year to operate the ISS. That sum, compounded over the station’s lifetime, has represented a significant fraction of the agency’s total budget. The calculation NASA is now making is whether the science and presence in orbit justify that cost when private companies are willing to build and run orbital infrastructure and sell NASA access to it. The model shifts from landlord to tenant. Instead of owning and operating the station, NASA becomes a customer, buying crew time and research capacity from whoever builds the next generation of platforms. The $150 billion question is whether that works.

The ISS cost over $150 billion to build and costs $3 billion per year to run. NASA’s bet is that private companies can do this cheaper — and that competition will do what government monopoly could not.

Haven-1 is not the only contender. Axiom Space, which has already been awarded a NASA development contract, is launching a Payload Power Thermal Module to the ISS in early 2027; roughly nine months later that module will undock and rendezvous with Axiom’s habitat module to form an independent two-module station by early 2028. Starlab — a joint venture between Voyager Technologies and Airbus, with Northrop Grumman providing autonomous docking capability, and a design that will launch in a single flight aboard SpaceX’s Starship — is targeting 2028. Blue Origin’s Orbital Reef, billed as a “mixed-use business park in orbit,” is in development. Between 2026 and 2031, NASA plans to provide $1–1.5 billion in funded Space Act Agreements to support commercial station development — a programme it has recently revised, shifting from firm fixed-price contracts after officials concluded that a fully commercial business case does not yet exist in low Earth orbit.

The gap between the ISS’s retirement and the arrival of its successors is a genuine concern. NASA’s planning assumes overlap between commercial station operations starting in 2028 and ISS retirement in 2030. If commercial timelines slip — as they have already for Haven-1, which moved from a May 2026 target to Q1 2027 — that overlap narrows. Axiom Space chief executive Michael Suffredini has said publicly that the market may not be large enough for more than one commercial station. The company has faced funding difficulties. The question of whether the economic case for private space stations is as solid as the engineering ambition is one that has not yet been answered by the market.

What space is actually for

The commercial case for space stations rests on a set of claims about what can be done in orbit that cannot be done on Earth — and what that is worth. The most substantive of these claims concerns microgravity manufacturing.

Gravity shapes almost everything about how materials form and how biological processes operate. Protein crystals grown on Earth are distorted by gravity-driven convection; the same proteins grown in microgravity form larger, more perfect structures that reveal their molecular architecture more clearly. Drug development that depends on understanding a protein’s shape can benefit from having better data. Major pharmaceutical companies and JAXA have conducted crystallisation experiments on the ISS. One JAXA study of proteins associated with Duchenne muscular dystrophy led directly to a drug candidate — TAS-205 — now in Phase 3 clinical trials. California startup Varda launched a manufacturing satellite in June 2023, producing crystals of ritonavir — an HIV medication — in orbit and returning the capsule to Earth in February 2024.

Optical fibres grown in microgravity achieve purity levels impossible on Earth. Alloys and semiconductors manufactured without sedimentation can have more homogeneous microstructures. What changes with commercial stations is not the science. It is the access — more capacity, lower cost, broader range of customers, and the possibility of operating manufacturing processes at scale rather than as research demonstrations.

Space tourism is the other commercial pillar. Axiom Space has already sold private astronaut missions to the ISS, at prices reported to start around $55 million per person. Haven-1 will accommodate private visitors. Blue Origin’s Orbital Reef explicitly frames its design around commercial hospitality. The market here is narrower — it requires the kind of wealth that makes $55 million an acceptable discretionary expenditure — but it is real, and it serves the less glamorous purpose of underwriting operational costs that the research market alone may not cover.

The science done on the ISS has already produced pharmaceutical advances, materials breakthroughs and discoveries in human physiology that could not have been achieved on Earth. The question is whether the commercial model can sustain access to that environment after the ISS is gone.

The honest uncertainty is about the size and reliability of the non-government market. NASA’s contribution — buying crew time, hosting government science, funding development — is the anchor. If that anchor holds, the economics of commercial stations are plausible. If NASA’s budget is cut, or its priorities shift, or the agency decides it can achieve its goals without a permanent crewed outpost in low Earth orbit, the business cases of every company in this race become significantly harder.

Who inherits the sky

The ISS was many things, but one of them was political. The same hardware that carried American astronauts carried Russian cosmonauts, European scientists, Japanese engineers, Canadian robotics technologists. The station’s assembly required fifteen nations to agree on standards, interfaces, safety protocols and operational procedures. It survived the post-Cold War period, survived Shuttle disasters, survived the tensions that followed Russia’s 2014 invasion of Ukraine — though the latter ultimately prompted the partners to plan for Russian withdrawal and ISS retirement. Whatever its limitations as a scientific platform, it demonstrated that sustained human cooperation in space was possible.

The commercial stations being built now are American. They are funded by American venture capital and American government contracts. They will be operated under American regulatory frameworks and serve primarily American and allied institutional customers. China, which has built its own Tiangong station — permanently crewed since 2022 — is not a customer. The transition from government-operated to commercially-operated orbital infrastructure in the Western world is also, in a geopolitical sense, a transition from international to national. That is not necessarily wrong. But it is a change that deserves to be named.

The science also changes. The ISS was an instrument of national prestige and international diplomacy, which meant science happened on it even when the immediate commercial return was unclear. Commercial stations will prioritise the experiments customers pay for. That will include pharmaceutical companies crystallising proteins, universities running long-duration biology experiments, technology companies testing hardware in the radiation and vacuum of orbit. It will also mean that certain categories of basic science, whose value is not easily priced and whose benefits arrive decades later, will struggle to find a berth.

What the commercial model has to prove

The case for commercial stations is clear enough on paper. A research environment that currently costs $3 billion per year to run could be made more accessible, more competitively priced, and faster to iterate — if the market works as advertised. The pharmaceutical researcher, the materials scientist, the physician studying bone density in long-duration spaceflight, the engineer testing life-support systems for deep-space missions: all of them benefit if orbital access becomes cheaper and easier rather than harder and more expensive. The platform itself does not matter. What matters is whether the transition sustains the access.

The honest question is whether private operation adds something or merely substitutes one expensive monopoly for another. The ISS cost $3 billion per year under government management. The benchmark for commercial stations is not zero. It is whether, over time, the cost per hour of orbital research comes down, whether the range of customers expands beyond governments and billionaires, and whether the scientific output justifies the investment.

The harder question — the one the market cannot easily answer — is what happens to research that does not have an immediate commercial return. A drug candidate for a neglected tropical disease. A long-term study of bone loss that informs treatment for earthbound osteoporosis patients. A materials science experiment whose applications are twenty years away. These are the kinds of research that the ISS accommodated because governments decided they mattered. Whether commercial operators will find business models that serve them too is not yet clear.

My Opinion

The commercial argument is not wrong on its own terms. Making orbital access cheaper and more competitive is good for science. What I’m not convinced by is the assumption that the market will fund the research that takes twenty years to return a result — the kind that doesn’t fit a quarterly report or a drug application. Space has always held something that cannot easily be priced: the pull of the unknown, the technologies that come back to Earth from solving problems nobody had imagined, the slow accumulation of knowledge about what the species can survive and where it might eventually go. The ISS existed because governments decided that the unknown was worth funding. I’d like the decade ahead to find a way to carry that conviction into whatever comes next — not because it’s commercially compelling, but because it’s how we stay curious about a universe that doesn’t owe us answers.

What the decade ahead has to answer

A commercial model optimises for what can be priced. Several of the most consequential questions about the next era in orbit cannot be priced — not yet, possibly not ever.

1. The ISS was built by fifteen nations working together across geopolitical fault lines. The commercial stations being built now are almost entirely American. What does the transition from international to nationally anchored orbital infrastructure mean for science, for diplomacy, and for who gets to go to space?

2. Haven-1 slipped from May 2026 to early 2027. Axiom is facing funding pressures. The gap between ISS retirement and commercial successors may be tighter than planned. What happens to continuous human presence in low Earth orbit if the commercial stations arrive late?

3. The ISS funded science that had no immediate commercial return because governments decided that science mattered. Commercial stations will fund the experiments that customers pay for. Is there a type of research you think is important that might not survive that transition?

4. A private astronaut mission to the ISS costs around $55 million per person. Haven-1 will be no different in its early missions. At what point — if ever — does the price drop far enough that the question of who gets to go to space becomes meaningfully different?

The last astronaut to walk on the Moon did so in 1972. For more than fifty years, human beings have not left low Earth orbit. The ISS has been the closest thing to a frontier that most of us will ever witness — a small, fragile, extraordinarily expensive outpost at the edge of the world’s reach.

What replaces it will be built not by nations but by companies, and financed not by ideology but by markets. Whether that produces something better, something worse, or simply something different from what the ISS represented will become clear in the decade ahead.

Leave a comment

Sources & references

Haven-1 and Vast

• Haven-1 — Wikipedia

• Vast Delays Haven-1 Launch to 2027 — Payload Space

• Vast Advances Haven-1 Into Integration Phase — Vast Space

Axiom Space

• Axiom Station — Wikipedia

• Vast and Axiom awarded new private missions to ISS — NASASpaceFlight

• Axiom Space Secures $350M in Financing — Axiom Space

• Axiom Space puts a brave face on claims of dire finances — The Register

• Axiom Space raises $350M backed by Qatar fund — CNBC

Starlab

• Starlab Space Station — Wikipedia

• Private Starlab space station moves into full-scale development — Space.com

• Voyager Space Announces Teaming Agreement with Northrop Grumman for Starlab — PR Newswire

NASA Commercial LEO Destinations programme

• Commercial Space Stations — NASA

• NASA releases details on revised next phase of commercial space station development — SpaceNews

• Payload Field Guide: Commercial LEO Destination — Payload Space

Varda Space / ritonavir

• HIV drugs manufactured in space return to Earth — Healio

• Return of the Ritonavir — Varda Space Industries

• Rocket Lab successfully returns spacecraft capsule to Earth — Rocket Lab

JAXA protein crystallisation / Duchenne muscular dystrophy

• Creating new and better drugs with protein crystal growth experiments on the ISS — NASA

• Creating New and Better Drugs with Protein Crystal Growth Experiments — SpaceNews

ISS context and commercial transition

• Commercial Efforts Flail to Replace Aging International Space Station — Engineering

  ---

  You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

A case documented in the medical literature described a patient with pancreatic cancer who missed out on a potentially life-prolonging treatment because their scan result was delayed by eighteen months. The scan had been taken. The images existed. No one had looked at them in time.

That case was not unusual enough to be a scandal. It was unusual enough to make the journal. The background against which it sits — a radiology system under sustained and worsening pressure, in which delays are the norm rather than the exception — has become so familiar that it no longer registers as news. It is simply the condition patients navigate when something goes wrong inside their heads and they need to know what it is.

In September 2025, more than 386,000 people in England were waiting longer than six weeks for a diagnostic test. The total number waiting for any diagnostic test exceeded 1.7 million. The Royal College of Radiologists has documented a 30% shortfall of clinical radiologists in England — roughly 2,000 doctors short of what is needed to meet current demand — with the gap projected to reach 39% by 2029. Forty-six percent of NHS acute trusts are not meeting even the interim waiting time target.

A system that can read a brain MRI scan in seconds has arrived. Not a prototype. Not a laboratory demonstration. A system trained on more than 220,000 real clinical scans, tested across more than 30,000 studies in a running health system over a full year, published in February 2026 in Nature Biomedical Engineering. The question is not whether the technology works. It is why, given that it works, the gap between scan and diagnosis remains so wide for so many patients.

What the AI can actually do

The system is called Prima. It was developed by a team at the University of Michigan led by neurosurgeon Todd Hollon, and it works differently from most AI diagnostic tools that have preceded it.

Earlier AI models in radiology were built for specific tasks: detecting a particular kind of lesion, flagging a certain cancer type, screening for a defined abnormality. They were useful for the thing they were designed for and nothing else. Prima was trained on everything — every MRI study taken at the University of Michigan Health system since radiology records were digitised, covering more than 220,000 studies and 5.6 million imaging sequences. It was trained not just on the images themselves but on the patients’ clinical histories and the reasons each scan had been ordered by a physician. It works, as its developers describe it, the way a radiologist works: integrating what the image shows with what the patient’s medical history tells you.

Across more than 50 different radiological diagnoses — stroke, brain haemorrhage, tumours, multiple sclerosis, aneurysms and more — Prima outperformed other state-of-the-art AI models, achieving a mean diagnostic accuracy of 92% across that range. It can identify which cases require urgent attention and immediately alert the appropriate specialist: a stroke neurologist if the scan suggests a stroke, a neurosurgeon if it suggests a bleed. Feedback is available as soon as imaging is complete. The result that would otherwise take days or weeks to arrive takes seconds.

Prima was trained on every MRI taken at a major health system since records were digitised. It works the way a radiologist works — integrating image data with clinical history. The result takes seconds.

UCL and King’s College London published parallel work in December 2025. Their model, trained on 60,000 brain MRI scans and the radiology reports written about them, accurately distinguished normal from abnormal scans approximately 19 times out of 20. A randomised multicentre trial across UK hospitals started in 2026 to assess how the system performs in live clinical workflows. These are not the same research programme, but they point in the same direction: AI that can read a brain scan fast and flag what matters is operationally real, not theoretical.

For acute neurological conditions, stroke above all, the stakes of speed are measurable in the specific. In stroke care, the standard principle is time is brain: for every hour of untreated ischaemic stroke, roughly 120 million neurons are lost. The gap between scan and treatment decision is a gap in which irreversible damage accumulates. A system that compresses that gap from hours to seconds is not a marginal improvement in workflow efficiency. It is a different category of clinical event.

Why the gap persists

If the technology exists and the problem is severe, the natural question is why the gap between them remains so large. The answer is not simple and it is not flattering to any single part of the system.

Regulatory approval is one obstacle. AI diagnostic tools that make clinical decisions about individual patients are classified as medical devices in the UK and the EU. The approval process is rigorous — necessarily so, given the consequences of getting it wrong — but it is not designed for the pace at which AI research is producing results. The models that receive approval tend to be narrow-task systems, because narrow tasks are easier to validate; a foundation model like Prima, which makes diagnoses across 50 conditions, presents a different and more complex challenge. The UCL team’s next step — a randomised multicentre trial in UK hospitals — is the appropriate pathway, and it is one that takes years.

Integration is a second obstacle. An AI model that produces a result in seconds is only useful if that result reaches the clinician who needs it, in a format they can act on, connected to the patient record system they are already using. NHS radiology departments typically operate on legacy IT infrastructure. The connections between imaging systems, electronic patient records, and clinical communication tools are incomplete, inconsistent, and often not interoperable. The radiologist shortage means that even when AI flags a case as urgent, there may not be a specialist immediately available to review the flag and make the treatment decision.

Liability sits beneath both. When an AI system incorrectly categorises a scan — flags something urgent that is not, or misses something that is — the question of who is responsible is unresolved in most legal and professional frameworks. The EU AI Act, whose main provisions came into application in August 2026, classifies medical AI diagnostic tools as high-risk and requires documentation of training data, bias checks, and human oversight policies — though the provisions specifically covering CE-marked medical devices are not due to apply until 2027. Radiologists operating under professional accountability for their diagnoses have every reason to be cautious about delegating to a system whose error profile is not yet fully understood.

The technology compresses the gap from hours to seconds. The regulatory pathway, the IT infrastructure, and the liability framework compress it from years to months. That asymmetry is the problem.

The Royal College of Radiologists has been clear that AI represents one of the most significant changes in medical care since the inception of the NHS — and equally clear that far from replacing radiologists, AI creates new roles and responsibilities for them. The challenge is not the technology’s relationship with the profession. It is the system’s capacity to absorb a disruptive tool without the disruption falling, as it usually does, on the patients.

Who bears the cost of the gap

Between October 2024 and September 2025, an average of more than 500,000 people were waiting each month for a CT or MRI scan in England. That figure increased by nearly 19,000 per month compared to the previous year. More than 74,000 of those waiting longer than six weeks were waiting for a CT or MRI scan necessary to diagnose or rule out cancer and other time-critical conditions.

The people who bear the cost of the gap are not evenly distributed. Delays in imaging disproportionately affect patients in regions with the fewest radiologists, which are not the regions with the least need. The East of England had the highest proportion of patients waiting too long for a diagnostic test in September 2025. Rural and lower-income areas consistently face longer waits than urban centres with large teaching hospitals. The patients most likely to experience an eighteen-month wait for a scan result are not the patients with the most access to private imaging, second opinions, or the capacity to advocate loudly for themselves in a system under pressure.

The documented case of the pancreatic cancer patient is one example of a pattern that oncologists and radiologists describe consistently. One consultant oncologist, quoted in a Royal College of Radiologists report, put it plainly: delays in scans and treatment were resulting in missed or late cancer diagnoses, with some patients’ conditions deteriorating to a stage where treatment was no longer possible. These are not patients who would be well-served by a technology that exists in a Michigan laboratory and a multicentre trial and not yet in their district general hospital.

My Opinion

Wait times are widely reported. What is less discussed is that even reaching the top of the queue offers no guarantee — patients are routinely displaced at the last moment when capacity shifts. Technology addresses demand; it does not address what happens after the scan. A faster result still requires a specialist to act on it, a treatment pathway to follow, and a system prepared to move quickly.

I have seen, first hand, what happens when effective health technology runs into departmental self-preservation. In much of the NHS, budgets are allocated on the basis of patient volumes. A department that uses AI to reduce the number of patients requiring manual review is a department that may face a reduced budget in the next cycle. The incentive runs directly against adoption. The most promising diagnostic AI I have encountered in NHS settings has repeatedly stalled not because it failed clinically but because it succeeded — and success threatened the funding model. Until the NHS addresses how it allocates resource to departments that adopt technology effectively, the gap between what is technically possible and what patients actually receive will remain wide.

Between the result and the patient

For a system 2,000 radiologists short of its stated need, an AI that effectively multiplies radiologist capacity is not a productivity tool — it is a patient safety intervention. The question is whether it is being deployed as one. There is a version of AI adoption in radiology that improves throughput statistics without improving what happens to the patient at the end of the queue: a system that accelerates the reporting of straightforward scans while leaving the complex and urgent ones with the same human bottleneck has improved the metric without improving the care. The test is whether the AI is being deployed where the clinical need is highest — in stroke pathways, in urgent cancer triage, in rural hospitals with no overnight radiologist — not where the implementation is easiest.

The infrastructure for deploying it that way does not yet exist in a consistent form across the NHS. What is needed is not just a model that works in a research setting, but a system that responds to the specific circumstances of each patient — their geography, their condition’s urgency, the resources available at their local trust. The Community Diagnostic Centres being built across England are a step toward that. The NHS 10-Year Workforce Plan consultation, in which the Royal College of Radiologists called for AI to be central to workforce strategy, is a step toward that. But the distance between the step and the patient who waited eighteen months for a scan that was already taken remains, for now, very wide.

The decisions that haven’t been made

The technology, the shortage, and the incentive problem have been outlined, what follows are not rhetorical questions. Each one has an answer — and each answer requires someone in government, in NHS leadership, or in clinical governance to decide that the gap matters enough to close it.

• If you or someone close to you has been on a waiting list for an MRI or CT scan, how long did it take? Was the delay ever explained? And did you know that 500,000 people are in that position every month in England?

• The technology to read a brain MRI in seconds exists. The shortage of radiologists to read them at all is severe and worsening. If you were designing a health system’s response to that combination, where would you start — with the AI, with the workforce, or with the infrastructure that connects them?

• The documented case of the patient who waited eighteen months for a scan result is one example. How many similar cases are undocumented — unreported because the patient did not know the delay had happened, or because no one connected the delayed report to the deteriorating outcome?

• AI that triages brain scans is being developed primarily in large academic health systems in wealthy countries. The global radiologist shortage is most severe in low and middle-income countries. What would a deployment pathway that addressed that imbalance look like — and who would need to build it?

A brain MRI takes around thirty minutes to acquire. Prima reads the result in seconds. The average NHS patient waits weeks. The technology does not close that gap automatically. Deploying it at the scale and in the places where it would matter most requires decisions about regulation, infrastructure, liability and investment that have not yet been made.

Those decisions are not technical. They are political. And they will be made — or not made — while patients wait.

Leave a comment

Sources and references

• Prima AI system: Hollon et al., “Learning neuroimaging models from health system-scale data”, Nature Biomedical Engineering, February 2026.
  https://www.nature.com/articles/s41551-025-01608-0

• Michigan Medicine press release: “An AI model that can read and diagnose a brain MRI in seconds.” https://www.michiganmedicine.org/health-lab/ai-model-can-read-and-diagnose-brain-mri-seconds

• UCL and King’s College London brain MRI model: UCL News: “AI model helps speed up diagnosis of brain disorders”, December 2025. https://www.ucl.ac.uk/news/2025/dec/ai-model-helps-speed-diagnosis-brain-disorders

• King’s College London: “AI brain scan model identifies stroke, brain tumours and aneurysms.” https://www.kcl.ac.uk/news/ai-brain-scan-model-identifies-stroke-brain-tumours-and-aneurysms

• NHS diagnostic waiting times (September 2025): Royal College of Radiologists: “Nearly half of NHS trusts missing test waiting time target as backlogs grow.”
  https://www.rcr.ac.uk/news-policy/latest-updates/nearly-half-of-nhs-trusts-missing-test-waiting-time-target-as-backlogs-grow/

• NHS Digital: “Diagnostic waiting times and activity for September 2025.”
  https://digital.nhs.uk/data-and-information/publications/statistical/nhse-diagnostics-waiting-times-and-activity-data/for-september-2025

• Radiologist workforce shortfall: Royal College of Radiologists: “2024 workforce census reports lay bare the challenges facing radiology and clinical oncology.”
  https://www.rcr.ac.uk/news-policy/latest-updates/2024-workforce-census-reports-lay-bare-the-challenges-facing-radiology-and-clinical-oncology/

• Stroke neuron loss: Saver JL, “Time Is Brain — Quantified”, Stroke, 2006.
  https://www.ahajournals.org/doi/10.1161/01.str.0000196957.55928.ab

• EU AI Act — medical device provisions: Kennedys Law: “The EU AI Act implementation timeline: understanding the next deadline for compliance”, 2026.
  https://www.kennedyslaw.com/en/thought-leadership/article/2026/the-eu-ai-act-implementation-timeline-understanding-the-next-deadline-for-compliance/

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

It leaves no trace. No alerts fire. No systems misbehave. The attack, if it is happening, looks identical to normal network traffic. The adversary captures encrypted data — patient records, financial transactions, strategic communications, diplomatic cables — and stores it. Then they wait. They wait for a quantum computer capable of breaking the encryption to exist. When it does, they decrypt everything they collected. The breach happened years earlier. The victim had no idea.

This is called harvest now, decrypt later. It is not a theoretical future scenario. The US Department of Homeland Security, the UK’s National Cyber Security Centre, the European Union Agency for Cybersecurity, and the Australian Cyber Security Centre have all issued official guidance stating that adversaries are already doing this. State actors and advanced persistent threat groups are collecting encrypted data at scale, banking on quantum decryption becoming available within the next decade or two.

The good news, and there genuinely is good news, is that the solution exists and has been standardised. In August 2024, after an eight-year international evaluation process, the National Institute of Standards and Technology (NIST) published three post-quantum cryptography standards that organisations can implement today to protect their systems against quantum attacks. The standards are ready. The algorithms are tested and approved. The path is clear.

The bad news is that most organisations have not yet completed, or in many cases begun, their cryptographic inventories.

Why current encryption breaks

Almost all public-key cryptography — the kind that secures your bank login, your email, your medical records, and the connections between servers that underpin the internet — relies on the difficulty of certain mathematical problems. RSA encryption, which is used to protect the majority of internet traffic, depends on the fact that factoring the product of two very large prime numbers is computationally infeasible for classical computers. The numbers are so large, and the factoring process so slow, that it would take longer than the age of the universe for a classical machine to break a properly implemented RSA key. The security is not derived from complexity of design. It is derived from the limits of classical computation.

A quantum computer running an algorithm called Shor’s algorithm — published by Peter Shor in 1994 (SIAM Journal 1997), thirty years before any machine existed that could run it at useful scale — can factor large numbers exponentially faster than any classical approach. What takes a classical computer millions of years takes a quantum computer, at sufficient scale, hours. RSA breaks. Elliptic curve cryptography, which is used in everything from TLS web security to blockchain transactions, breaks. Diffie-Hellman key exchange, used to establish secure connections, breaks. The mathematical foundations that the entire architecture of digital security is built on become vulnerable.

The security of RSA encryption is not derived from complexity of design. It is derived from the limits of classical computation. A quantum computer changes those limits. When it does, the encryption breaks.

The timeline for when a quantum computer capable of doing this will exist is genuinely uncertain. The most optimistic estimates suggest the early 2030s. More cautious assessments say the mid-to-late 2030s. Some researchers place it later still. What is not uncertain is that quantum computing is advancing rapidly, that the below-threshold error correction result demonstrated by Google in December 2024 validates the theoretical path, and that once a cryptographically relevant quantum computer exists, breaking encryption will be a matter of running software, not of further hardware development.

The harvest now, decrypt later threat means this uncertainty is beside the point for data that needs to remain confidential for more than a decade. A diplomatic communication encrypted today that must remain secret for twenty years is already at risk — because if an adversary captures and stores it now, they do not need to decrypt it today. They need to decrypt it before the twenty years are up. Medical records, legal correspondence, intellectual property, strategic financial plans, national security communications — all of these have lifespans that could extend into the quantum era. The relevant question is not when quantum computers arrive. It is how long your data needs to remain protected.

The standards are ready. The migration is not.

NIST’s three finalised post-quantum cryptography standards — ML-KEM (for key encapsulation), ML-DSA (for digital signatures), and SLH-DSA (a hash-based signature scheme) — are built on mathematical problems that quantum computers cannot efficiently solve. They have undergone eight years of rigorous international evaluation, involving hundreds of cryptographers across academia, industry and government. They are not experimental. They are ready to implement.

The transition timeline is clear. NIST’s guidance calls for quantum-vulnerable algorithms to be deprecated and ultimately removed from its standards by 2035, with high-risk systems expected to migrate considerably earlier. The NSA’s Commercial National Security Algorithm Suite 2.0 framework requires quantum-safe algorithms for all new US national security systems by January 2027, full application migration by 2030, and complete infrastructure migration by 2035 — with compliance cascading through defence contractors, federal agencies and regulated industries. The UK’s NCSC has published equivalent guidance with target dates for migration activities. The European and Australian cybersecurity agencies have done the same.

The problem is execution. Identifying where quantum-vulnerable algorithms are used across an enterprise turns out to be surprisingly difficult. Cryptography is embedded in operating systems, in hardware, in libraries, in protocols, in custom-built applications, in supplier systems, in legacy infrastructure that predates anyone currently working at the organisation. The protocols most at risk — TLS, which secures most internet traffic; IKE, used in VPNs; S/MIME, used in email; code signing — depend on public-key cryptography and will all need to be migrated. Some of these migrations are straightforward software updates. Others require hardware replacement, which takes years and significant capital expenditure.

The complexity is compounded by interdependency. The first step in any migration is a cryptographic inventory: a complete map of every system, protocol and library that uses cryptography. Most organisations do not have one. Building it requires examining every layer of the technology stack, including components provided by third parties who may themselves not yet have assessed their quantum exposure. The Federal Reserve’s research division published analysis in 2025 examining the harvest now, decrypt later threat to distributed ledger networks, finding that blockchain transactions committed today cannot be retroactively re-encrypted — making the permanence that gives blockchains their value also their greatest vulnerability to quantum attack.

There is also a practical distinction about symmetric encryption. AES-256 and the SHA-3 family of hash functions are not broken by quantum computers — or rather, they are weakened but not broken, in a way that is addressed simply by using larger key sizes that are already standard practice. The algorithms that break catastrophically under quantum attack are the asymmetric, public-key schemes. Organisations that have already moved to AES-256 throughout are not in as exposed a position as those still running older symmetric standards. But the public-key infrastructure that secures their communications, authenticates their users and signs their code is still vulnerable.

The standards to protect against quantum attacks were finalised in August 2024. The first step is a cryptographic inventory — a complete map of every system using quantum-vulnerable algorithms. Most organisations do not have one.

The new post-quantum algorithms come with larger key sizes and signature sizes than their classical predecessors, and some introduce performance overhead. Migrating to ML-KEM and ML-DSA is not a drop-in replacement for RSA everywhere — it requires careful implementation, particularly in constrained environments like embedded firmware, Internet of Things devices, and legacy systems with limited processing capacity. The industry’s guidance is to begin with hybrid deployments where possible: running the new post-quantum algorithms alongside existing classical schemes, so that security is maintained against classical attacks during the transition and gradually shifted to quantum-resistant-only. This is more complex than a clean cutover, but it is the responsible approach for systems that cannot afford any gap in protection.

What is actually at risk

The data categories most exposed to the harvest now, decrypt later threat share one characteristic: they need to remain confidential for a long time. Government and military communications — diplomatic cables, intelligence assessments, operational plans — are the most frequently cited category, because state actors with the capability to harvest at scale are the most likely harvesters, and national security data has the longest confidentiality requirements. The exposure is direct and serious.

Healthcare data carries similar risks. Medical records are retained for decades and contain some of the most personal information people generate. A patient’s HIV status, mental health history, genetic conditions, reproductive choices — all encrypted in hospital systems today, all potentially sitting in an adversary’s storage waiting for the tools to read them. The same applies to legal communications protected by attorney-client privilege, to research data underpinning pharmaceutical competitive advantage, to financial records containing the kind of material non-public information that could move markets if exposed years later.

The most alarming aspect of harvest now, decrypt later attacks is that they leave no visible signs. No files are corrupted. No systems are disrupted. No ransom note arrives. The breach is silent, and the damage is deferred. An organisation can be entirely unaware that its most sensitive communications from the past decade have been collected and are sitting in an archive waiting for decryption. By the time the damage becomes visible — when a quantum computer makes the archive readable — it is too late to protect the data that was already taken.

This is not a risk profile that fits the usual mental model of a cyber attack. Most security investment, most incident response planning, most board-level conversation about cyber risk is oriented around attacks that are visible and immediate. Harvest now, decrypt later is invisible and patient. The discipline it requires is different: not reactive, but anticipatory. The investment it requires must be made before any damage occurs, against a threat whose timing is uncertain, for a payoff that consists of nothing happening.

Who gets left exposed

What is not yet clear is who gets there first, and who is left last. The largest technology companies, banks and government agencies are already in the process of migration. What is not yet certain is whether the organisations that hold the most sensitive data about ordinary people — healthcare providers, legal systems, social service agencies — have the resources, the expertise, and the urgency to migrate before the threat becomes critical. A well-resourced adversary harvesting encrypted data at scale is not particularly interested in the records of major academic medical centres, which will likely be early adopters. It is interested in the records of district hospitals, GP practices and national health registries, which may be last.

Mosca’s Theorem, named after cryptographer Michele Mosca of the University of Waterloo’s Institute for Quantum Computing, states the problem precisely: if X is the number of years your data must remain secure, Y is the time required to complete your migration, and Z is the time before a cryptographically relevant quantum computer exists, then you must start migrating before X + Y exceeds Z. Every year of delay reduces the margin. For organisations with long-lived sensitive data, the margin may already be thin.

An organisation that has not yet begun its cryptographic inventory does not know which of its systems are most exposed. That inventory — a complete map of every system, protocol and library that uses cryptography — is the first step NIST, the NCSC, and every cybersecurity agency has now recommended. Identify the systems. Assess which data has the longest confidentiality requirement. Prioritise in order of exposure and sensitivity. Do not wait for a quantum computer to exist before beginning.

My Opinion

Every organisation has a security budget. What priority it receives depends on the nature of the business, the regulatory environment, the threats faced, and how many incidents have already landed. That calculus has always been reactive — weighted toward visible risks with measurable consequences. Post-quantum cryptography breaks that model completely. There is no incident to detect, no alert to fire, no visible breach to respond to. The only way to know you were compromised may be a ransom demand arriving a decade from now, or a disclosure you will never see. Cyber security needs a new conversation in the PQC world, and it needs to start with the organisations that cannot afford to have it — because those are exactly the ones holding the data that is most worth stealing.

Before the computers arrive

1. Does the organisation you work for, or depend on, have a cryptographic inventory — a complete map of every system using quantum-vulnerable encryption? If not, does it have a plan to produce one, and a timeline for when that will be done?

2. What data held by organisations you interact with — your bank, your GP, your employer, your government — has a confidentiality requirement that extends ten, fifteen, or twenty years into the future? Is any of it still protected by RSA or elliptic curve cryptography?

3. The harvest now, decrypt later attack leaves no visible trace. How would your organisation know if it had already happened? What would the discovery process look like, and is there a version of it that is not simply ‘when the quantum computer arrives and someone tells us’?

4. he organisations most likely to complete PQC migration early are the ones with the largest security budgets. The organisations holding the most sensitive data about the most people are often not those organisations. What does an equitable migration strategy look like — one that protects not just well-resourced institutions but the full range of data that needs protecting?

The breach that hasn’t happened yet does not have to happen at all.

Leave a comment

Sources & References

Standards and official guidance

• NIST Post-Quantum Cryptography Standards (August 2024): ML-KEM (FIPS 203), ML-DSA (FIPS 204), SLH-DSA (FIPS 205) https://csrc.nist.gov/projects/post-quantum-cryptography

• NIST — HQC selected as fifth algorithm (March 2025): https://www.nist.gov/news-events/news/2025/03/nist-selects-hqc-fifth-algorithm-post-quantum-encryption

• NIST IR 8547 (Initial Public Draft, November 2024): Transition to Post-Quantum Cryptography Standards — deprecation timeline to 2035 https://csrc.nist.gov/pubs/ir/8547/ipd

• NSA — Commercial National Security Algorithm Suite 2.0 (CNSA 2.0): Quantum-safe deadlines for national security systems; January 2027 procurement gate https://media.defense.gov/2022/Sep/07/2003071836/-1/-1/0/CSICNSA_2.0_FAQ.PDF

• UK National Cyber Security Centre — Timelines for migration to post-quantum cryptography (March 2025):Three-phase roadmap: assessment by 2028; prioritised upgrades 2028–2031; full migration by 2035 https://www.ncsc.gov.uk/guidance/pqc-migration-timelines

• US Department of Homeland Security — Post-Quantum Cryptography: Official DHS guidance including harvest now, decrypt later threat confirmation https://www.dhs.gov/quantum

• ENISA — Post-Quantum Cryptography: Current State and Quantum Mitigation: EU Agency for Cybersecurity guidance; EU coordinated roadmap targets high-risk systems by 2030, full transition by 2035 https://www.enisa.europa.eu/publications/post-quantum-cryptography-current-state-and-quantum-mitigation

• Australian Cyber Security Centre — Post-Quantum Cryptography: Official ACSC guidance on harvest now, decrypt later threat https://www.cyber.gov.au/resources-business-and-government/maintaining-devices-and-systems/system-hardening-and-administration/post-quantum-cryptography

Research and primary sources

• Shor, P. (1994 / 1997) — Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer: Note: first presented at FOCS 1994 (the date used in the article); the formal journal paper was published in SIAM Journal on Computing, 1997. https://arxiv.org/abs/quant-ph/9508027 SIAM Journal version: https://epubs.siam.org/doi/10.1137/S0097539795293172

• Google Quantum AI — Willow processor announcement (December 2024): Below-threshold error correction result — first quantum chip to demonstrate error rates that improve as qubit count scales https://blog.google/innovation-and-ai/technology/research/google-willow-quantum-chip/

• Federal Reserve Finance and Economics Discussion Series 2025-093: “Harvest Now, Decrypt Later” — analysis of harvest threat to distributed ledger networks; blockchain immutability as a specific vulnerability https://www.federalreserve.gov/econres/feds/files/2025093pap.pdf

• Michele Mosca — Mosca’s Theorem: Co-founder and Deputy Director, Institute for Quantum Computing, University of Waterloo https://uwaterloo.ca/institute-for-quantum-computing/

Market data

• MarketsandMarkets — Post-Quantum Cryptography Market (2025): PQC market projected to reach $2.84bn by 2030, CAGR 46.2% https://www.marketsandmarkets.com/PressReleases/post-quantum-cryptography.asp

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Somewhere beneath the Vijf Eikentunnel in Rijen, a railway underpass in the Netherlands, concrete is doing something concrete has never done before. A crack has formed — the kind that appears in any heavily loaded structure over time, the kind that, left alone, lets water in and corrosion begin and the slow degradation that ends, years or decades later, in a repair bill or a closure or a failure. Except this concrete is not waiting to be repaired. Dormant bacteria embedded in the mix have been activated by the water entering the crack. They are feeding on calcium-based nutrients and producing limestone. The crack is sealing itself.

The idea that a material could detect its own damage and fix it without any external intervention sounds like a property borrowed from living organisms. That is not a coincidence. The research behind self-healing materials draws explicitly on the biology of healing — on how skin closes a wound, how bone knits after a fracture, how trees seal damage to their bark. The scientific ambition is to replicate those capabilities in the inert materials that hold the built world together: concrete, steel, polymers, coatings.

The built world needs this badly. Corrosion and material degradation cost the global economy an estimated $2.5 trillion every year — roughly 3.4% of global GDP, according to a 2016 NACE International study, the most comprehensive global assessment of corrosion costs undertaken. That figure covers the direct costs of repair and replacement and a portion of the indirect costs of disruption, but not the full human cost of infrastructure failures, delayed maintenance, or the carbon embedded in the concrete and steel that gets thrown away and rebuilt. The scale of the problem is so large and so distributed that it rarely registers as a single issue. Cracks in bridges, rust in pipelines, spalling on car park decks, moisture ingress in tunnel linings — these are experienced as individual nuisances, not as a systemic failure of how we build things.

Self-healing materials are a systemic answer to a systemic problem. The question is how close they are to being usable at scale, and what it would take to get them there.

How materials learn to heal

There are two distinct approaches to making a material that can repair itself, and the distinction matters for understanding what each approach can and cannot do.

The first approach is extrinsic: healing agents are stored within the material and released when damage occurs. The most widely studied version of this embeds microcapsules — tiny containers filled with a liquid resin, a polymer, or a mineral-producing compound — directly into the material during manufacture. When a crack propagates through the material, it ruptures the capsules along its path. The healing agent flows into the crack, reacts with a catalyst or with the surrounding material, and solidifies. The crack is sealed. The material continues to function. The bacteria-based concrete works on a similar principle, with dormant bacterial spores taking the place of chemical capsules: water activates the bacteria, the bacteria produce calcite crystals, the crystals seal the gap.

The second approach is intrinsic: the molecular structure of the material itself contains bonds that can break and reform. When damage occurs at the microscopic level, these dynamic bonds — hydrogen bonds, reversible covalent bonds, supramolecular interactions — release and reconnect, the material flows slightly toward the damaged region and the damage heals. This approach works well in polymers and elastomers, where molecular mobility is sufficient to allow the necessary flow. It does not, in general, work in rigid materials like concrete or steel, where the molecular architecture does not allow that kind of movement.

Dormant bacteria embedded in concrete activate when water enters a crack. They produce limestone crystals that seal the gap. The material fixes itself without any human intervention. This is not a laboratory demonstration. It has been built into real structures.

The most advanced real-world deployments of self-healing materials are in concrete, and the most credible research has come from Delft University of Technology in the Netherlands, where professor Henk Jonkers and his team developed the bacteria-based approach over more than a decade. The world’s first building constructed with bacteria-based self-healing concrete was a lifeguard station by a lake in the Netherlands, a demonstration project built by the Delft group. Two further full-scale demonstrator projects followed.

Basilisk, the commercial spinout from the Delft research, expanded deployment of its bacteria-based healing agent into large-scale infrastructure projects in 2025. These are not controlled laboratory experiments. They are structures in the field.

The laboratory results are substantive. For cementitious mixes with inorganic self-healing admixtures designed to seal cracks up to 0.3 millimetres wide, permeability-based healing indices — a measure of how effectively the crack has been sealed — regularly exceed 90% at 28 days. Superabsorbent polymer systems tested against 0.3-millimetre cracks lowered total water flow through the crack by up to 98%.

Beyond concrete, self-healing polymer coatings on car surfaces can repair minor scratches using near-infrared light from sunlight. Self-healing tyres — which use a soft polymer layer that flows into punctures and reseals them — are available commercially from several major manufacturers. Self-healing polymer electrolytes for lithium-ion batteries are an active area of research, addressing the micro-cracking of electrodes that limits battery longevity.

The gap between the laboratory and the bridge

The science works. The engineering challenges that stand between a working laboratory material and wide deployment in real infrastructure are substantial but not unsolvable, and the most important one is cost.

Self-healing concrete costs more to produce than conventional concrete. The bacteria, the encapsulating agents, the additional admixtures — all of these add to the price per cubic metre. The premium varies by system and scale, but it is real, and it creates an immediate procurement problem. Infrastructure projects are costed on initial build price, not on lifecycle cost. A bridge specified with self-healing concrete costs more upfront. The maintenance savings accrue over decades, to a different budget, potentially under a different political administration. The person signing the initial contract and the person who would benefit from lower repair bills twenty years later are usually not the same person, and often not the same organisation.

This is not a technical problem. It is a procurement and accounting problem. Whole-life costing — evaluating infrastructure based on the total cost over its operational lifetime rather than just the initial build price — is standard guidance in most developed countries’ infrastructure policy frameworks. It is applied inconsistently. The Treasury Green Book in the UK explicitly requires whole-life cost assessment for public projects. Modelling of US concrete lifecycle emissions found that a 50% increase in structural longevity would have reduced associated CO₂ emissions by an estimated 14% over the twentieth century. Self-healing concrete that extends maintenance intervals well beyond that would compound the reduction further. A bridge built for a 50-year maintenance cycle rather than a 25-year one does not just save money. It reduces the carbon embedded in repeated repair cycles, the disruption of lane closures, and the probability of a failure caused by a crack that was never found.

The second constraint is crack width. Bacteria-based and capsule-based self-healing systems work best on cracks up to around 0.3 to 0.5 millimetres wide. Beyond that threshold, the healing agent cannot bridge the gap effectively. This matters because the cracks that most threaten structural integrity are not always small ones. A mature concrete structure that has been under load for thirty years may develop cracks that exceed what current self-healing systems can address. The technology is most powerful as a preventive intervention — stopping small cracks from becoming large ones — not as a remedial one for structures that are already significantly damaged.

The third constraint is one-time use. Most extrinsic self-healing systems — the capsule-based and bacteria-based approaches — can heal a given location once. The capsules rupture and are gone. The bacteria are consumed. If a crack opens again at the same location, the healing capacity has been depleted. Intrinsic polymer systems, where dynamic molecular bonds reform after each damage event, can in principle heal the same location multiple times. Research into multi-cycle concrete healing systems is ongoing, but the one-time-use limitation of the most commercially advanced systems is a real constraint on where and how they are deployed.

What infrastructure could look like

The consequences of widespread adoption of self-healing materials are not just financial, though the financial case is real. Cement production is responsible for approximately 8% of global CO₂ emissions — one of the largest single industrial sources of greenhouse gases. A consequential driver of cement consumption is the repair and replacement of structures that have degraded faster than their design life suggested. Self-healing concrete that doubles the effective maintenance interval of a structure halves, roughly, the cement demand associated with keeping that structure operational.

Infrastructure degradation disproportionately affects communities where the capacity to fund maintenance is limited — rural road networks, ageing urban bridges in lower-income municipalities, water and sewage infrastructure in countries where regular inspection and repair is not operationally feasible. A pipeline that monitors and seals its own small leaks does not require the inspection regime that a conventional pipeline does. A bridge deck that arrests crack growth autonomously does not generate the maintenance costs that have accumulated as structural degradation.

In England and Wales, a 2025 industry assessment found that 17% of the local road network is in poor condition — structural deterioration that has built up because maintenance funding has not kept pace with deterioration. That gap is not primarily a funding problem. It is also a materials problem.

The field is also moving rapidly beyond concrete into applications with different but equally consequential implications. Self-healing coatings on offshore wind turbines would reduce maintenance costs for one of the most expensive and difficult-to-access classes of infrastructure. Self-healing insulation on undersea power cables would extend their operational life considerably. Self-healing polymer housings for electronic devices would reduce the rate at which functional devices are discarded due to physical damage — one contributor to the 62 million tonnes of electronic waste generated globally in 2022, a figure rising by 2.6 million tonnes every year, according to the Global E-waste Monitor 2024. Self-healing battery electrodes, if they reach sufficient maturity, could extend the calendar life of electric vehicle batteries, reducing the extraction demand for lithium and cobalt that currently constrains sustainable EV adoption.

Cement production causes roughly 8% of global CO₂ emissions. A consequential part of that is repair and replacement of degraded structures. Self-healing materials that double maintenance intervals could cut that demand substantially.

What it would take to build differently

Self-healing materials help structures do their job for longer, with less human intervention. The bridge that arrests its own crack growth reduces the workload on engineers and maintenance teams whose schedules currently scale with every year of infrastructure age. The pipeline that seals its own pinhole leaks reduces the deployment of crews to locate and repair them. The scale of what needs doing is large, and the technology for at least some of the most important use cases has cleared its proof-of-concept stage.

Whether self-healing materials extend what infrastructure can do — allowing structures to be built in environments or to lifespans that would otherwise be impractical — or whether they are primarily cost-reduction tools for existing practice, the honest answer is both. A marine structure with self-healing concrete can be specified for a longer service life in a more aggressive environment than one without it. A pipeline coating that repairs micro-abrasions can be run in conditions that would quickly degrade conventional coatings. These are real expansions of what is buildable. But much of the near-term value will come from the more straightforward task of making ordinary infrastructure last longer and cost less to maintain.

What the industry has not yet seriously addressed is whether self-healing systems should be deployed uniformly at all. At present, they are — the same bacteria concentration, the same capsule density, regardless of where the highest damage risk sits in a given structure. A bridge deck sees very different stress profiles across its span. A pipeline coating experiences different corrosion rates at different locations based on soil chemistry and moisture. A system that modulated its healing capacity to match the risk profile of the structure — higher concentration in the highest-risk zones, lower elsewhere — would do more with the same material. This is technically feasible in principle and not yet routinely implemented in practice.

What the procurement system doesn’t capture

1. Infrastructure procurement in most countries evaluates initial cost, not lifecycle cost — even where policy frameworks require whole-life assessment. Self-healing materials are more expensive upfront and cheaper over time. What would it take to change the procurement incentive, and who would need to make that change?

2. Corrosion and material degradation cost $2.5 trillion per year globally. That figure is distributed across thousands of individual failures, repair contracts, and maintenance budgets. Is there a version of infrastructure ownership — public, private, or hybrid — that better captures the value of materials that last longer?

3. The communities with the most degraded infrastructure are often the ones with the least capacity to pay for repairs. Self-healing materials are currently priced as a premium product. What would a deployment model look like that directed this technology toward the places with the highest need, rather than the highest willingness to pay?

4. Self-healing materials borrow the concept of biological healing and apply it to inert matter. The research at Delft started with a question: if bone can heal, why can’t concrete? What other properties of living systems — adaptation, response to damage, distributed sensing — are materials scientists currently trying to replicate, and which of those would you most want to see applied to infrastructure?

My opinion

Electric vehicles are heavier than the combustion cars they replace — sometimes by several hundred kilograms — and roads specified for the last generation of vehicle are already absorbing loads beyond their original design. The UK road network, in many places, looks like potholes held together by asphalt. That is not a funding problem in isolation. It is also a materials problem, and one that becomes more urgent with every year that the vehicle fleet changes. Self-healing materials are not a long-horizon technology. They exist, they work, and they are cheaper than the alternative of letting the infrastructure fall further behind.

The lifeguard station by the lake in the Netherlands is still standing. The self-healing concrete in its foundations has been doing its work, quietly, since it was built — sealing the small cracks before they became large ones, without any human action, without any inspection finding them first.

That is a plain description of a consequential idea. Infrastructure that degrades slowly and repairs itself is not just cheaper to maintain. It is safer, more resilient, and less carbon-intensive. The hard part is not the science. It is getting the science out of the demonstration project and into the specification.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

The town of Wylfa sits on the northern tip of Anglesey, a Welsh island connected to the mainland by two bridges and a reputation for nuclear power. The original Wylfa power station opened in 1971 and ran for more than four decades, supplying electricity to households and businesses across the region. When it closed in 2015, the local economy lost something difficult to replace — not just jobs, but a particular kind of work: skilled, well-paid, long-term.

Plans for a new nuclear plant at Wylfa have circled since before the original station was switched off — Horizon Nuclear Power, then Hitachi, then a succession of other investors, none of whom made it past the preliminary stages. In June 2025, something different: the UK government selected Rolls-Royce SMR as its preferred partner to build the country’s first small modular reactors. By April 2026, a commercial contract had been signed and Wylfa confirmed as the site for up to three SMR units. If the timeline holds, the grid connection comes in the mid-2030s.

That “if” is doing a lot of work. The nuclear industry has a long and painful history of timelines that do not hold, costs that do not stay contained, and promises that outlast the political will required to keep them. The people of Wylfa — and the communities near every proposed SMR site — have reason to ask hard questions before deciding whether this time is genuinely different.

The honest answer is: it might be. And that is a more interesting answer than either the enthusiasts or the sceptics want to give.

Why nuclear has a second chance

To understand what small modular reactors are trying to solve, it helps to understand what broke the large ones.

Nuclear power has three advantages over almost every other energy source that are not marginal. It produces virtually no carbon emissions during operation. It generates electricity around the clock, regardless of whether the wind blows or the sun shines. And its energy density is extraordinary — a single kilogram of uranium fuel contains roughly two million times the energy of a kilogram of coal. These are the reasons nuclear provides about a tenth of the world’s electricity despite decades of political difficulty and public mistrust.

The problem is construction. Large nuclear plants take fifteen years to build in Western countries. The most recent to complete in the United States, the Vogtle plant in Georgia, came in seven years late and roughly $22 billion over its original $14 billion budget. The Hinkley Point C plant currently under construction in Somerset — the UK’s first new nuclear station in a generation — has seen its estimated cost rise from £18 billion when contracts were signed in 2016 to £35 billion in equivalent terms, approaching £49 billion at current prices. These are not outliers. They are the pattern.

The causes are structural: one-off bespoke designs that require learning everything from scratch on each new site; an atrophied supply chain that lost its skills during the decades-long pause in construction; and regulatory frameworks designed around 1970s technology that treat every new application as if nothing has been learned since. The result is that nuclear power, which ought to be the most reliable weapon in the decarbonisation arsenal, has spent thirty years being too expensive and too slow to be deployable at scale.

Nuclear’s advantages are extraordinary. Its construction record in the West is, on the evidence, one of the most expensive and difficult in modern engineering.

Small modular reactors are an attempt to escape that trap by changing the nature of the construction problem. Rather than building a bespoke cathedral on each site, the idea is to build a reactor in a factory — standardised, modular, replicable — and ship it to where it’s needed. A conventional reactor might generate 1,000 to 1,600 megawatts of electricity. An SMR typically produces between 50 and 470 megawatts per unit. You can add more units as demand grows, and the first unit can begin generating revenue while subsequent ones are still being assembled.

The factory model matters for cost. Nuclear’s expense problem stems partly from the fact that large reactors have never been built in sufficient numbers to drive down costs through learning and repetition. The steel and concrete and specialist components are expensive partly because the people and processes that handle them do so rarely. A factory model — producing the same design, repeatedly, under controlled conditions — is how aviation and automotive manufacturing drove down costs over decades. It is, at least in theory, how nuclear might do the same.

Where the theory meets the track record

There is a problem with the factory model, and it is not a small one: it has not yet produced a nuclear reactor on time or on budget, anywhere in the world.

As of early 2026, three SMRs have been built globally — one in China, two in Russia — plus one still under construction in Argentina. None were built on time. Russia’s floating plant, the Akademik Lomonosov, began construction in 2007 and reached operation only in 2019, at roughly three times its original budget. Argentina’s CAREM reactor broke ground in 2014 with a planned completion of 2017 and was still incomplete as of 2026, construction halted twice by funding crises, with costs significantly above initial estimates. China’s pebble-bed reactor at Shidaowan took approximately eleven years from construction start to commercial operation. These are first-of-a-kind projects, and some overrun is expected when building genuinely new technology. But the pattern of delay holds across different countries, different designs, and different political contexts.

The most instructive cautionary tale is NuScale Power. The American company was the first to receive US Nuclear Regulatory Commission design certification for an SMR, achieving that milestone in 2022 — a genuine achievement after years of rigorous review. NuScale appeared to be on the cusp of proving that the new model worked. Then, in 2023, its flagship project — a planned plant in Idaho — was cancelled. The project began with an estimated cost of around $3.6 billion for a 720-megawatt plant. Before cancellation, with the design rescaled and costs revised, that figure had risen to $9.3 billion. The customers, a consortium of public utilities, could not make the economics work. NuScale has since received regulatory approval for an uprated 77-megawatt design and is pursuing new projects, but the Idaho cancellation cast a long shadow over the industry’s cost claims.

The sceptical case against SMRs is not that the physics is wrong — it is that the economics may be structurally difficult. Smaller reactors lose the cost advantages that come with scale: building one 1,000-megawatt plant is cheaper per unit of output than building four 250-megawatt ones, because you only need one set of site preparation, one control room, one regulatory approval process. A 2025 study by Kim and Macfarlane — the latter a former chair of the US Nuclear Regulatory Commission — found that most cost projections for SMRs assumed high-volume serial manufacturing that has never materialised in nuclear, and that more realistic estimates placed the cost of electricity well above $100 per megawatt-hour, in some cases significantly more. For context, onshore wind in the UK currently costs around £40–50 per megawatt-hour.

There is also the waste question. Some analyses suggest that SMRs could produce greater volumes of radioactive waste per unit of energy generated than large conventional reactors — a consequence of their smaller scale and, in some advanced designs, the use of different fuel cycles. This is contested by the industry, but it is not a settled matter, and it matters to any community being asked to accept a reactor nearby.

None of this means SMRs cannot work. It means the confident promises being made — by governments, by technology companies, by investors — are running ahead of the evidence. The International Energy Agency’s scenarios have commercial SMR deployment beginning around 2030 — meaningful on paper, but, in the context of a global electricity system measured in thousands of gigawatts, still a rounding error in the near term.

What hangs on getting this right

The stakes are high enough that the question of whether SMRs work is not abstract.

Renewable energy — wind and solar — has become extraordinarily cheap and is now the cheapest source of new electricity generation in most of the world. But it is intermittent. The sun does not always shine; the wind does not always blow. Storing renewable energy at grid scale, for days or weeks at a time, remains an unsolved problem at the volumes the transition requires. This is sometimes called the baseload problem: the need for power sources that generate electricity continuously, regardless of weather, to underpin a grid that cannot afford to go dark.

Nuclear is one of the very few zero-carbon technologies that addresses the baseload problem directly. It generates power around the clock, in any weather, for decades. Hydropower does the same, but its geography is fixed. Large-scale geothermal is similarly constrained. If the world is serious about reaching net zero while keeping the lights on reliably, the honest arithmetic suggests that some form of nuclear will be part of the answer — and that SMRs, if they can deliver on their promise, could be a significant part.

The tech sector has reached a version of this conclusion independently. Amazon has committed over $500 million to SMR development. Google has partnered with Kairos Power to bring 500 megawatts of SMR capacity online by 2035, with the first unit targeting 2030, to power its data centres. Microsoft has signed agreements in the same space. The irony is pointed: AI’s insatiable demand for always-on power is one of the forces driving investment in nuclear energy’s revival.

The geopolitical dimension adds a further layer. Russia and China have both invested heavily in nuclear technology as a strategic export — Russia through Rosatom, which has construction contracts in over twenty countries; China through its state-owned enterprises. The UK-US Atlantic Partnership for Advanced Nuclear Energy, signed in September 2025, explicitly includes joint safety assessments and a shared commitment to eliminate dependence on Russian nuclear fuel by 2028. The race to build credible Western SMR supply chains is not purely a climate story. It is also a question of whose technology runs the world’s energy infrastructure in the decades ahead.

Against this backdrop, the communities living near proposed SMR sites find themselves in a familiar position — asked to accept proximity to technology that carries real and perceived risks, on the promise of jobs and cheap energy, with a track record that gives them every reason to ask for guarantees that are difficult to give. Research published in Energy Research & Social Science in 2025 put it plainly: the burdens and benefits of nuclear transitions are rarely distributed fairly, and existing frameworks to address this are inadequate. Wylfa’s community has been promised a new nuclear future before. The question is whether this version of the promise is better-designed, or just better-marketed.

What the technology still owes

At its best, an SMR helps communities, grids and industries access clean, reliable power that they cannot get from renewables alone. The UK government’s £2.5 billion commitment, the US government’s backing of multiple designs, and the private investment pouring in from the tech sector all reflect a real assessment that this technology could help solve a real problem. That is not spin. It is the outcome of serious people looking hard at the arithmetic of net zero.

Whether it adds something genuinely new is a sharper question. The factory model is the central claim — that modular construction changes the economics of nuclear in the same way that modular construction changed the economics of consumer electronics and automotive manufacturing. The honest answer is that this claim is plausible but unproven at commercial scale. The first Western SMR to be built on time and on budget will be the most important data point in energy policy this decade.

The question most consistently overlooked is whether this technology responds to the specific circumstances of the communities and contexts it serves — to their energy needs, their employment profiles, their risk tolerances, their histories with the industry. The SMR industry’s governance frameworks, on the evidence of current proposals, are not yet doing this. Regulatory processes are improving — the US ADVANCE Act of 2024 directed the NRC to streamline its approach, and by the end of 2025 it had met 30 of its 36 planned deliverables. But the frameworks for distributing the benefits and managing the disruptions of deployment — particularly for the communities most directly affected — remain underdeveloped.

The people of Wylfa have skills, knowledge and a relationship with nuclear power that spans generations. They are not a passive audience for a technology decision made elsewhere. Whether the SMR programme being planned for their island treats them as participants or subjects will say something important about whether the industry has learned anything from its history.

My opinion

I have worked inside the nuclear industry and I know what it delivers when it works. I also know what it costs when it goes wrong — not just financially, but in the lives of the communities closest to it. That makes it impossible for me to land cleanly on either side of this argument.

The NIMBY response to nuclear is rational, not irrational. If you live near a proposed site, you are being asked to accept proximity to technology whose consequences, in the worst case, outlast your lifetime and your children’s. That is a different category of ask than living near a wind farm.

But the power needs are also real. The grid cannot be decarbonised on intermittent sources alone. If small modular reactors can be built reliably, at reasonable cost, with genuine engagement of the communities they affect — that is worth pursuing. The question is whether the industry can meet that bar. It has not yet, but that is not the same as saying it cannot.

We also do not yet know if there will be any long-term consequences of renewable generation - what happens to wind patterns that are distributed by wind farms? what happens to ocean currents when hydro wave generation removes some of the kinetic energy; what happens to the earth’s core temperature when sunlight is blocked from hitting the ground by solar, or by heat extraction? We’ve made serious investments in all these technologies without the evidence of consequences - we know the consequences for nuclear.

What I am certain of is that complacency is the wrong response in either direction. The technology needs to prove itself commercially. The governance needs to prove itself ethically. Those are not the same challenge, and the industry tends to treat the first as if it solves the second.

The questions Wylfa should ask

• The UK government has pledged £2.5 billion and named a preferred supplier. But the first SMR is a decade away from the grid. How much confidence is there that the political will, the budget and the regulatory process will all hold for ten years?

• If renewables are now the cheapest source of new power, and storage technology is improving rapidly, is it possible that SMRs solve a problem that will largely have resolved itself by the time they arrive? Or is the baseload need real and irreducible?

• The tech companies investing in SMRs are doing so to power data centres. If artificial intelligence is driving both the energy demand that makes nuclear necessary and the investment that makes nuclear viable, what does that say about who the energy system is being built for?

• If you lived in Wylfa — or near any proposed SMR site — what would you need to know, and who would you need to hear from, before deciding whether to support the project?

The phrase “too cheap to meter” was coined in 1954 by Lewis Strauss, the chairman of the US Atomic Energy Commission, describing what nuclear power might one day become. It has since become one of the most famous over-promises in the history of technology. Small modular reactors are not making that promise. They are making a more modest one: that nuclear power can be built faster and cheaper than it has been. That is a lower bar. It is also one the industry has not yet cleared.

Whether it clears it this decade matters more than almost anything else in energy policy.

Leave a comment

Sources & References

• Wylfa power station: UK Government, GOV.UK: “Wylfa closes after almost 45 years” — https://www.gov.uk/government/news/wylfa-closes-after-almost-45-years

• ANS Nuclear Newswire, 10 June 2025: “UK’s own Rolls-Royce wins SMR competition” — https://www.ans.org/news/2025-06-10/article-7102/uks-own-rollsroyce-wins-smr-competition/

• NucNet: “UK Picks Rolls-Royce For Domestic Small Modular Reactor Rollout” — https://www.nucnet.org/news/uk-picks-rolls-royce-for-domestic-small-modular-reactor-rollout-6-2-2025

• New Civil Engineer, 13 April 2026: “Rolls-Royce SMR secures Wylfa contract and £599M government loan” — https://www.newcivilengineer.com/latest/rolls-royce-smr-secures-wylfa-contract-and-599m-government-loan-13-04-2026/

• Neutron Bytes, 19 April 2026: “UK & Rolls Royce Sign Deal for Three SMRs at Wylfa” — https://neutronbytes.com/2026/04/19/uk-rolls-royce-sign-deal-for-three-smrs-at-wylfa/

• IEA, Nuclear Power page (nuclear generated 9% of global electricity in 2024): https://www.iea.org/energy-system/electricity/nuclear-power

• IEA, Global Energy Review 2025: https://www.iea.org/reports/global-energy-review-2025/electricity

• Utility Dive, 2026: “After 2 years, ratepayer pain and political fallout from Georgia’s nuclear plant Vogtle” — https://www.utilitydive.com/news/after-2-years-ratepayer-pain-political-fallout-georgia-nuclear-vogtle/817792/

• The Current GA, May 2026: “Two Years After Completion, Plant Vogtle Still Looms Over the Nuclear Debate” — https://thecurrentga.org/2026/05/12/two-years-after-completion-plant-vogtle-still-looms-over-the-nuclear-debate/

• EnergyTransition.org, April 2026: “The billion-dollar boondoggle: how Vogtle became the US’s monument to nuclear folly” — https://energytransition.org/2026/04/the-billion-dollar-boondoggle-how-vogtle-became-the-uss-monument-to-nuclear-folly/

• New Civil Engineer, February 2026: “Hinkley Point C’s cost climbs to £35bn with confirmation Unit 1 will power up in 2030” — https://www.newcivilengineer.com/latest/hinkley-point-cs-cost-climbs-to-35bn-with-confirmation-unit-1-will-power-up-in-2030-20-02-2026/

• Construction Wave, February 2026: “Hinkley Point C costs approach £49bn as project faces M&E delays” — https://constructionwave.co.uk/2026/02/23/hinkley-point-c-costs-approach-49bn-as-project-faces-me-delays/

• US Department of Energy: “NRC Certifies First U.S. Small Modular Reactor Design” — https://www.energy.gov/ne/articles/nrc-certifies-first-us-small-modular-reactor-design

• Utility Dive, 2023: “NuScale, UAMPS terminate small modular reactor project in Idaho” — https://www.utilitydive.com/news/nuscale-uamps-terminate-small-modular-nuclear-reactor-smr-project-idaho/699281/

• Boise State Public Radio, November 2023: “NuScale nuclear reactor project in Idaho canceled” — https://www.boisestatepublicradio.org/news/2023-11-10/idaho-small-nuclear-reactor-project-canceled

• E&E News: “NuScale cancels first-of-a-kind nuclear project as costs surge” — https://www.eenews.net/articles/nuscale-cancels-first-of-a-kind-nuclear-project-as-costs-surge/

• US Department of Energy: “NRC Approves NuScale Power’s Uprated Small Modular Reactor Design” — https://www.energy.gov/ne/articles/nrc-approves-nuscale-powers-uprated-small-modular-reactor-design

• World Nuclear News: “China’s demonstration HTR-PM enters commercial operation” — https://www.world-nuclear-news.org/Articles/Chinese-HTR-PM-Demo-begins-commercial-operation

• NextBigFuture, December 2023: “China’s Pebble Bed Reactor Finally Starts Commercial Operation” — https://www.nextbigfuture.com/2023/12/chinas-pebble-bed-reactor-finally-starts-commercial-operation.html

• Wikipedia: Akademik Lomonosov —https://en.wikipedia.org/wiki/Akademik_Lomonosov

• Bellona.org, 2015: “New documents show cost of Russian floating nuclear power plant skyrockets” — https://bellona.org/news/nuclear-issues/2015-05-new-documents-show-cost-russian-nuclear-power-plant-skyrockets

• Power Magazine: “Russia Sees Floating Nuclear Power Plant Costs Balloon” — https://www.powermag.com/russia-sees-floating-power-plant-costs-balloon/

• Wikipedia: CAREM — https://en.wikipedia.org/wiki/CAREM

• Nuclear Engineering International: “Argentina’s CAREM-25 SMR faces setbacks” — https://www.neimagazine.com/news/argentinas-carem-25-smr-faces-setbacks/

• Buenos Aires Herald: “Construction of first Argentine-made nuclear power reactor halted amid layoffs” — https://buenosairesherald.com/business/construction-of-first-argentine-made-nuclear-reactor-halted-amid-layoffs

• ScienceDirect, Progress in Nuclear Energy, 2025: “Challenges of small modular reactors: A comprehensive exploration of economic and waste uncertainties associated with U.S. small modular reactor designs” — https://www.sciencedirect.com/science/article/abs/pii/S0149197025003877

• CleanTechnica summary, September 2025: “Small Modular Reactors and the Big Questions of Cost & Waste” — https://cleantechnica.com/2025/09/10/small-modular-reactors-and-the-big-questions-of-cost-waste/

• UK Government / BEIS, 2024: “Onshore Wind and Solar PV: Cost of Electricity Report Update 2024” — https://assets.publishing.service.gov.uk/media/68ba91f411b4ded2da19fe92/onshore-wind-and-solar-pv-cost-electricity-report-update-2024.pdf

• IEA: “The Path to a New Era for Nuclear Energy” — https://www.iea.org/reports/the-path-to-a-new-era-for-nuclear-energy

• IEA data chart: “Small modular reactor global installed capacity by scenario and case, 2025–2050” — https://www.iea.org/data-and-statistics/charts/small-modular-reactor-global-installed-capacity-by-scenario-and-case-2025-2050

• ESG Today: “Rolls-Royce Signs Deal with UK to for First Fleet of Small Modular Nuclear Reactors” (references Amazon/X-energy investment) — https://www.esgtoday.com/rolls-royce-signs-deal-with-uk-to-deliver-fleet-of-small-modular-nuclear-reactors/

• Google/Kairos Power: “Google and Kairos Power partner to develop nuclear energy” — search Google’s official press release at blog.google for the specific announcement.

• GOV.UK: “Golden age of nuclear delivers UK-US deal on energy security” — https://www.gov.uk/government/news/golden-age-of-nuclear-delivers-uk-us-deal-on-energy-security

• Al Jazeera, 18 September 2025: “US and UK sign major nuclear power deal: What does it include?” — https://www.aljazeera.com/news/2025/9/18/us-and-uk-sign-major-nuclear-power-deal-what-does-it-include

• Nuclear Innovation Alliance: “Regulatory Implementation Summary: NRC Progress Under the ADVANCE Act”(December 2025) — https://nuclearinnovationalliance.org/regulatory-implementation-summary-nrc-progress-under-advance-act

• NRC ADVANCE Act dashboard: https://www.nrc.gov/about-nrc/governing-laws/advance-act/dashboard

• NRC History: “Too Cheap to Meter: A History of the Phrase” — https://www.nrc.gov/reading-rm/basic-ref/students/history-101/too-cheap-to-meter

• Wikipedia: Too cheap to meter — https://en.wikipedia.org/wiki/Too_cheap_to_meter

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

A pathologist examining a breast cancer biopsy can read a great deal from the slide. The cells — stained pink and purple — arrange themselves in patterns that decades of training make legible. Some areas are dense with malignant cells: the tumour’s centre. Others show the advancing edge, where cancer is pushing into healthy tissue. Some regions are infiltrated by immune cells; others appear almost entirely free of them.

What the traditional biopsy slide cannot show is what those cells are actually doing — which genes each one is expressing, what proteins it is making, what signals it is sending to its neighbours, why the immune cells in one region are actively attacking the tumour while the immune cells in another region appear to have gone quiet. The stain shows structure. It cannot show function.

For most of the history of molecular biology, researchers addressed this limitation by doing what they could: taking the tissue, grinding it up, and measuring the average gene expression of the resulting mixture. This approach called bulk sequencing produces useful information. It also, inevitably, hides the detail that matters. A tumour cell that is resistant to a drug and a tumour cell that is sensitive to the same drug look identical in a bulk sample. Their individual signals cancel each other out. What the measurement captures is the average, and in a heterogeneous tissue, the average misleads.

Single-cell and spatial biology exist to fix that problem. Single-cell technologies analyse gene expression at the resolution of individual cells. Spatial technologies add location: they preserve the tissue architecture and record not just what each cell is doing but where in the tissue it sits. Together, they produce something that did not exist before — a map of a living tissue that shows, at microscopic resolution, which genes are active in which cells in which locations, and what those cells are saying to each other.

Nature Methods named spatially resolved transcriptomics its Method of the Year in 2020. Since then, the technology has moved from the leading edge of basic research to commercial platforms available in well-equipped laboratories worldwide. In February 2025, Illumina announced a spatial technology product planned for commercial release in 2026 with a capture area nine times larger than existing technologies and four times greater resolution. The field has tipped from pioneering to mainstream. What that means for understanding disease — and ultimately for treating it — is the subject of this article.

The tumour is not the problem. The neighbourhood is.

The central insight that spatial biology has delivered to cancer research can be stated simply: a tumour is not a mass of identical malignant cells. It is an organised community of different cell types — cancer cells in various states, immune cells of multiple kinds, structural cells called fibroblasts, blood vessel cells, and others — all occupying distinct spatial niches within the tissue, interacting constantly with their immediate neighbours, and behaving very differently depending on where they are.

This was suspected for a long time. What spatial biology has done is make it visible and measurable.

One of the most striking findings from spatial transcriptomics research is that the same tumour can simultaneously contain cells that are highly sensitive to a treatment and cells that are highly resistant — and that this is not purely a matter of genetics. Location matters. A cancer cell sitting in the centre of a tumour, surrounded by other cancer cells and relatively few immune cells, behaves differently from a genetically identical cancer cell at the tumour’s leading edge, where it is in constant contact with immune signals and structural proteins from the surrounding healthy tissue. The edge cell may be more aggressive and invasive; the centre cell may be more metabolically active but less mobile. Their gene expression profiles differ. Their drug sensitivities differ. The treatment that kills one may leave the other intact.

A tumour contains cells that are sensitive to a drug and cells resistant to the same drug — not just because of genetics, but because of location. Cells at the leading edge behave differently from cells at the centre. Spatial biology makes this visible.

A 2023 paper in Nature Communications on oral squamous cell carcinoma demonstrated that the gene expression profile of cells at the tumour’s leading edge — what researchers call the leading edge signature — was consistently associated with worse clinical outcomes across multiple cancer types, while the tumour core signature was associated with improved prognosis. The spatial position of the cancer cells relative to the boundary between tumour and healthy tissue carries predictive information that no amount of bulk sequencing could have extracted, because bulk sequencing cannot preserve the information about where each cell was sitting.

Immune cells show equally striking spatial organisation. Immunotherapy — the class of treatments that effectively unlocks the immune system to fight cancer — works by removing the signals that cancer cells use to suppress immune activity. It works well for some patients and not at all for others. Spatial biology research is revealing why. In patients who respond to immunotherapy, spatial analysis shows specific arrangements of immune cells clustered in particular locations around the tumour — what researchers call immune hubs or tertiary lymphoid structures. In patients who do not respond, those structures are absent or differently arranged. The presence or absence of the structures is not detectable by any analysis that does not preserve the spatial information. A blood sample, a bulk biopsy, even conventional single-cell sequencing without location data cannot tell you whether the right immune cells are in the right places.

Spatial transcriptomics of small cell lung cancer tumours has identified two distinct subtypes of cancer cells within the same tumours: one associated with treatment resistance and high proliferative activity, the other associated with treatment sensitivity and greater immune cell contact. The resistant subtype appears to reshape its local environment to drive away immune cells — to domesticate the macrophages in its vicinity and convert them into accomplices in suppressing the immune response. This finding was invisible to conventional analysis. It required spatial mapping of the tumour to become apparent.

The platforms are ready. The clinical pipeline is not.

The commercial field for spatial biology has developed quickly and is now genuinely competitive. 10x Genomics offers the Xenium In Situ platform, which can detect up to 5,000 genes across tissue samples at subcellular resolution. Bruker’s CosMx platform can profile over 19,000 genes — the entire protein-encoding transcriptome. Visium HD, also from 10x Genomics, resolves spatial data to 2-micrometre bins. Illumina’s forthcoming platform will use a capture area nine times larger than current technologies, integrated with its Connected Multiomics software suite for analysis.

The availability of these platforms has shifted the technology from something that required specialist physics laboratories to something that any well-resourced research institution can operate. The research output has followed. A review of spatial omics in cancer research published in Cancer Cell in January 2026, by researchers at the MD Anderson Cancer Center, described how spatial technologies now map tumour architecture with a fidelity that earlier methods could not achieve, resolving functional niches and spatial communities, converting spatial patterns into mechanistic insights.

The transition from research tool to clinical instrument, however, has several steps remaining. The data volumes generated by spatial transcriptomics are enormous. A single spatial experiment can produce hundreds of gigabytes of data. Analysing it requires computational infrastructure and bioinformatics expertise that most clinical settings do not yet have. The integration of spatial transcriptomics data with conventional pathology, with genomics, with clinical records, and with the kind of standardised workflows that regulatory agencies require before a diagnostic test can be approved — all of this is in progress but not yet complete.

The cost has been falling but remains significant. Running a spatial biology experiment on a clinical sample costs more than conventional sequencing and takes longer. For research settings, this is manageable. For routine clinical use — for the pathologist examining a cancer biopsy to guide treatment decisions for a patient who needs an answer within days — the workflow, cost and analytical requirements are still a barrier. The question of where spatial biology first enters clinical practice is the most practically important one in the field right now.

The most plausible near-term pathway is through what researchers call companion diagnostics: tests that are performed alongside a specific treatment to determine who is most likely to benefit. The patterns that spatial biology can detect — the presence or absence of immune structures around a tumour, the proportion of cells with certain gene expression signatures at the tumour edge, the spatial arrangement of drug-resistant subpopulations — are exactly the kind of information that could guide decisions about whether to use immunotherapy, which chemotherapy to start with, or whether a tumour is likely to respond to a newly approved targeted agent.

Why drug resistance looks different from here

Drug resistance is the central unsolved problem of cancer treatment. Most patients who are initially helped by a drug eventually stop responding to it. Understanding why, and predicting in advance who will stop responding and when, is the problem that determines whether a patient can be switched to an effective treatment before their cancer becomes untreatable.

Spatial biology is offering new ways to think about this problem. The conventional model of drug resistance focuses on genetics: a cell acquires a mutation that makes it insensitive to the drug, and that cell proliferates while the drug kills everything else. That model is real and important. But spatial biology research is demonstrating that it is incomplete. Resistance is also ecological.

A resistant cell does not only survive because of its own genetic properties. It also survives because of where it is in the tissue — who its neighbours are, what signals they are sending, whether the local immune environment is permissive or hostile. Research in melanoma has shown that distinct resistant fates — different ways a cell can become drug-resistant — are not randomly distributed through a tumour. They are spatially organised, associated with specific microenvironments. The cells in regions of necrosis behave differently from the cells adjacent to functional blood vessels. The resistant fate a cell adopts appears to be partly determined by its location.

This has immediate implications for treatment design. If resistance is partly spatial, then treating a tumour as a single uniform target with a drug calibrated to the average cell will predictably fail. The cells most resistant to treatment are often those in locations that the drug reaches least effectively and where the microenvironment most actively suppresses the immune system. A treatment strategy informed by spatial biology might target not just the cancer cells but their supporting environment: the macrophages they have domesticated, the fibroblasts that form barriers, the immune exclusion zones that protect the most resistant subpopulations.

If resistance is partly spatial, treating a tumour as a uniform target will predictably fail. The cells most resistant to treatment are often in the locations least accessible to drugs and most protected by their microenvironment.

Spatial analysis of breast tumours combined with pharmacogenomic profiling has identified that cancer cells in immunosuppressive microenvironments — the regions of the tumour where the immune system has been most effectively shut down — show specific sensitivity to cell cycle arrest agents and a class of drugs targeting the PI3K/AKT/mTOR pathway. This is information that could not be derived from bulk sequencing, because the immunosuppressive microenvironment and the cancer cells within it cannot be distinguished in a ground-up tissue sample. Seeing the map reveals what the average conceals.

Beyond cancer, spatial biology is beginning to reveal the tissue architecture of conditions that have been poorly understood precisely because they are diseases of organisation rather than diseases of individual cells. Pulmonary fibrosis — scarring of the lung tissue — involves molecular dysregulation at specific locations where epithelial cells are actively remodelling. Identifying exactly where in the tissue this remodelling is happening, and which cells are driving it, is exactly what high-resolution spatial transcriptomics can provide. Early data from the Illumina platform has shown that the scale and sensitivity of the technology enables the whole transcriptome to be studied across large tissue sections, identifying localised molecular dysregulation that earlier technologies could not resolve.

Who the insight reaches

Spatial biology is doing something specific for researchers and eventually clinicians: it makes visible biology that was previously invisible. Not incompletely understood — invisible. The question of whether two cancer cells in different parts of a tumour are behaving differently could not be asked until there was a technology that could preserve spatial information and measure gene expression simultaneously. Now it can be asked and answered. That matters for everyone whose work depends on understanding why treatments fail.

The distance between the research laboratory and the patient is where the real question sits. At present, the insights from spatial biology are largely flowing through academic research and pharmaceutical development. The direct beneficiary is the next generation of clinical trials — better-designed studies, better patient stratification, drugs tested against tumours whose architecture has been mapped rather than averaged. That is real progress. But the pathway from spatial biology discovery to a changed treatment decision for a specific patient is still long, and the technology to shorten that pathway — affordable, fast, clinically integrated spatial diagnostics — is not yet routinely available.

Whether spatial biology becomes a tool for everyone or a research instrument for well-resourced institutions is the question that will determine its real value. The core insight — that a patient’s tumour has a spatial architecture that determines how it will respond to treatment — is as true for a patient at a regional cancer centre as for one at a major academic medical centre. The Illumina platform under development is explicitly designed to reduce cost and integrate with existing sequencing infrastructure, which suggests the field is at least trying to address the accessibility barrier. The data infrastructure required to analyse the results — the computational pipelines, the bioinformatics expertise, the clinical interpretation frameworks — presents a more persistent challenge.

What the map cannot yet tell us

1. Spatial biology reveals that the same tumour can simultaneously contain cells that respond to a treatment and cells that resist it — and that location, not just genetics, partly determines which is which. How does that change how you think about the phrase ‘this treatment works for this cancer’?

2. If the spatial architecture of a tumour predicts whether immunotherapy will work, and that spatial information can only be obtained by analysing the biopsy tissue before treatment starts — who should have access to that analysis? What would it mean for cancer equity if spatial diagnostics are only available at specialist centres?

3. Spatial biology generates enormous datasets from individual tissue samples. Making sense of that data requires AI tools, bioinformaticians, and infrastructure that most clinical settings do not currently have. What is the bottleneck that most needs addressing before this technology reaches routine clinical use?

4. The insight that drug resistance is partly ecological — shaped by the local environment around cancer cells rather than only by the cells’ own genetics — suggests that some treatments may need to target the microenvironment, not just the tumour. What does that mean for how cancer drugs are designed and tested?

My Opinion

The gap between what spatial diagnostics can reveal and where that capability is available is not a technical problem. It is the same problem that follows every powerful diagnostic technology: it arrives at the best-resourced centres first and, without deliberate intervention, tends to stay there. Cancer does not follow a map of specialist centres. It arrives wherever it arrives — in rural hospitals, in regional oncology units, with patients who cannot travel to the institutions that define the frontier. The field needs to be working towards diagnostics that are portable, affordable, and deployable anywhere, not because that would be admirable, but because the technology’s actual value depends on it.

The pathologist looking at the traditional biopsy slide can see structure. Spatial biology lets you see function. It lets you read which cells are talking to which other cells, and what they are saying, and what happens in the neighbourhood where the immune system has given up.

The dimension is location. And it turns out to matter enormously.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

In the spring of 2023, a patient at Memorial Sloan Kettering Cancer Center in New York received an injection that had been made specifically for them. The sequence of molecules inside it had been designed using a biopsy of their own tumour. No one else in the world would receive exactly this treatment. The vaccine’s job was not to prevent a disease — it was to teach the immune system to recognise and destroy cancer cells that were already there.

The patient had pancreatic cancer. Pancreatic cancer kills around 87% of the people diagnosed with it within five years. It is resistant to most forms of treatment. The immune system usually does not recognise it as a threat. The mRNA vaccine was attempting to change that.

Three years later, the patient had not relapsed. In the trial they had been part of, vaccine-induced immune responses were still detectable in the blood — a signal that the immune system was still, years on, watching for the cancer to return. This is not a cure. It is not even yet a proven treatment outside of trials. But it is the kind of result that changes what researchers believe is possible.

The technology making it possible is the same one that protected hundreds of millions of people against COVID-19. mRNA — messenger ribonucleic acid — carries genetic instructions into cells, tells them to produce a specific protein, and then disappears. For infectious disease, the instruction was: produce a fragment of the virus’s spike protein, and let the immune system learn to recognise it. For cancer, the instruction is more personal: produce antigens that match the specific mutations in this patient’s tumour, so the immune system learns to hunt down cells carrying those mutations. The concept is the same.

Beyond COVID: what mRNA can and cannot do

The COVID vaccines demonstrated something that decades of laboratory research had suggested but never proved at scale: mRNA can be manufactured rapidly, delivered safely into human cells, and used to produce a protein the body has never encountered before. Before 2020, no mRNA medicine had ever been approved for human use. By the end of 2021, hundreds of millions of doses had been administered. The speed was the point — mRNA therapies can be designed in days once the target sequence is known, compared with the months or years required to develop traditional protein-based vaccines.

The platform is now being turned on cancer, autoimmune diseases, and genetic conditions where the body produces too little or too much of a particular protein. The logic is the same in each case: if you can specify the instruction, you can produce the protein.

Cancer is the most advanced frontier. There are now more than 120 clinical trials of mRNA cancer vaccines currently under way, targeting melanoma, lung, colorectal, pancreatic, brain, prostate and breast cancers, among others. The most significant results to date come from melanoma. A personalised mRNA vaccine called mRNA-4157, developed by Moderna in collaboration with the immunotherapy drug pembrolizumab, reduced the risk of recurrence or death by 44% compared to pembrolizumab alone in a Phase 2 trial. Patients given the combination were less likely to see their cancer return at five years than those who received the immunotherapy drug on its own. Phase 3 trials are now enrolling globally.

mRNA carries instructions into cells. For COVID, the instruction was: learn to recognise a virus. For cancer, the instruction is: learn to hunt down cells carrying these specific mutations.

The pancreatic cancer results are more tentative but potentially more significant, given how few treatment options exist. BioNTech’s personalised vaccine, autogene cevumeran, in combination with chemotherapy and immunotherapy, produced durable immune responses in pancreatic cancer patients — and in follow-up data published in 2025 and 2026, seven of the eight patients whose immune systems responded to the vaccine were still alive four to six years after surgery. Of the eight who did not respond, only two were. Pancreatic cancer’s resistance to immunotherapy had long been assumed to be near-absolute. These results suggest it may be conditional.

Beyond cancer, mRNA is in clinical trials for influenza, HIV, tuberculosis, and a range of rare genetic conditions where specific proteins are absent or dysfunctional. The questions are about speed, cost, and who gets access.

The personalisation problem

The most powerful version of mRNA cancer therapy is also its most limiting feature. A personalised vaccine — one designed around the specific mutations in a single patient’s tumour — is, by definition, a product that cannot be manufactured at scale. It requires a biopsy, genomic sequencing to identify which mutations are present and likely to trigger an immune response, computational modelling to select the right antigens, and then manufacture of the mRNA sequence that encodes them. This has to happen quickly enough to be clinically relevant, which in practice means within weeks of surgery.

The current manufacturing timeline for a personalised mRNA vaccine runs between six and ten weeks from biopsy to treatment-ready product, depending on the programme, and researchers are working to compress it further. The logistics are not incidental to the clinical result. In the BioNTech pancreatic trial, tumour tissue removed in New York was shipped to manufacturing facilities in Germany. The time between tissue leaving the body and the vaccine being ready had to be minimised. This is not a process that transfers easily to a district general hospital in the English Midlands, or a regional cancer centre in West Africa.

Cost is the other constraint. Manufacturing a personalised mRNA cancer vaccine currently costs upwards of $100,000 per patient. That figure reflects the early-stage nature of the technology — the expectation is that automation and manufacturing improvements will bring it down — but it is the figure that exists today, not the figure that might exist in 2035. For a health system trying to decide where to invest limited resources, a treatment that costs more than most people in the world earn in a lifetime, and that has not yet completed Phase 3 trials, is not an immediate commissioning priority.

A personalised cancer vaccine is designed around a single patient’s tumour mutations. That is its power. It is also why manufacturing it at scale, at a cost that health systems can absorb, is the defining challenge.

There is a further complication that does not feature in most of the coverage of mRNA therapeutics, but is directly relevant to where the technology goes next. In the United States, federal investment in mRNA vaccine research — which underpins much of the academic pipeline — was significantly cut in 2025. The National Cancer Institute saw funding reductions of around 31% in the first three months of that year. Research programmes specifically developing mRNA cancer vaccines faced additional scrutiny from the Department of Health and Human Services. The political pressure on mRNA technology in the US, driven by anti-vaccine sentiment that predated and outlasted COVID, is a genuine threat to the public research infrastructure on which the field depends.

The UK has moved in the opposite direction. In August 2025, the government announced £30 million for a UK RNA Biofoundry specifically to accelerate RNA therapy development. The Cancer Vaccine Launch Pad, a collaboration between the NHS and BioNTech, is fast-tracking thousands of patients into mRNA cancer vaccine trials targeting lung, breast, head and neck, and other cancers. The global market for mRNA cancer therapies is projected to reach $30 billion by 2033. The investment is arriving. Whether it arrives equitably — or concentrates in the health systems that can most readily afford it — is the question that will define the technology’s impact.

The distance between discovery and access

The patient at Memorial Sloan Kettering who received the pancreatic cancer vaccine in 2023 lived close to one of the world’s best-funded cancer research institutions, in a country with a health insurance system that covers experimental treatments for some patients in some circumstances, and was enrolled in a clinical trial with costs covered by the research programme. These were not incidental conditions. They were why that patient was treated and not someone else.

Pancreatic cancer does not concentrate in wealthy neighbourhoods. Neither does melanoma, lung cancer, glioblastoma, or any of the other cancers now being targeted by mRNA trials. But the trials are running almost exclusively in high-income countries, at major academic medical centres, with manufacturing infrastructure that does not exist in the places where cancer burden is rising fastest. By 2050, global cancer incidence is projected to reach 35 million cases per year — a significant portion of that growth driven by population ageing in low and middle-income countries. The treatments being developed now may arrive in those countries decades after they become available in New York or London, if they arrive at all.

This is not unique to mRNA. It is the pattern of almost every major therapeutic innovation. But mRNA was supposed to be different — its manufacturing speed and relative simplicity were explicitly cited, during the pandemic, as reasons to believe it could close the gap between rich and poor countries in access to medicines. That aspiration now needs to be tested against the reality of personalised cancer treatment, where the manufacturing process is the opposite of simple, the cost per patient is the opposite of affordable, and the regulatory frameworks governing it are only beginning to be developed.

My Opinion

The field’s working assumption is that personalised mRNA vaccines will become affordable as manufacturing scales. I think that is optimistic in a way that lets the field avoid a harder question.

A vaccine designed around one patient’s specific tumour mutations is not a product that scales the way a platform does. The personalisation is the therapy. You cannot remove it without removing what makes it work.

If the goal is equitable treatment, the investment needs to run in two directions simultaneously: into making personalised approaches cheaper and faster for the patients who can access specialist centres, and into shared-antigen mRNA approaches that target common mutations across large patient populations. The second track is less scientifically compelling. It is more likely to reach the person in Lagos.

The field rarely discusses this openly. The personalised vaccine is the story. Population-level mRNA is not. That asymmetry in attention is worth examining.

What the technology can and cannot yet do

mRNA cancer therapy helps the body do something it already knows how to do — mount an immune response — but has failed to do against a tumour that has learned to hide. The technology does not replace the immune system. It gives it better information. That is precisely the kind of assistance that medicine has sought for decades in treating cancers that resist other approaches. The question is not whether it works in the biology. The question is whether it works for the people who need it.

A melanoma patient in Manchester enrolled in a Cancer Vaccine Launch Pad trial is being given a capability their immune system would not otherwise have had. A melanoma patient in Lagos is not. The science adds something genuinely new. The system around it does not yet extend that addition to the people most likely to be without other options.

Personalised mRNA vaccines are, in the narrow technical sense, the most individual cancer treatment ever designed — each one built for one person, from that person’s own tumour biology. The treatment is exquisitely adapted to the biology. It is not yet adapted to the world.

The questions the science leaves open

The biology is not the hardest part. These are not questions about whether the technology works — the early evidence says it does. They are questions about investment, access, geography, and design: about who makes the decisions that will determine whether the science reaches everyone it could reach, or only the people who were already closest to it.

• The UK Cancer Vaccine Launch Pad is enrolling NHS patients in mRNA cancer trials right now. If you or someone you know has been affected by one of the cancers being targeted — melanoma, lung, breast, head and neck — do you know how to find out whether a trial is relevant?

• The manufacturing cost of a personalised mRNA vaccine is currently around $100,000 per patient. What would it take — in automation, in manufacturing scale, in public investment — to bring that number to one that health systems in lower-income countries could absorb?

• The US has cut federal investment in mRNA research. The UK has increased it. What does it say about the future geography of cancer treatment that the direction of public investment in the science varies so sharply between countries?

• The patient in New York survived because they were in the right place, enrolled in the right trial, at the right time. What would it take to design a version of this technology — and the healthcare system around it — where that sentence did not define who gets treated?

The mRNA platform can encode almost any protein the body can produce. The immune system can, it turns out, be taught to recognise almost any mutation cancer can generate. These two facts together are the reason that researchers who have spent their careers treating pancreatic cancer — a disease long described as an almost certain death sentence — are now using words like ‘paradigm shift’.

The science is moving faster than most of the systems around it. The history of medical innovation suggests the systems do not catch up on their own.

Leave a comment

Sources and references

• Pancreatic cancer survival rate: Pancreatic Cancer Action Network press release, 2025: five-year relative survival rate stalls at 13%. American Cancer Society, Survival Rates for Pancreatic Cancer, 2025. cancer.org

• mRNA-4157 melanoma trial (KEYNOTE-942): Individualised neoantigen therapy mRNA-4157 (V940) plus pembrolizumab versus pembrolizumab monotherapy in resected melanoma: a randomised, phase 2b study. The Lancet, 2023. Updated three-year results presented at ASCO, 2025. Five-year recurrence-free survival data: CancerNetwork, 2025.

• BioNTech autogene cevumeran — pancreatic cancer: Three-Year Phase 1 Follow-Up Data, BioNTech press release, 2025. Updated follow-up (4–6 years): Memorial Sloan Kettering Cancer Center news release; reported in CNN Health, April 2026. AACR Annual Meeting 2025: immune response correlates with clinical benefit in pancreatic cancer patients.

• US federal cancer research funding cuts: US Senate Minority Staff report, May 2025: federal cancer research funding reduced by approximately 31% in the first three months of 2025. Reported in Axios, PBS NewsHour, OncLive (May 2025).

• UK RNA Biofoundry: GOV.UK press release, 28 August 2025: £29.6 million investment in UK RNA Biofoundry at CPI’s RNA Centre of Excellence, Darlington. gov.uk

• Global cancer incidence projections to 2050: IARC / GLOBOCAN 2024 projections: global cancer incidence projected to reach 35 million cases annually by 2050, a 77% increase from 2022. Reported by UN News, February 2024. (Note: the Global Burden of Disease Study 2023 uses a lower estimate of 30.5 million; the IARC figure is the more widely cited projection.)

• mRNA cancer vaccine market projection ($30 billion by 2033): Noted in commentary published in The Lancet Oncology, 2025, drawing on market research projections. The underlying figure originates from commercial market research.

• Manufacturing cost and timeline: Manufacturing costs upwards of $100,000 per patient: reported across multiple clinical development sources including Scientific American (2025) and RNA cancer vaccine pipeline reviews (PMC, 2025). Timeline of 6–10 weeks from biopsy to treatment-ready product: Moderna manufacturing data; autogene cevumeran trial documentation showing approximately 9-week average for the BioNTech pancreatic cancer programme.

• 120+ clinical trials: Current Progress and Future Perspectives of RNA-Based Cancer Vaccines: A 2025 Update. Cancers (MDPI), May 2025. PMC12153701

• Cancer Vaccine Launch Pad: NHS England / BioNTech Cancer Vaccine Launch Pad: fast-tracking patients into mRNA cancer vaccine trials targeting melanoma, lung, breast, head and neck, and other cancers. nhs.uk

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

The pattern has been documented often enough that it has a name. An AI agent is given the instruction to find options — for a flight, a supplier, a candidate. It checks availability, compares prices, retrieves payment details from a linked account, and completes the task. A confirmation arrives before any human has made a decision. The outcome may be correct. The price reasonable. And yet a decision was made that no person made, and if it had gone wrong — wrong date, wrong airport, wrong account charged — the question of what to do next would be genuinely unclear.

This is the agentic AI moment. Not the science fiction version, where a machine plots against its creators. The mundane, already-here version, where AI systems that can act are being given access to the things that matter — calendars, inboxes, bank accounts, business systems — and the line between “help me with this” and “do this” is blurring faster than the accountability structures can keep up.

From chatbot to colleague

The shift from conversational AI to agentic AI is not incremental. It is a change in the nature of what the technology does.

A conversational AI — the kind most people encountered first — responds. You ask it something, it answers, the interaction ends. It can be wrong, it can be inconsistent, but it cannot do anything in the world beyond produce text. The consequences of its failures are bounded.

An agentic AI operates differently. Given a goal, it plans a sequence of steps to reach it, calls on tools — APIs, databases, software interfaces, payment systems — executes each step, monitors the results, and adjusts if something fails. It does not wait for instruction at each stage. It works. This is why the transition from 2023’s chatbots to 2025’s agents felt, to people paying attention, like something more than an upgrade. OpenAI released Operator, Anthropic extended Claude to computer use, Google launched Project Mariner — all in the same period, all enabling AI to interact with real systems on a user’s behalf. The architecture changed. The capability changed with it.

The adoption numbers reflect this. A Zapier survey of enterprise leaders in early 2026 found around 72% were using or testing agentic systems. Gartner forecast that 40% of enterprise applications would embed task-specific AI agents by the end of the year. The agentic AI market, valued at roughly $7.3 billion in 2025, is projected by Fortune Business Insights to reach $139 billion by 2034 — though other analysts project considerably higher figures. These are not pilot-project figures. This is live deployment across large enterprises.

The capability that makes agentic AI effective — the ability to act without waiting for permission at each step — is precisely what makes it difficult to govern.

The use cases making the business case are concrete enough. An AI agent that handles the full lifecycle of an insurance claim — reading the form, cross-referencing policy, assessing evidence, initiating payment — compresses a process that once took days into minutes. Walmart’s agentic supply chain system detects demand signals, adjusts procurement plans, and reroutes logistics without human triggers, with reported reductions in out-of-stock incidents across pilot regions. In software development, agents now write, test and deploy code: Stack Overflow’s 2025 developer survey found 84% of developers using AI tools, tools that now generate around 41% of all code written. Research published by Landbase, a sales automation firm, puts average returns from mature agentic deployments at 171% — roughly three times the return from traditional automation.

These are genuine improvements. They are also the easy half of the story.

The accountability gap

The hard half concerns what happens when the agent is wrong.

In 2024, a Canadian tribunal held Air Canada liable after its customer service chatbot gave a grieving passenger incorrect information about bereavement fare discounts. Air Canada’s defence — that the chatbot was a separate legal entity responsible for its own statements — was rejected without hesitation. The airline was responsible. The principle was unambiguous: organisations are accountable for what their AI systems say and do. The machine cannot be sued. The humans behind it can.

That principle becomes considerably more complicated when the AI is not talking but acting. A wire transfer sent to the wrong account. A contract clause accepted on behalf of a business. A medical appointment cancelled because an agent misread a scheduling instruction. In each case, the action has real-world consequences that may be difficult or impossible to reverse, the chain of decisions that led to it may involve multiple models, APIs and data sources, and the human who nominally authorised the deployment may have had no visibility into the specific step that went wrong.

The governance figures sit uncomfortably alongside the adoption figures. A 2025 industry survey found only 44% of organisations had formal AI governance policies in place. Only 11% of enterprises had agentic systems actually running in production rather than pilot. Deloitte found that 42% of organisations were still developing their agentic strategy, and 35% had no formal strategy at all. A Fortune report published in March 2026 — drawing on research from Wharton and Accenture — put the problem plainly: AI agents were spreading across the enterprise value chain, often ahead of formal strategy and governance.

The Wharton Accountable AI Lab framed it as a supply chain problem. Accountability is not a single organisation’s responsibility — it runs from the model developers who trained the system, through the platform providers who packaged it, to the organisations that deployed it, and finally to the humans who were supposed to be supervising it. When something goes wrong, everyone points to someone else. And in a system where a single agent action might involve five external data sources, three AI models, and four different APIs, the pointing is genuinely difficult to resolve.

The human in the loop — or not

The phrase the industry has settled on is “human in the loop” — the idea that a person remains involved in agentic decisions, available to check, correct or override. It is a reassuring phrase. In practice, it describes a spectrum, and much of the deployment happening now sits at the end where the human is less supervisor than emergency contact.

Microsoft has coined a different formulation: the “agent boss” — the person who builds agents, delegates to them, and manages them. The role is real and the framing is honest. But it also makes visible something the more comfortable language obscures. If one person is now the agent boss for a fleet of AI systems each capable of taking hundreds of actions per day, the ratio of human attention to autonomous action has shifted dramatically. The human is not in each loop. They are above all the loops, available to intervene if something surfaces — which requires that something surface, which requires that the systems are designed to surface it.

Fortune’s March 2026 report documented what it called “shadow AI” — the use of AI tools, including agentic systems, that employees bring into the workplace outside formal IT channels, beyond the visibility of any governance structure. Microsoft’s 2025 research found around three-quarters of knowledge workers were already using AI tools; a significant proportion were doing so with tools their employers had not sanctioned and could not audit. An orphaned agent — one running in production with no designated owner, accessing systems its original designers never intended — does not typically go rogue. It accumulates risk silently, its permissions unreviewed, its behaviour unchanged as the business context around it evolves.

The human is not in each loop. They are above all the loops — available to intervene if something surfaces. Which requires that something surface.

None of this is an argument against agentic AI. The productivity gains are real. The compression of tedious, error-prone, time-consuming processes is real. The question is not whether to use agents. It is whether the organisations deploying them are building accountability in from the start, or treating governance as something to address after the value has been captured. The evidence, on balance, suggests the latter is happening more often than the former.

The test most deployments are failing

Most current deployments pass the first question easily. The AI that clears an inbox backlog, processes a claims queue, or monitors a live system for failures is doing work that was real, time-consuming and often unrewarding. People who use these tools report genuine relief. That part is not in doubt.

The harder question is whether the agent adds something — capability, judgement, insight — or whether it substitutes for human understanding in ways that hollow out the understanding itself. There is a difference between an AI that surfaces relevant information so a person can make a better decision, and an AI whose recommendations are so compelling, and whose interface so frictionless, that human approval becomes a formality. A hospital AI diagnostic study cited in the Wharton and Accenture report found exactly this pattern: at high-performing sites, clinicians understood the system’s confidence scores and knew when to question their findings; at lower-performing sites, they either followed the tool uncritically or ignored it entirely. The tool was the same. The human relationship with it was not.

The third question — whether the agent responds to the specific person it serves, or treats every user as the average case — is the one most consistently failed. The documented pattern of agents interpreting “find” as “do” is a small example: the system optimised for task completion and treated the distinction as a rounding error. At the scale of enterprise deployment, that rounding error multiplies. The question of whether the people affected by agentic decisions — employees, customers, citizens — are treated as participants or as variables is not a technical question. It is a design choice.

My opinion

The accountability question underneath agentic AI has not been seriously answered yet, and I don’t think it will be until something goes badly wrong at a scale that cannot be managed quietly. The opportunity is real — the productivity gains are documented and the compression of genuinely tedious processes matters. But the responsible structure for deciding who owns the outcome when an agent acts on a person’s behalf has not been established. Was it the person who clicked through the terms? The company that deployed the agent? The vendor who built the platform? The model developer underneath all of it? Every layer points to the next. Right now, that costs individual people inconvenience and embarrassment. The price of that uncertainty will rise.

Who answers for it

The capability that makes agentic AI effective — the ability to act without waiting for permission at each step — is precisely what makes it difficult to govern. The technology is not the problem. The gap between how fast it is being deployed and how slowly the accountability structures are following it is.

The consequences of that gap are still emerging. Not all of them will be recoverable.

Leave a comment

Sources and references

• Agentic AI adoption — 72% of enterprise leaders: Zapier, State of Agentic AI Adoption Survey 2026. Survey of 500+ enterprise leaders. zapier.com/blog/ai-agents-survey/

• Gartner — 40% of enterprise applications to embed AI agents by end of 2026: Gartner press release, 26 August 2025: “Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025.” gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026

• Agentic AI market size — $7.3 billion (2025) projected to $139 billion (2034): Fortune Business Insights, Agentic AI Market Size, Share & Forecast Report 2026–2034. fortunebusinessinsights.com/agentic-ai-market-114233. Note: other analysts project higher 2034 figures (Precedence Research: $199 billion; Grand View Research: $183 billion by 2033).

• OpenAI Operator, Anthropic Claude computer use, Google Project Mariner: Released in late 2024 and early 2025. Confirmed via product announcements from OpenAI, Anthropic, and Google.

• Walmart agentic supply chain system: Walmart Global Tech blog; Supply Chain Dive, “4 ways Walmart is scaling AI to unify its supply chain,” 2025. tech.walmart.com; supplychaindive.com/news/4-walmart-supply-chain-ai-uses/760891/

• Developer AI tool usage — 84% of developers; 41% of code: Stack Overflow, 2025 Developer Survey. stackoverflow.com

• Returns on agentic AI deployments — 171% average ROI: Landbase, 39 Agentic AI Statistics Every GTM Leader Should Know in 2026. landbase.com/blog/agentic-ai-statistics. Note: Landbase is a commercial sales automation firm; this is industry research rather than independent analysis.

• Air Canada chatbot liability ruling: Moffatt v Air Canada, British Columbia Civil Resolution Tribunal, 14 February 2024. Reported by CBC News: cbc.ca/news/canada/british-columbia/air-canada-chatbot-lawsuit-1.7116416

• AI governance policies — 44% of organisations: Corroborated across multiple 2025 industry surveys including IAPP AI Governance Profession Report 2025 (iapp.org) and SailPoint enterprise security research.

• Only 11% of enterprises with agentic systems in production; Deloitte — 42% developing strategy, 35% no strategy: Deloitte, 2025 Emerging Technology Trends Survey (500 US technology leaders, June–July 2025). deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html

• Fortune / Wharton / Accenture report — March 2026: “Intelligence may be scalable, but accountability is not,” Fortune, 26 March 2026. Joint research from Accenture’s global products practice and the Wharton AI & Analytics Initiative. fortune.com/2026/03/26/ai-agents-accountability-accenture-wharton-report/

• Wharton Accountable AI Lab: Led by Professor Kevin Werbach, Department of Legal Studies and Business Ethics, Wharton School, University of Pennsylvania. ai-analytics.wharton.upenn.edu/wharton-accountable-ai-lab/

• Hospital AI diagnostic study — high- vs low-performing sites: Cited in the Wharton and Accenture report (Fortune, March 2026). See above.

• Shadow AI / knowledge workers using unsanctioned AI tools: Microsoft, 2025 Work Trend Index. Microsoft’s research found approximately 75% of knowledge workers using AI tools at work, with a significant proportion using tools outside formal IT channels. microsoft.com/en-us/worklab/work-trend-index

• Microsoft “agent boss” framing: Microsoft Copilot product communications and AI documentation, 2025–2026. microsoft.com/en-us/microsoft-copilot

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

In September 2023, as Hurricane Lee was curving northward west of Bermuda, meteorologists at NOAA and the National Hurricane Center were working to determine where it would make landfall: New England or farther east, in Canada. The sooner they could call it, the earlier they could issue warnings. Evacuations for a major hurricane require at least three full days. Every additional day of warning is not a planning luxury. It is a measurable difference in how many people can get out.

The traditional forecast models locked in on Nova Scotia about six days before landfall. A different model — an experimental AI system called GraphCast, developed by Google DeepMind and running as a live test on the European Centre for Medium-Range Weather Forecasts website — had already called it three days earlier. Nine days out, GraphCast was pointing to Nova Scotia with the same confidence that conventional models would not reach for another 72 hours.

Three days. In the context of hurricane preparedness, that is not a technical curiosity. It is the difference between an evacuation that works and one that does not. It is the difference between a ferry that leaves the island while the roads are still passable and one that does not. It is, in the most direct sense, lives.

This is the AI weather forecasting story — not the version about supercomputers and equations, but the version about what happens in the gap between when a forecast becomes reliable and when people have enough time to act. That gap is shrinking. The implications run further than most discussions of AI and climate have yet acknowledged.

How weather prediction works, and why it is hard

Traditional weather forecasting is an extraordinary intellectual achievement. Starting in the early twentieth century, researchers proposed that the behaviour of the atmosphere could be modelled mathematically — that the physical equations governing fluid dynamics, thermodynamics and radiation could be translated into algorithms, fed with observations from weather stations, balloons, satellites and ocean buoys, and used to project how the atmosphere would evolve over days ahead. They were right. By the 1960s, when computers became powerful enough to run these models, numerical weather prediction began to demonstrate real skill. It has been improving ever since at a rate of roughly one useful day per decade — today’s six-day forecast is as accurate as the five-day forecast was ten years ago.

These traditional models are based on physics. They represent the atmosphere as a three-dimensional grid, apply equations at each grid point, and propagate the simulation forward in time. The equations are derived from first principles — conservation of mass, energy and momentum. When the model produces a prediction, you can trace exactly why: this air mass moved because of that pressure gradient, that front formed because of these temperature differences. The physics provides both the accuracy and the interpretability.

The cost is computation. Running a ten-day global forecast using the European Centre for Medium-Range Weather Forecasts’ flagship model — widely regarded as the world’s best — requires hours of runtime on one of the largest supercomputers in the world, consuming enormous amounts of energy. The ECMWF’s computing infrastructure costs hundreds of millions of pounds to build and operate. Most countries cannot afford anything comparable, which means most of the world’s weather forecasting depends on a handful of institutions with sufficient resources to run the physics.

GraphCast produces a ten-day global forecast in under a minute on a single machine. The equivalent physics-based forecast requires hours on one of the world’s largest supercomputers. The cost gap is measured in orders of magnitude.

AI weather models work differently. Rather than encoding the physics of the atmosphere from first principles, they learn from data. GraphCast was trained on four decades of ECMWF reanalysis data — a historical reconstruction of global weather conditions combining observational records with traditional models to produce a consistent dataset — and learned to predict how weather states evolve by recognising patterns in that history. It does not calculate what the atmosphere will do based on physics. It recognises what atmospheric configurations have historically led to what subsequent states, and applies that learned relationship to new initial conditions.

The result, in terms of accuracy, is remarkable. A paper published in Science in November 2023 showed that GraphCast outperformed ECMWF’s best deterministic model on more than 90% of 1,380 verification targets. Huawei’s Pangu-Weather, published in Nature in 2023, achieved comparable results and runs 10,000 times faster than traditional ensemble models. ECMWF moved its own AI-based model — AIFS, the Artificial Intelligence Forecasting System — to operational status in February 2025, making it the first major meteorological agency to deploy an AI model as part of its official forecast suite. NOAA launched its own suite of AI-driven global weather models in late 2025, built on GraphCast foundations and fine-tuned with NOAA’s own data, requiring up to 99.7% fewer computing resources than their traditional counterparts.

What AI can and cannot do

The models are as notable for what they cannot do as for what they can.

The models excel at medium-range forecasting — the three-to-ten-day window where most consequential weather decisions are made. They struggle with extreme events that fall outside their training data. This is not a minor limitation. Hurricane Otis, which intensified from a tropical storm to a Category 5 hurricane in roughly twelve hours before striking Acapulco in October 2023 — killing more than fifty people and destroying most of the city’s infrastructure — is precisely the kind of event that AI models trained on historical patterns may fail to capture. Rapid intensification over warm ocean water, in conditions that push against the edge of what the historical record contains, is where pattern recognition may fail and physics-based modelling may have an irreplaceable advantage.

The AI models also depend entirely on traditional physics-based systems for their input data. They cannot generate a forecast from raw observations alone — they require a processed, quality-controlled atmospheric state as their starting point, which is produced by the same supercomputer-intensive data assimilation systems that underpin conventional forecasting. If those systems fail or degrade, AI forecasting degrades with them. The models are not independent of the infrastructure they appear to make redundant.

The most sophisticated agencies are already pursuing a hybrid approach that addresses both sides of this. NOAA’s HGEFS — the Hybrid Global Ensemble Forecast System — combines AI-based ensemble forecasts with traditional physics-based ensemble modelling. In early testing, this hybrid consistently outperformed both the AI-only and physics-only systems. The combination captures the AI’s computational efficiency and its skill at medium-range pattern recognition, while the physics-based component provides a check on situations where the atmospheric state is genuinely novel.

There is a further limitation that ECMWF’s own documentation is direct about: AI models do not yet explain their forecasts in the way that physics-based models can. A traditional model can trace the chain of physical causation that led to a prediction. An AI model can say what it predicts but cannot always say why in terms a forecaster can interrogate. When an AI forecast diverges from a physics-based forecast — when the two disagree significantly — that disagreement is valuable information about forecast uncertainty. But resolving the disagreement requires human expertise. The role of the meteorologist is shifting: from running models and reading their output, toward interpreting the ensemble of AI and physics forecasts, identifying where they diverge, and communicating uncertainty to the people who need to make decisions.

What changes when forecasts get earlier and cheaper

Emergency preparedness is the most direct consequence. Hurricane evacuation orders typically need to be issued three to four days in advance to be effective — that is the minimum time required to move a large population out of a coastal zone before a major storm arrives. A model that identifies the landfall region nine days out rather than six gives emergency managers an extra three days to coordinate, to communicate, to move people who cannot move themselves. AI models are extending that window.

Agriculture is the second major domain. Planting, irrigation, harvest and frost-protection decisions all depend on weather forecasts. Farmers in rain-fed agricultural systems — which account for the majority of global food production — manage risk primarily through weather information. A more accurate ten-day forecast, available to a farmer with a smartphone, changes the calculation for whether to plant this week or next, whether to apply irrigation before a forecast rain event or wait. Google DeepMind’s WeatherNext 2 now powers weather information across Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API. The agricultural implications of that scale of deployment in low-income farming communities have not yet been fully studied, but access to forecast information that was previously available only to well-resourced agricultural operations is now reaching smallholder farmers for the first time.

Energy system management is the third. Renewable energy generation is inherently variable — wind and solar production depend directly on weather conditions. Grid operators managing the balance between generation and demand need accurate forecasts of wind speeds and solar irradiance to plan how much backup capacity to hold in reserve and when to bring it online. GenCast, Google DeepMind’s probabilistic AI forecasting model, reduces wind power forecasting errors by up to 20% within a two-day lead time compared to traditional ensemble models. At grid scale, a 20% reduction in forecast error translates directly into lower costs for consumers and more efficient integration of renewable energy.

The implications for climate science itself are harder to measure but potentially the most significant of all. The same AI techniques that underpin operational weather forecasting are now being applied to climate modelling at longer timescales — predicting seasonal and sub-seasonal patterns months ahead. NOAA’s Atlantic Oceanographic and Meteorological Laboratory and its Weather Program Office’s Subseasonal to Seasonal Research programme are working toward extending extreme weather forecast lead times from the current two-to-four-day window to two-to-four weeks. If achievable, that shift would change disaster preparedness entirely. A community that knows six weeks in advance that conditions are likely to produce a severe hurricane season, or an anomalous flooding pattern, or an unusual drought, can prepare differently than one that gets four days’ notice.

A more accurate ten-day forecast, on a smartphone, changes the calculation for a farmer deciding whether to plant this week or next. That is not a marginal improvement. At the scale of global food production, it is consequential.

Traditional numerical weather prediction at global scale requires infrastructure that only a handful of institutions worldwide can afford. GraphCast-quality ten-day forecasts can run on a single laptop. NOAA’s AI models require a fraction of a percent of the computing resources of their traditional counterparts. The same countries that have historically depended on ECMWF or NOAA data for their weather services — because they cannot afford to run their own global models — can now run inference on AI models with modest infrastructure. This is not a complete solution to the observation network problem (AI models still need input data from the global observing system that wealthy countries have historically funded), but it opens a door to greater national meteorological capacity that was previously closed by computational cost.

The forecast that arrives and the one that doesn’t

AI weather forecasting helps people anticipate the future well enough to make better decisions, and the improvement is not marginal. Three extra days of hurricane warning, more accurate flood alerts for river communities downstream from atmospheric river events, earlier seasonal outlooks for farmers facing planting decisions under climate uncertainty — these are real improvements. The technology is operational. These improvements are already being delivered.

The question of who receives that improvement is harder. AI weather forecasts are available through Google Maps, through NOAA’s public systems, through the ECMWF website. The infrastructure to receive a forecast is a smartphone. The infrastructure to act on it — the evacuation route, the early warning system connected to the government emergency network, the insurance product that adjusts premiums based on forecast risk — is not universally available. The forecast that tells a coastal community in Bangladesh nine days in advance that a major cyclone is forming in the Bay of Bengal is only useful if there is a system in place to act on that information. The observation networks that produce accurate initial conditions for AI models are denser in wealthy countries than in poor ones. The forecast quality is not equal.

The problem that matters most is the one where AI models are least equipped. The climate system is changing. Atmospheric rivers are intensifying. Hurricanes are undergoing rapid intensification more frequently. The Arctic sea ice that provides boundary conditions for mid-latitude weather is disappearing. AI models trained on historical patterns may be systematically less reliable in the conditions that climate change is producing — precisely the conditions for which the most accurate forecasts matter most. The hybrid approach being pursued by NOAA and ECMWF — combining AI pattern recognition with physics-based modelling — is the appropriate response. But it requires sustained investment in both the AI systems and the physical science infrastructure they depend on. Cutting observation networks to fund AI forecasting would be a category error.

My Opinion

The hybrid approach at NOAA and ECMWF is correct — not because it hedges between old and new, but because the two systems are doing different jobs. AI recognises patterns in historical data at speed; physics-based models reason about conditions that appear nowhere in any training set, which in a changing climate are often the conditions that matter most. Cutting observation networks and physical science funding in the name of AI efficiency would be a category error: the AI models run on the observational infrastructure they appear to make redundant. A forecast that reaches a community with no evacuation route and no warning system is not better forecasting. It is better knowledge of what is coming with no way to act on it.

Questions for the people who have to act

The improvements in AI weather forecasting are already in the systems most organisations use. What follows are questions for the people who have to act on what those systems produce — emergency managers, grid operators, farmers, and policy makers deciding how much to trust a forecast they cannot interrogate.

1. Three extra days of hurricane warning can make the difference between an effective evacuation and an ineffective one. What other decisions — in emergency management, in agriculture, in energy — depend on the accuracy and lead time of weather forecasts in ways you haven’t previously thought about?

2. AI weather models are computationally cheap to run but still depend on expensive observation networks and traditional data assimilation systems for their input. If the political pressure to cut meteorological spending increases — as it has in the US with proposals to reduce NOAA’s capacity — what happens to the quality of AI forecasts that everyone now depends on?

3. The climate system is changing in ways that push beyond the patterns AI models were trained on. Rapid intensification, unprecedented rainfall events, Arctic conditions without historical precedent — these are exactly the situations where pattern-recognition AI may fail precisely when accurate forecasting matters most. How should society balance the efficiency gains from AI forecasting with the continued investment in physics-based modelling that handles genuinely novel situations?

4. AI weather forecasting is now delivered through consumer products — smartphone apps, mapping services, voice assistants. Most people receive forecasts without knowing whether they came from a physics-based model or an AI. Does that transparency matter? Should it?

Weather forecasting has been improving for more than a century, driven by better physics, better observations, and better computers. AI represents a different kind of improvement in that progression — not a replacement of what came before, but an acceleration that has opened global weather forecasting capacity previously closed by computational cost.

The people who most need accurate weather forecasts are not the researchers at ECMWF or the meteorologists at NOAA. They are the farmer deciding whether to plant, the emergency manager deciding when to order an evacuation, and the family on a coastline in the path of a hurricane trying to understand whether they have enough time.

Three days earlier matters. It has always mattered. The question is who has the infrastructure to act on it.

Leave a comment

Sources & references

• Hurricane Lee (2023) and GraphCast prediction: National Hurricane Center Tropical Cyclone Report — Hurricane Lee (AL132023). National Oceanic and Atmospheric Administration. https://www.nhc.noaa.gov/data/tcr/AL132023_Lee.pdf

• GraphCast performance benchmarks: Lam, R. et al. (2023). “Learning skillful medium-range global weather forecasting.” Science, 382(6677). doi: 10.1126/science.adi2336. Published November 2023. Confirms >90% of 1,380 verification targets outperformed ECMWF HRES.

• Pangu-Weather: Bi, K. et al. (2023). “Accurate medium-range global weather forecasting with 3D neural networks.” Nature, 619, 533–538. Published July 2023. doi: 10.1038/s41586-023-06185-3.

• ECMWF AIFS operational status: “ECMWF’s AI forecasts become operational.” ECMWF, February 2025. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational. AIFS version 1.0.0 implemented 25 February 2025.

• NOAA AI-driven weather models: “NOAA deploys new generation of AI-driven global weather models.” NOAA press release, 17 December 2025. https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models

• Hurricane Otis (2023): National Hurricane Center Tropical Cyclone Report — Hurricane Otis (EP182023). National Oceanic and Atmospheric Administration. https://www.nhc.noaa.gov/data/tcr/EP182023_Otis.pdf

• GenCast: Price, I. et al. (2024). “Probabilistic weather forecasting with machine learning.” Nature, 637, 84–90. doi: 10.1038/s41586-024-08252-9. Confirms ~20% CRPS improvement over ENS at 2-day lead times.

• WeatherNext 2: Google DeepMind. “WeatherNext 2: Google DeepMind’s most advanced forecasting model.” Google Blog, November 2025. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-2/

  ---

  You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

In July 2025, a randomised controlled trial set out to measure what AI coding tools actually do to developer productivity. The developers tested were not novices — they were experienced open-source contributors working on codebases they knew well, using the tools that have driven AI adoption at scale: Cursor Pro with Claude 3.5 and 3.7 Sonnet. Before the experiment began, they predicted the tools would make them 24% faster. After completing the tasks, they still believed AI had made them 20% faster. The measured result was that they were 19% slower than developers who had worked without AI. The gap between what the tools felt like and what they produced was 39 to 44 percentage points.

This was not a finding about novices or edge cases. The Model Evaluation and Threat Research group, which conducted the trial, ran it on real issues in real codebases — the kind of work that makes up most professional software development. The minutes saved on initial code generation were consumed by reviewing, fixing, or discarding AI output that was close to correct but not correct enough. The feeling of speed was real. What it was measuring was not productivity.

That gap — between the experience of using AI coding tools and their measurable effect — is what makes the current moment unusual. The tools have been adopted at extraordinary speed, the productivity dashboards show dramatic improvements, and the downstream costs are accumulating in places the dashboards do not look.

What the numbers actually say

The scale of AI code generation adoption is not in doubt. GitHub Copilot reached 20 million users by July 2025 and had 4.7 million paid subscribers by January 2026 — a 75% year-on-year increase. It is deployed at 90% of Fortune 100 companies. By early 2026, 84% of developers reported using AI tools in their development process. GitClear’s analysis of 211 million changed lines of code found that approximately 41% of all code committed globally in 2025 was initially generated or suggested by AI.

By April 2026, Google’s CEO Sundar Pichai confirmed that 75% of all new code at Google is AI-generated — reviewed and approved by engineers before it ships, up from just over 25% in mid-2024.

The productivity data from vendor-run studies is compelling. GitHub’s study of 4,800 developers, conducted with Accenture, found task completion 55% faster with Copilot. Pull request time dropped from 9.6 days to 2.4 days. Successful builds increased 84%. Developers using AI daily merge approximately 60% more pull requests than light users. DX, a developer analytics platform tracking more than 135,000 developers, found that developers using AI tools regularly self-report saving an average of 3.6 hours per week.

These numbers are real. They are also, largely, measuring the things AI code generation is good at: boilerplate, test scaffolding, documentation, refactoring familiar patterns, completing obvious continuations of existing code. The 55% faster figure comes from a task involving JavaScript HTTP server code — a type of work that is highly amenable to AI suggestion, with clear patterns and extensive training data. The same gains do not generalise to all development work, and the vendor-run studies are not designed to measure the downstream costs.

Experienced developers using AI tools felt 20% faster. They were 19% slower. The 39-percentage-point gap between perception and measurement is the most important data point in software development in 2025.

A separate analysis by Faros AI, examining telemetry from more than 10,000 developers across 1,255 teams, found that AI adoption correlated with a 9% increase in bugs per developer. Individual output increased; the organisation’s code quality fell.

The security picture is starker. Veracode tested more than 100 large language models against 80 coding tasks specifically designed to carry a known security dimension — the kind of tasks where a wrong choice introduces a real vulnerability. In 45% of cases, the models chose the insecure path, introducing vulnerabilities from the OWASP Top 10; newer and larger models did not perform significantly better than their predecessors. CodeRabbit’s analysis of 470 real-world pull requests found that AI-generated code introduced 2.74 times more cross-site scripting vulnerabilities than human-written code, and 1.9 times more insecure direct object references. Aikido Security’s survey of 450 organisations found that one in five had experienced a serious cybersecurity incident caused by AI-generated code in 2025.

What the technology cannot see

This is not the first time the industry has tried to automate code production. In the 1980s and 1990s, Computer Aided Software Engineering — CASE — made the same promise: feed your models and specifications into a repository, generate the code. The output was usually unusable. Teams spent longer unpicking what the tools produced than they would have spent writing from scratch. The tools were not wrong about the mechanics of generation. They were wrong about what the hard part of software engineering actually is.

An AI coding assistant faces a version of the same problem at larger scale and higher speed. It has been trained on enormous quantities of code from public repositories. It has learned to predict what code is likely to follow given what has already been written. It is, in that sense, a very sophisticated pattern-completion engine. What it does not have is any understanding of the specific system it is generating code for — the business logic, the security requirements, and the architectural decisions that are specific to this system rather than to code in general.

SQL injection is one of the oldest and most well-documented security vulnerabilities in web development. It appears in the training data with high frequency, both as the vulnerability and as its correct prevention through parameterised queries. AI models should, in principle, know how to avoid it. They often do not. The reason is that the model is optimised to produce code that completes the task as stated — and the task as stated rarely includes ‘and also, make absolutely sure no user input can be injected into this query’. The model takes the shortest path to a passing result. Security requirements that are implied rather than explicit tend to be skipped.

GitClear’s longitudinal analysis of 211 million lines of code identified a more structural problem. The proportion of code that constitutes refactoring — the disciplined work of improving existing code’s structure without changing its behaviour — fell from 25% of all commits to under 10% as AI adoption increased. Simultaneously, code duplication increased fourfold. The pattern is consistent with what happens when code generation is optimised for volume and speed: new code is added rather than existing code improved, similar patterns are copy-pasted rather than abstracted, and the codebase becomes progressively harder to understand and change.

The term that circulated widely in 2025 was ‘vibe coding’, coined by Andrej Karpathy to describe the practice of accepting AI suggestions without reading the generated code carefully — programming by feel, copying error messages back to the AI until it works. As a personal approach to low-stakes projects, this is a reasonable efficiency trade. As a pattern spreading through professional development teams, it has measurable consequences. OX Security examined more than 300 repositories and identified ten recurring anti-patterns present in 70 to 100% of AI-generated code, including incomplete error handling, weak concurrency management, and inconsistent architectural decisions. These are not exotic failure modes — they are the standard failure modes of code written by someone who knows what the output should look like but does not fully understand the system it runs in.

The junior developer question is also changing in ways that are not straightforward. Demand for very junior developers — those whose primary value is writing boilerplate and implementing well-specified features — has softened as AI handles more of that work. Gartner has forecast that 80% of software engineers will need to upskill in AI by 2027. The role that remains, and that is if anything growing in value, is the developer who can specify problems precisely enough for AI to address them, review AI output critically enough to catch its characteristic failure modes, and design systems architecturally rather than producing code line by line. That is a different job description from the one most developers were hired to do five years ago.

Who benefits, and the costs they cannot see

The clearest beneficiaries of AI code generation are not the developers who were already most productive. They are the people who previously could not build software at all. In Y Combinator’s Winter 2025 batch, 25% of startups had codebases that were 95% AI-generated, with small teams — often fewer than ten people — shipping products with paying customers at a pace that would have been impossible two years earlier. A non-developer who can specify what they need in plain language, and who has enough domain expertise to evaluate whether the output works, can now produce functional software at a speed that was impossible without technical co-founders. The barrier that used to require a technical co-founder or a development team has dropped to domain expertise and the ability to specify what you need.

The costs tend to fall elsewhere. Production systems with AI-generated code are accumulating a backlog of maintainability problems, duplicated logic, and architectural inconsistencies that compound over time — what the industry has taken to calling a technical debt tsunami. Analysis by Codebridge describes a predictable pattern in AI-assisted projects without proper governance — what it calls the 18-month wall: velocity gains in months one to three are followed by a plateau, then a decline, and by 18 months the team can no longer fully understand their own system. The code shipped faster. The code that was shipped has become progressively harder to change.

The clearest beneficiaries of AI code generation are the people who previously could not build software at all. The costs — security vulnerabilities, technical debt, degraded maintainability — tend to fall on the teams who inherit the codebase.

The security consequences are not evenly distributed either. An organisation with a mature security engineering function — automated scanning in the continuous integration pipeline, mandatory human review for AI-generated code, developers trained to recognise the characteristic vulnerability patterns that AI produces — can capture the productivity benefit while managing the risk. An organisation without those controls, which describes many of the smaller and less well-resourced companies where AI code generation is growing fastest, is generating code with known vulnerability patterns at scale. The one-in-five cybersecurity incident rate from Aikido Security’s survey is not a future risk. It is a 2025 reality.

What ‘more productive’ actually means

There is a genuine productivity case for AI code generation. For handling tedious scaffolding, repetitive testing, and documentation that would otherwise be skipped, the tools do what they claim — they reduce the time spent on work that does not require judgement, and free the developer to focus on what requires care. The 55% faster result in GitHub’s study is real. The question is whether the developer who is 55% faster at writing the initial code is also better, faster, or more thoughtful at reviewing it.

Whether AI code generation expands what developers can do — building more complex systems, serving more users — or whether it primarily produces more code without proportional improvement in outcomes is not the same question. The technical debt and security findings suggest that ‘more code’ and ‘better software’ are not the same thing. An organisation that measures developer productivity by lines of code or pull requests merged will see dramatic improvements. An organisation that measures by system reliability, security record, and long-term maintainability will see a more complicated picture.

The question that most organisations are not yet asking is whether their approach to AI code generation matches the actual risk. The risk profile is not uniform. A SQL query generated by AI in a financial application has a different security consequence from a CSS class generated by AI in a marketing site. A snippet of authentication logic reviewed by a senior security engineer has a different risk profile from the same snippet reviewed by a junior developer under time pressure.

The organisations managing this well tend to share two things: automated scanning on AI-generated code before it reaches production, and a differentiated approach — more scrutiny where the risk is high, less where it is low. Most apply uniform policies, either permitting AI use broadly or restricting it broadly. Neither matches the actual risk profile.

What the productivity dashboard doesn’t show

1. The METR trial found experienced developers were 19% slower with AI tools while believing they were 20% faster. If you use AI coding tools, how confident are you that your sense of your own productivity is accurate? What would you need to measure to find out?

2. One in five organisations reported a security incident caused by AI-generated code in 2025. Does your organisation scan AI-generated code for security vulnerabilities before it reaches production? If not, what is the process that is supposed to prevent those vulnerabilities from being deployed?

3. AI tools handle the initial code generation faster, but the work of specifying problems precisely, reviewing output critically, and designing systems that remain comprehensible as they grow — these are becoming more important, not less. Is the way your organisation hires, trains, and develops software engineers keeping up with that shift?

4. Startups in Y Combinator’s 2025 batch were building products with 95% AI-generated codebases and reaching real revenue. What does it mean for the software industry that the barrier to building has dropped this dramatically — and for whom has it actually dropped?

The tools are installed. The feeling of acceleration has not gone away. What is accumulating alongside it — the vulnerabilities, the duplicated logic, the refactoring debt — does not appear in the metrics being used to justify the adoption.

Leave a comment

Sources and further reading

1. METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
Joel Becker, Nate Rush, Elizabeth Barnes, David Rein — Model Evaluation and Threat Research, 10 July 2025
metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study
The randomised controlled trial underlying the 19% slower / 20% faster perception gap. 16 experienced open-source developers, real issues from their own codebases, Cursor Pro with Claude 3.5 and 3.7 Sonnet.

2. Veracode — 2025 GenAI Code Security Report
Veracode, July 2025
veracode.com/resources/analyst-reports/2025-genai-code-security-report
100+ large language models tested against 80 coding tasks specifically designed to carry a known security dimension. 45% of cases: model chose the insecure path. Rate consistent across model generations.

3. CodeRabbit — State of AI vs Human Code Generation Report
CodeRabbit, December 2025
coderabbit.ai/whitepapers/state-of-AI-vs-human-code-generation-report
470 real-world pull requests (320 AI-co-authored, 150 human-only). AI code: 2.74× more cross-site scripting vulnerabilities, 1.91× more insecure direct object references, 1.7× more total issues.

4. Aikido Security — 2026 State of AI in Security & Development
Aikido Security, 2025
aikido.dev/state-of-ai-security-development-2026
Survey of 450 CISOs, security leaders, developers, and AppSec engineers across Europe and the US. 20% reported a serious security incident caused by AI-generated code.

5. GitClear — AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones
GitClear, 2025
gitclear.com/ai_assistant_code_quality_2025_research
211 million changed lines of code analysed (2020–2024). Code duplication increased fourfold. Refactoring fell from approximately 25% of commits to under 10% as AI adoption increased.

6. Faros AI — The AI Productivity Paradox
Faros AI, 2025
faros.ai/ai-productivity-paradox
Telemetry from more than 10,000 developers across 1,255 teams. AI adoption correlated with a 9% increase in bugs per developer. Daily AI users merge approximately 60% more pull requests than light users.

7. DX — AI-Assisted Engineering: Q4 Impact Report 2025
DX (Developer Experience), 2025
getdx.com/blog/ai-assisted-engineering-q4-impact-report-2025
435 companies, 135,000+ developers, July–October 2025. Developers using AI tools regularly self-report saving an average of 3.6 hours per week. Note: self-reported perception, not independently measured.

8. GitHub / Accenture — GitHub Copilot Impact Study
GitHub and Accenture, 2024
github.blog/news-insights/research/survey-ai-wave-grows
4,800 developers. Task completion 55% faster with Copilot. Pull request time: 9.6 days → 2.4 days. Successful builds increased 84%. Note: benchmark task involving JavaScript HTTP server code — a domain highly amenable to AI suggestion.

9. GitHub Copilot user and subscriber figures
GitHub announcement, July 2025; Microsoft FY26 Q2 earnings call, 28 January 2026
dataconomy.com/2025/07/31/github-copilot-now-has-over-20-million-users
20 million total users by July 2025. 4.7 million paid subscribers by January 2026 — 75% year-on-year growth. Deployed at 90% of Fortune 100 companies.

10. Google — Sundar Pichai at Google Cloud Next 2026
Google, 22 April 2026
blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai
75% of all new code at Google is AI-generated, reviewed and approved by engineers before it ships. Up from just over 25% in mid-2024.

11. Gartner — Generative AI Will Require 80% of Engineering Workforce to Upskill Through 2027
Gartner, 3 October 2024
gartner.com/en/newsroom/press-releases/2024-10-03-gartner-says-generative-AI-will-require-80-percent-of-engineering-workforce-to-upskill-through-2027

12. OX Security — State of AI-Generated Code Report
OX Security, October 2025
prnewswire.com/news-releases/ox-report-ai-generated-code-violates-engineering-best-practices-302592642.html
300+ open-source repositories analysed. Ten recurring anti-patterns present in 70–100% of AI-generated code, including incomplete error handling, weak concurrency management, and inconsistent architectural decisions.

13. Codebridge — The Hidden Costs of AI-Generated Software: Why It Works Isn’t Enough
Codebridge
codebridge.tech/articles/the-hidden-costs-of-ai-generated-software-why-it-works-isnt-enough
Analysis of the 18-month wall pattern in AI-assisted projects without proper governance: velocity gains in months one to three, followed by plateau and decline.

14. Garry Tan — Y Combinator W25 batch AI-generated code
Garry Tan, X, 4 March 2025
x.com/garrytan/status/1897303270311489931
*25% of YC Winter 2025 batch startups had codebases that were 95% LLM-generated.

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

In a field hospital in an active conflict zone, a patient arrives with a severe bacterial infection. The hospital has the diagnostic equipment to identify the pathogen. What it does not have is the specific therapeutic protein — the monoclonal antibody, the recombinant enzyme, the vaccine dose — that the patient needs. The nearest manufacturing facility that could produce it is a continent away, behind a cold chain that has been disrupted, in a centralised facility that requires months of lead time to retool for a new product.

This is not an edge case. It describes, in concentrated form, the structural problem of biological medicine manufacturing: the gap between knowing what a patient needs and being able to produce it where the patient is. Most biological medicines are made by growing engineered living cells in large fermentation tanks, harvesting and purifying the protein they produce, and shipping it through refrigerated supply chains to the places where it is needed. The system works well in wealthy countries with stable logistics and well-capitalised health systems. It works considerably less well everywhere else.

Cell-free biomanufacturing is a different approach, and it is further along than most people know.

Biology outside the cell

Every living cell is a protein-making machine. The genetic instructions encoded in DNA are transcribed into messenger RNA, which is then translated by ribosomes into proteins. Cell-free protein synthesis, an active research area since at least the 1960s, captures that production machinery and runs it outside any living cell. The process begins by breaking open a batch of cells, typically engineered bacteria such as Escherichia coli, though yeast, plant, and mammalian cell extracts are also used, and extracting the ribosomes and associated enzymatic machinery. Add the genetic instructions for the protein you want; the machinery reads them and produces it. No fermentation tank. No living cell line. No months of retooling to switch to a new product.

The practical advantages over conventional biomanufacturing are specific. Because there is no living cell to maintain, proteins that would kill a conventional host cell line before producing useful quantities can be made cell-free without this constraint. Process development is faster: switching from producing one protein to another requires changing only the DNA or mRNA template, not engineering a new cell line and running the months of validation work that goes with it.

The property with the most direct implications for global health is lyophilisation. Cell-free reactions can be freeze-dried into a stable powder that requires no refrigeration. Researchers at the Jewett Laboratory, now at Stanford, have demonstrated conjugate vaccines produced this way at approximately $0.50 per dose, stable at 37 degrees Celsius — ambient temperature in much of the world without reliable cold chains — for up to four weeks. Separate work from the same group, funded by DARPA and published in Biotechnology and Bioengineering in 2025, demonstrated scalable cell-free production of T7 RNA polymerase — a critical enzyme in mRNA vaccine synthesis — achieving over 90% purity at one-litre scale.

Commercial applications are beginning to emerge. Sutro Biopharma uses a proprietary cell-free platform to produce antibody-drug conjugates and bispecific antibodies, with candidates in clinical trials. Ipsen Biopharm uses cell-free synthesis specifically because it minimises containment risk when producing highly toxic botulinum toxin — a case where the open system’s safety advantages are directly relevant to manufacturing practice. GreenLight Biosciences applied cell-free systems to RNA vaccine production before its 2023 acquisition, after which the platform was redirected to RNA-based agricultural biocontrol — a reminder that commercial priorities and public health priorities do not always point in the same direction.

What the technology cannot yet reach

Technology Readiness Level 5 is an honest assessment of where most cell-free biomanufacturing sits in 2026. The technology has been validated in laboratory settings and small-scale demonstrations. It has not yet been validated in the large-scale, regulated manufacturing environments that pharmaceutical products require before they can be administered to patients.

The most significant technical constraint is yield. Cell-free systems typically produce lower quantities of protein per unit volume than conventional cell-based fermentation at commercial scale. For high-value, low-volume applications — speciality biologics, antibody-drug conjugates, research reagents — this is manageable. For mass vaccine manufacturing at the scale of hundreds of millions of doses, the yield gap remains a real cost constraint.

Post-translational modifications present a related challenge. Many therapeutic proteins require specific chemical modifications after assembly — glycosylation being the most important, where sugar molecules are attached at specific positions and affect the protein’s stability, activity, and immunogenicity. Recent work from the Jewett group, published in ACS Synthetic Biology in 2025, achieved over 85% glycosylation efficiency and yields of up to 450mg per litre of glycoprotein in a two-step cell-free platform. Progress, but still an active engineering challenge rather than a solved problem.

Good Manufacturing Practice validation, the regulatory standard required before medicines can be produced for clinical use, has not yet been established for cell-free biomanufacturing systems. The mRNA vaccine precedent is informative in both directions: a sufficiently urgent need can accelerate regulatory development dramatically, and absent that urgency, the pathway tends to lag the science by years.

The $0.50 per dose figure for lyophilised conjugate vaccines is also derived from raw material costs at laboratory scale — a proof of concept, not a commercial cost estimate. At manufacturing scale, enzyme costs, lysate production, energy, and quality control all need to be incorporated. The expectation is that cell-free systems will be cost-competitive for certain product classes, particularly those requiring rapid product switching, but this has not yet been demonstrated in practice.

Who gets to make medicine

The current geography of biological medicine manufacturing is highly concentrated. The large fermentation-based facilities required to produce biologics at scale are predominantly located in the United States, Europe, and a small number of other high-income countries. The capital cost, typically in the range of hundreds of millions to over a billion dollars, and the technical expertise required to operate such a facility means that most low- and middle-income countries have no domestic biomanufacturing capacity for the medicines their populations need most.

The COVID-19 pandemic made this structural dependency visible in its most acute form. When mRNA vaccine production was concentrated in a handful of facilities, the question of who received vaccines first was not determined by medical need. It was determined by geography, procurement power, and the location of manufacturing capacity. Countries without domestic production capability were last in line, and in some cases, doses pledged through COVAX arrived too late to prevent the deaths they were intended to prevent.

Cell-free biomanufacturing addresses this problem directly. A cell-free system producing lyophilised vaccine components requires no fermentation infrastructure, no refrigeration during storage or transport, and significantly less technical overhead to operate than a conventional bioreactor facility. The direction the research points toward is distributed, decentralised manufacturing: regional production facilities, field hospital capabilities, in principle point-of-care production in the settings where patients need treatment.

The access implications extend beyond pandemic preparedness. Many neglected tropical diseases — conditions that cause enormous suffering in low-income countries but attract little pharmaceutical development investment because the affected populations cannot pay prices that support conventional manufacturing economics — could in principle be addressed by cell-free produced therapeutics if the manufacturing cost barrier were reduced. A vaccine against enterotoxigenic E. coli, one of the leading causes of diarrhoeal disease and child mortality in the developing world, was among the first demonstrated applications of low-cost lyophilised cell-free vaccine production. The connection between manufacturing economics and which diseases receive medicines is direct.

The test that matters most is whether the technology is deployed where the access gap is sharpest, rather than where the regulatory environment is most familiar. A reduction in the capital and manufacturing threshold changes little if the pathway to GMP certification, WHO prequalification, and COVAX procurement simply replicates the infrastructure requirements of conventional manufacturing. Regulatory agencies in low-income countries, COVAX procurement frameworks, and WHO prequalification processes will all need to develop explicit pathways for cell-free produced biologics. The geographic promise of the technology does not translate into geographic reality without that work.

The field hospital in the conflict zone, the outbreak in the region without cold chain infrastructure, the neglected disease in the country without manufacturing capacity — these are not hypothetical future scenarios. They are the current state of biological medicine access for a substantial portion of the world’s population.

Cell-free biomanufacturing does not yet solve those problems. It is an early-stage technology with real constraints and a significant gap between laboratory demonstration and clinical deployment. It is, however, the first credible technical approach to a problem that has always been treated as structural and immovable: that making biological medicines requires the kind of industrial infrastructure that is simply not available in most of the world.

The geography of manufacturing is the geography of access. The question the science cannot answer on its own is who will fund the work to change it.

My Opinion

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Satellite internet from low Earth orbit is no longer experimental. Starlink has roughly 9,000 satellites in orbit, operating close enough to the surface to keep latency low enough for video calls, cloud software, and large file transfers — in places that will never see a fibre cable. The coverage footprint reaches an estimated 3.1 billion people. As of December 2025, the subscriber base was nine million.

Those two numbers describe the same technology. They are very different things.

Coverage is the geographic area within which the signal can, in principle, be received. Subscribers are the people who can afford to use it, who have the hardware, and who live in markets the provider has chosen to serve. The engineering problem is largely solved. What hasn’t been solved, and what the headline numbers tend to obscure, is the price, the politics, and the question of who gets to decide.

The constellation that already exists

SpaceX’s Starlink is the dominant reality of satellite internet in 2026. Around 9,000 Starlink satellites are in low Earth orbit operating between 340 and 1,200 kilometres above the surface, low enough to keep signal latency to 20–50 milliseconds rather than the 600-millisecond round trips of the old geostationary satellites that made satellite internet frustratingly slow throughout the era of geostationary connectivity. As of December 2025, Starlink had roughly nine million subscribers across more than 155 countries, with coverage reaching an estimated 3.1 billion people. Residential plans in most markets cost around $110 per month, plus an upfront hardware cost of $350–$600 for the dish and router. Speeds typically range from 50 to 220 Mbps under good conditions.

The nearest competitor in low Earth orbit is Eutelsat OneWeb, formed from the merger of the European satellite operator Eutelsat and the British-Indian OneWeb constellation. OneWeb operates 648 satellites, primarily serving enterprise, government and maritime customers, with particular strength in polar regions. It is not a consumer product in the way Starlink is. Amazon Leo, rebranded from Project Kuiper in November 2025, is the most significant new entrant. With approximately 300 production satellites in orbit as of April 2026 and an FCC-authorised constellation of 3,236 planned, Amazon began an enterprise preview in late 2025 and is rolling out service more broadly in 2026 as satellite density increases. Its integration with Amazon Web Services gives it something Starlink lacks: a natural fit with the infrastructure decisions that large organisations have already made. Telesat Lightspeed, Canada’s planned LEO constellation, is targeting 2027.

A separate development sits alongside the LEO constellations: AST SpaceMobile, which is building satellites that connect directly to standard mobile phones, no specialised dish required, via partnerships with AT&T, Verizon and other carriers. The first generation of service supports messaging and basic data. More capable versions are planned. If AST SpaceMobile delivers on its roadmap, the access barrier shifts entirely: connectivity from orbit available to anyone with an ordinary mobile handset.

Taken together, the technology represents a real shift. Low Earth orbit satellite internet provides broadband speeds with latency approaching that of ground-based cable, to places that will never see a fibre rollout. In Kenya, where Starlink launched in 2023, analysis by Rest of World found that the monthly subscription cost is lower than that of the leading fixed-line providers. In five African countries — Kenya, Ghana, Zimbabwe, Mozambique and Cape Verde — Starlink is cheaper than the cheapest available broadband alternative. In Indonesia, nearly 60% of Starlink’s subscriber base is in rural areas, nearly four times the proportion seen among fixed wireless or fixed-line users, reflecting how the technology has positioned itself in markets where ground-based infrastructure is absent, according to an Opensignal analysis published in November 2025.

The three gaps the headline numbers hide

Stating that Starlink reaches 3.1 billion people is a statement about coverage, the geographic area within which the signal can, in principle, be received. It says nothing about who can afford to use it. The two numbers are very different.

In Nigeria, as of late 2025, the standard Starlink residential subscription costs the equivalent of $39 per month. The national minimum wage is approximately $48 per month. The monthly subscription alone therefore consumes more than 80% of minimum wage earnings — before accounting for the hardware cost of the dish, which runs to around $406. In an analysis published by TechCabal, the conclusion was direct: Starlink in Nigeria is a service for high earners. It addresses a real connectivity gap and is faster than most alternatives, but it is not a vehicle of mass digital inclusion at current pricing. The people it most visibly fails to serve are precisely the ones whose connection to the global economy would benefit most from affordable internet.

The community hub model is a partial answer. A single Starlink terminal shared across a school, a health clinic, or a rural community centre provides access at lower per-user cost than individual subscriptions. Governments in Mexico, parts of Sub-Saharan Africa and Pacific Island nations have subsidised terminal deployments for schools and public institutions. These approaches work within the technology’s constraints, but they provide intermittent shared access rather than the persistent personal connectivity that changes what an individual can do - hold a remote job, study independently, run a small business.

In Nigeria, the standard Starlink monthly subscription consumes more than 80% of the national minimum wage before hardware. The technology covers the ground. The economics do not reach the people who most need it.

The second gap is geopolitical. Starlink is owned and operated by SpaceX, which is owned and personally controlled by Elon Musk. The service is now active in 160 markets, operates across every active military theatre where Western forces are present, and serves as critical communications infrastructure for Ukraine’s armed forces. In 2022, Musk declined to extend Starlink coverage to support a Ukrainian naval drone operation near Crimea, citing his personal assessment of escalation risk. In early 2025, US negotiators were reported to have threatened restrictions on Ukraine’s Starlink access during negotiations over a minerals deal — a threat the US government did not officially confirm, and that Musk denied SpaceX would act on, but one whose credibility was sufficient to generate significant anxiety in Kyiv. When a private company, and effectively a single individual, can decide whether a frontline military has communications access, the relationship has moved beyond commercial. It has become a question of which private actor holds sovereign functions that states have not figured out how to reclaim.

Countries that have noticed this dynamic are responding in predictable ways. China has authorised its own domestic LEO constellation. The European Union is accelerating IRIS², its planned governmental satellite constellation. India is developing its own LEO system. These are expensive and slow. They reflect a rational assessment that dependence on SpaceX for critical communications infrastructure carries political risk that states are not willing to accept indefinitely. For countries without the resources to build a rival system, the options are narrower.

The third consequence is further removed from any individual subscriber but harder to reverse. When no governance mechanism constrains how many objects private operators can place in orbit, the sky fills up faster than anyone has agreed to manage it. Starlink alone has launched more than 10,700 satellites, with an FCC-approved plan for nearly 17,000 second-generation satellites and a stated long-term ambition of 42,000. Amazon Leo plans 3,236. Other constellations add thousands more. The number of operational satellites in low Earth orbit has roughly tripled in five years. This creates a set of problems that have no straightforward commercial solution. Collision risk between satellites and between satellites and other space objects increases with orbital density. Light pollution from reflective satellite trails is now a documented problem for ground-based astronomy. The Kessler syndrome — a cascade of collisions generating debris that makes entire orbital shells unusable — is a theoretical endpoint that gets less theoretical as orbital density increases. Russia has reportedly been developing a weapon designed to flood Starlink’s orbital shells with clouds of high-density pellets. Analysts and intelligence assessments from two NATO nations have warned that such an attack could generate debris cascading out of control, threatening not only Starlink but every constellation in low Earth orbit, including those operated by Russia and China.

What connectivity from orbit actually changes

The changes that satellite internet delivers when access is genuinely affordable are concrete. In rural communities with reliable satellite connectivity, telemedicine consultations become possible, a GP or specialist in a city can see a patient in a remote village. Remote education stops being a compromise of intermittent video and becomes a sustained engagement. Agricultural information services that require real-time connectivity — weather alerts, market prices, crop disease identification through image recognition tools — reach the farmers who most need them. Disaster response changes when communications infrastructure can be re-established in hours via satellite rather than weeks via cable repair.

These are not hypothetical. They are documented in communities across Sub-Saharan Africa, Latin America, Southeast Asia and the Pacific where satellite connectivity has arrived and where the access barrier was cleared, often through subsidised hardware, government procurement, or NGO deployment. The lesson from these communities is not that the technology solves the connectivity problem. It is that the technology, when its economics are addressed, removes one of several barriers that have kept communities disconnected. The remaining barriers power access, device access, digital literacy, content in local languages, do not disappear because the signal arrives from orbit.

The consequences of the geopolitical concentration are slower to manifest but potentially more durable. A world in which critical communications infrastructure for developing nations runs through a constellation controlled by a single billionaire’s company is a world in which the decisions that individual makes about pricing, coverage, and access carry consequences that previously would have required state-level action. SpaceX can raise prices, withdraw from markets, or restrict usage in ways that no regulatory body currently has clear authority to prevent. The International Telecommunication Union governs spectrum allocation. It does not govern service continuity for nations that have come to depend on a private provider for essential communications.

The gap between coverage and access

Satellite internet works best for people when they can actually afford to use it. Where the economics work, where the subscription is affordable and the hardware accessible, the technology removes a real barrier that ground infrastructure failed to remove. A household with a working subscription has gained persistent connectivity: enough for remote work, independent study, a small business operated from somewhere that previously had none of those options. A school with a shared terminal has partial help. A farming community that cannot clear the hardware cost has nothing. The distinction matters because it determines whether the technology fulfils the claim made for it or whether it creates a new tier of the digital divide: connected enough to know what you are missing, not connected enough to benefit from it.

Whether satellite internet adds something that terrestrial networks cannot is answered most clearly at the coverage edge. In places where fibre will not arrive within this generation’s planning horizon, LEO satellite is not a supplement to terrestrial infrastructure. It is the infrastructure. The improvement is real and significant for those communities. For communities where ground-based options exist, satellite is generally a backup or a premium option, and the case is weaker.

The hardest question is not about the technology. The technology does not adapt to the economic circumstances of the people it claims to serve. It adapts to the willingness to pay of the people who can afford it. The pricing structure that works in Germany or Australia does not work in Nigeria or Papua New Guinea. SpaceX has shown some willingness to adjust regional pricing, plans in some African markets run as low as $10 per month for a capped data allowance, but the hardware cost remains a significant barrier even at reduced subscription rates. An adaptive approach would involve structured subsidies, national procurement at scale, or hardware financing models that distribute the upfront cost over time. Some governments are pursuing these. Most are not.

The engineering is largely solved. What isn’t solved is the price, the politics, and the question of who gets to decide.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Quantum computers have always promised extraordinary things. They have also always broken — not because of poor engineering, but because of the physics. The qubits in a quantum computer are extraordinarily fragile. They interact with their environment, lose their quantum state, and make errors at rates that would render any classical computer unusable. The field has known this since the beginning. It has also known what is required to fix it.

In December 2024, Google announced that its Willow quantum chip had crossed a threshold that quantum computing researchers had been working toward since the mid-1990s. Understanding why that matters — and why it is still only the beginning — requires a short account of what quantum computing actually is, and why errors are so central to the challenge.

Why errors are everything

A classical computer stores information as bits: each one is definitively a zero or a one. A quantum computer stores information as qubits, which can exist in what is called a superposition of zero and one simultaneously. This is not a metaphor — it is the actual quantum mechanical state of the physical system. What makes quantum computers potentially powerful is that a system of qubits can represent and process an exponentially large number of possible states at the same time, performing certain kinds of calculations far more efficiently than any classical approach.

The catch is that this quantum state is extraordinarily delicate. The moment a qubit interacts with its environment — with vibration, heat, stray electromagnetic radiation, even the act of measurement — the superposition collapses. The qubit behaves like a classical bit. The quantum advantage disappears.

This process is called decoherence. Managing it is the central challenge of quantum computing.

The solution theorists developed, quantum error correction, sounds counterintuitive at first. Instead of trying to build qubits that don’t make errors (which turns out to be essentially impossible with current technology), you build a logical qubit from many physical qubits, encoding the information redundantly and using the correlations between the physical qubits to detect and correct errors before they propagate. A logical qubit might require tens, hundreds, or eventually thousands of physical qubits to maintain reliably.

This creates a daunting scaling problem. If each logical qubit requires a hundred physical qubits, and a useful quantum algorithm requires a thousand logical qubits, you need a hundred thousand physical qubits all operating with error rates low enough for the correction scheme to work. For most of quantum computing’s history, this seemed so distant as to be almost theoretical. Then, in late 2024, Google published something that changed the conversation.

The error is not a bug to be patched. It is a consequence of the physics. Quantum error correction doesn’t prevent errors — it detects and corrects them before they matter, using redundancy across many physical qubits.

Google’s Willow chip demonstrated, for the first time in any quantum computing platform, that making a quantum error-correcting code larger actually reduced the error rate of the encoded logical qubit, and that this improvement was exponential. This is the property theorists had predicted was necessary for error-corrected quantum computing to work at scale, and it had never been conclusively shown experimentally. The larger the surface code, the more physical qubits devoted to error correction, the better the logical qubit performed.

The scaling law held. The path, at least in principle, was clear.

Willow’s 105 physical qubits also completed a specific benchmark calculation in roughly five minutes that would require a classical supercomputer an estimated 10 to the power of 25 years to perform. That number is so large as to be meaningless in everyday terms — the age of the universe is around 14 billion years, or roughly 10 to the power of 10. The point is not the absolute comparison. It is the direction and magnitude of the gap.

Three roads, one destination

The quantum computing field in 2025 and 2026 is characterised by genuine competition between categorically different technological approaches — not just different companies building the same kind of machine, but different physical architectures with different advantages, different error profiles and different paths to scale. Understanding them does not require a physics degree. It requires grasping one central distinction: whether error correction is done primarily in software, or baked into the hardware.

Google and IBM have both built their systems around superconducting qubits — tiny circuits, cooled to near absolute zero, that exhibit quantum behaviour. IBM’s roadmap is the most detailed in the industry. Its Nighthawk processor, released in late 2025, delivered 120 high-quality physical qubits and introduced couplers that extend connectivity across the chip. IBM targets verified quantum advantage, solving a problem better than any classical method, by the end of 2026, and a full fault-tolerant machine with 200 logical qubits by 2029. Its approach emphasises steady incremental progress and a clear engineering roadmap.

IonQ and Quantinuum build their qubits from individual ions — electrically charged atoms trapped in electromagnetic fields. Trapped-ion systems have better error rates per operation than superconducting systems, and any two trapped-ion qubits can interact directly without the connectivity constraints of a chip. The trade-off is speed: trapped-ion operations are slower. QuEra and Atom Computing use neutral atoms, manipulated with precisely aimed laser beams, which offer their own connectivity advantages. IEEE Spectrum’s Top Tech 2026 report named neutral-atom systems among the most promising near-term candidates for error-corrected demonstrations.

Microsoft’s approach is categorically different from all of the above. In February 2025, the company announced Majorana 1, a chip built around what it calls topological qubits — a form of qubit that encodes information in the global geometric properties of a physical system rather than in the state of individual particles. The theoretical appeal is significant: a topological qubit should be inherently resistant to local errors because its quantum information is distributed, not localised. Braid the Majorana particles around each other and you have performed a computation; local noise cannot accidentally unbraid them.

The announcement attracted intense scrutiny. The Nature paper came with an editorial note stating that the results did not constitute evidence for topological modes as claimed in Microsoft’s press release. At the American Physical Society’s global summit in March 2025, physicists including Henry Legg of St Andrews and Eun-Ah Kim of Cornell went on record with doubts about Microsoft’s verification methods — the field’s independent assessment concluding that the evidence fell short of establishing topological modes, and that the topological approach remained the furthest from practical demonstration of the competing architectures. Microsoft’s researchers disputed this forcefully. Scott Aaronson, a prominent quantum computing theorist at the University of Texas, captured the independent position precisely: the approach is worth pursuing, but the evidence so far is partial and further independent replication is required.

By mid-2025, Microsoft had published additional work demonstrating a tetron qubit device, the next step in its roadmap, showing that the error protection mechanism behaved as theoretically predicted. The picture was clearer. The independent verification was not yet there.

What becomes possible

The question of what a working quantum computer could actually do is easier to answer than the question of when one will exist.

Drug discovery and materials science. A fault-tolerant quantum computer could simulate chemical reactions with a precision no classical supercomputer will ever achieve, regardless of how much faster they become. The implications for discovering new medicines, designing new catalysts, modelling protein folding are real. A simulation that shows exactly how a drug candidate binds to a protein in living biology is not a fantasy. It is the purpose for which quantum computing was originally conceived.

Cryptography. Most of the encryption protecting the internet — the secure connections used for banking, e-commerce, and private communication — relies on the computational difficulty of factoring very large numbers. A sufficiently capable quantum computer running Shor’s algorithm could break RSA encryption. This is not a near-term threat: the quantum computers that exist today are nowhere near the scale required.

But a 2025 analysis by Google Quantum AI revised earlier estimates of around twenty million physical qubits down to approximately one million. As error correction improves, the timeline for this threat becomes less abstract. The transition to post-quantum cryptographic standards — already underway, with NIST finalising new standards in August 2024 — is not an abundance of caution. It is a necessary precaution against a threat that will arrive eventually.

Optimisation. Many of the most economically important computational problems — logistics routing, financial portfolio optimisation, supply chain management, climate modelling — are optimisation problems, in which you are searching for the best solution across an enormous space of possibilities. Quantum approaches to these problems may deliver advantages that are harder to characterise than the clear theoretical guarantees of drug simulation or cryptography, but that could be commercially significant. IBM, working with partners including Boeing, Cleveland Clinic and Oak Ridge National Laboratory, is actively investigating where quantum advantage first becomes tangible in real industrial problems.

A fault-tolerant quantum computer could simulate chemical reactions at a level no classical machine will ever match. Drug discovery, materials science, and cryptography are the three applications where the stakes are clearest.

The honest position on timelines is the one offered by IEEE Spectrum’s Top Tech 2026 report: we will not get there in 2026. The quantum computers that exist today, including Google’s Willow and IBM’s Nighthawk, are capable of impressive demonstrations. They are not capable of the sustained, reliable, large-scale error-corrected computation that drug discovery or cryptographic breaking would require. IBM targets fault-tolerant machines by 2029. The same report puts the first practical, large-scale quantum applications in the early 2030s at the optimistic end — and considerably later at the realistic end.

Jensen Huang’s January 2025 estimate of fifteen to thirty years drew enough pushback that he publicly walked it back two months later; Bill Gates suggested the possibility of three to five. The range of expert opinion remains wide.

What has changed, what makes 2025 and 2026 different from the decade before them, is that the theoretical framework has been experimentally confirmed. The below-threshold demonstration matters not because it delivers useful computation today, but because it validates the entire foundation that error-corrected quantum computing depends on. The scaling law holds. The path is real. The question has shifted from whether this can work to when.

Who this is actually for

The case for quantum computing is most often made in terms of what it makes possible — better drugs, stronger materials, faster computation. The more important question is who it is being developed for. The molecular simulations quantum computers could enable would change what research chemists can do in ways classical computers structurally cannot. The relevant question is whether the technology is being developed with those intended beneficiaries clearly in view — the researchers, the patients who might receive better drugs, the communities whose data security depends on cryptographic standards. Or whether the primary beneficiaries turn out to be the companies and investors who will control access to it.

The access question is not resolved by Quantum-as-a-Service platforms existing. IBM, Google and Microsoft all offer cloud-based quantum access to researchers worldwide. Whether that access is available to the researcher in Singapore, in Nairobi, in São Paulo who needs quantum capabilities for their scientific work — whether they can afford it, and use it — is not a technical question. It is an economic and political one.

The difference between quantum capability being broadly accessible and quantum capability being concentrated in the hands of governments and large corporations is not determined by the physics. It is determined by pricing, by infrastructure, and by decisions that technology companies and policymakers are making now.

The most urgent question is specific: the RSA cryptography threat. Quantum computers will eventually be able to break the encryption that currently protects much of the world’s digital infrastructure. The timeline is not known precisely, but it is not infinite. Google has set a 2029 internal deadline for completing its own migration to post-quantum cryptography — well ahead of NIST’s 2035 guideline and NSA’s 2031 target. The organisations that are actively planning for this transition — implementing the NIST standards, auditing their systems, migrating to quantum-resistant encryption — are acting on the right information.

Those treating this as a distant concern are making a mistake that is difficult to correct retroactively: data that is encrypted today but vulnerable to future quantum attack is being harvested now by adversaries planning to decrypt it later. This is the harvest now, decrypt later threat, and it is operational, not theoretical.

The decisions that can’t wait

Google’s Willow demonstration crossed a threshold that quantum computing theorists had worked toward since the mid-1990s. How much does that milestone matter to you — does it feel like a genuine turning point, or another piece of hype in a field that has promised results for decades?

The encryption protecting your banking, your communications and your medical records is vulnerable to a sufficiently capable quantum computer. Post-quantum cryptography standards have been finalised. Does the organisation you work for, or depend on, have a plan for migrating to them? If not, why not?

Drug discovery, materials science, logistics — the applications that stand to benefit most from quantum computing are spread across industries and geographies. Who controls access to the quantum computers that make those benefits possible? Does it matter?

Microsoft’s topological qubit claims generated genuine scientific disagreement — between company researchers and independent physicists, in peer-reviewed journals and at conferences. How do you weigh corporate announcements about breakthrough technologies against the more cautious assessments of independent scientists? What standard of evidence would you need to believe this technology had arrived?

Quantum computing has been described as revolutionary for so long that the word has lost some of its force. It has been genuinely difficult, the promises have repeatedly outrun the delivery. That is the right response to the actual physics rather than a failure of ambition. The wait is now measured in years, not decades.

My opinion

The claim that quantum computing has the potential to change everything has been made before — about electricity, about computing, about the internet. Sometimes it was right.

The challenge here is not building the machine. It is understanding what the machine is doing. If the computation happens inside error-correcting codes that no individual can read back, how do we know when something is wrong? If a company uses quantum simulation to discover a drug, how does a regulator verify the reasoning?

Trust requires interpretability. Quantum computing, at scale, may not offer it in a form that governance can work with.

When the technology becomes commercially viable, and the evidence now suggests that it will, companies will need to test it, govern it, audit it, and account for what it decides. The questions of how to use it, how to verify it, how to control it and who decides are not technical questions. They are political and regulatory ones. The infrastructure to answer them does not yet exist. The physics has been ahead of the governance for a generation. The gap is still widening.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Alex is a robotics engineer. The robot Alex works on needs to process visual data continuously tracking its surroundings, identifying objects, navigating obstacles, responding to unexpected movement. The processing runs on a GPU. The GPU is good at the task. It is also expensive, hot, and power-hungry in a way that limits how long the robot can operate on a battery before it needs to be recharged. A robot that runs for eight hours and then needs to stop is a different kind of tool from one that runs for days.

The human visual cortex performs real-time visual processing continuously throughout an entire waking day on roughly 20 watts — the power of a dim light bulb. It does not process every photon that hits the retina. It responds to change: to movement, to novelty, to things that differ from what was there a moment ago. When nothing changes, the visual system does relatively little. When something changes, it fires. The computation is event-driven, sparse, and asynchronous. It is radically different from the way a GPU processes information, which polls every data point on a regular clock cycle whether anything has changed or not.

Neuromorphic computing is the attempt to build hardware that works the way the brain works. Not to simulate the brain, that is a different and much larger ambition, but to apply the brain’s computational principles to electronic hardware: spiking rather than continuous processing, event-driven rather than clock-driven with memory and computation co-located. The result, in current prototypes and early deployments, uses orders of magnitude less energy than conventional chips for the specific classes of problems it handles well.

In early 2024, Intel delivered a system called Hala Point to Sandia National Laboratories in New Mexico. It contains 1.15 billion artificial neurons across 1,152 chips, fits in a chassis about the size of a microwave oven, and consumes a maximum of 2,600 watts. Shortly after its arrival, Sandia’s researchers published results in Nature Machine Intelligence showing that the system could solve partial differential equations — the mathematical foundation of physics simulation, climate modelling, and structural mechanics — with remarkable efficiency. The finding was described by the lead researcher as surprising. The field had assumed since at least the 1990s that neuromorphic hardware was suited to pattern recognition and AI inference. The Sandia result suggested the range of problems it could address was considerably wider.

Why the brain is a better computer for some problems

Every conventional computer built since the 1940s is based on what is called the von Neumann architecture: a central processor that executes instructions, separate memory that stores data, and a bus connecting the two. The processor fetches data from memory, processes it, stores the result back in memory, and repeats. The bottleneck in this design, known as the von Neumann bottleneck, is the constant movement of data between processor and memory. Modern computers have become extraordinarily fast at this movement, but it consumes a significant fraction of total energy, and it scales poorly with the kinds of AI workloads that need to process large amounts of data in real time.

The brain does not work this way. Neurons are both the processors and the memory. Synaptic connections store learned information in the strength of the connections themselves, not in a separate memory system. Processing happens where the memory is. There is no data bus, no constant shuttling between computation and storage. Instead, neurons accumulate inputs from their neighbours until a threshold is reached — the neuron fires, sending a brief electrical pulse, a spike, to downstream neurons — and then return to rest. Most neurons are quiet most of the time. The brain’s energy consumption reflects this: approximately 20 watts for the entire organ, with most of it going to the small fraction of neurons active at any given moment.

Neuromorphic hardware encodes these principles in silicon. The basic unit is the artificial neuron: a circuit that accumulates inputs, fires a spike when a threshold is crossed, and then resets.

Synapses between neurons have weights, the strength of the connection, which encode learned information. The processing is asynchronous: there is no master clock dictating when each neuron must compute. A neuron fires when it has received enough input, not at a prescribed time. Communication between neurons happens through spikes, brief binary events, rather than through the continuous, high-precision numerical values that GPU computations use.

This spike-based communication is sparse: most connections carry no signal most of the time. The energy expenditure tracks the actual computation being performed, not the maximum possible computation the hardware could perform.

The human brain processes visual information continuously on roughly 20 watts. A GPU doing comparable work uses hundreds of times more. Neuromorphic hardware applies the brain’s computational principles — event-driven, sparse, memory-and-processing co-located — to silicon.

The advantages are most pronounced for sparse, event-driven workloads: sensory processing, pattern recognition in streaming data, anomaly detection, robotics navigation. These are tasks where a great deal of the input data is redundant, most of what a camera sees most of the time is the same as what it saw a moment ago, and where the valuable signal is the exception rather than the rule. Event-based cameras, which fire only when light levels at a pixel change rather than capturing full frames at fixed intervals, are a natural complement to neuromorphic processors: the data is already in spike format, and the neuromorphic hardware processes it without the overhead of decoding conventional frame-rate video.

The limitations are equally specific. Neuromorphic hardware struggles with dense, high-precision numerical computation: floating-point arithmetic, matrix multiplication at scale, the kinds of operations that underpin training large neural networks. The training of GPT-3 required roughly as much energy as powering 120 houses for a year; independent researchers estimate GPT-4 required 40 to 50 times more, though OpenAI has not released official figures. Neuromorphic hardware is not, at present, a candidate for replacing the GPU clusters that train large language models. Where it is a candidate is in inference, running a trained model to produce outputs, and in the specific problem domains where sparse, event-driven computation aligns with the structure of the data.

The field, the hardware, and the unexpected results

The major neuromorphic platforms as of 2026 reflect different architectural philosophies and different answers to the question of what brain-inspired hardware should prioritise.

Intel’s Loihi 2, the chip the Hala Point system is built from, is fabricated on a 4-nanometre process and contains 1 million programmable neurons and 120 million synapses per chip. It supports a range of neuron models, including graded spikes, pulses that carry multi-dimensional information rather than simple binary fire-or-not signals, which bridges some of the gap between spiking neural networks and conventional deep learning. Loihi 2 ships with Lava, an open-source software framework that allows researchers to develop applications for neuromorphic hardware without requiring deep hardware expertise. Intel’s Loihi 3, released in January 2026, scales the per-chip capacity to 8 million neurons and 64 billion synapses, and is the first in the series to be made commercially available outside research programmes, with general availability targeted for Q3 2026.

IBM’s TrueNorth demonstrated a different set of priorities: extreme energy efficiency for inference on fixed classification tasks, at the cost of flexibility. IBM’s NorthPole, a more recent architecture, maintains some TrueNorth influence while integrating better support for conventional neural network primitives. The University of Manchester’s SpiNNaker 2 system, developed through the European Human Brain Project and now commercialised through SpiNNcloud, uses a mesh of ARM processor cores configured to simulate spiking neurons, a more general-purpose approach that is more flexible but less energy-efficient than dedicated neuromorphic circuits. In early 2025, Sandia received a SpiNNaker 2 system capable of modelling 175 million neurons alongside its Hala Point system, using the two architectures’ complementary properties for different research tasks.

BrainScaleS-2 from Heidelberg University takes the most biologically faithful approach: analogue circuits that directly emulate the electrical behaviour of neurons and synapses. Analogue neuromorphic hardware can run neural simulations up to 10,000 times faster than biological real time, because the physical dynamics of the circuits naturally implement the neuron model without requiring digital computation. This gives it unique advantages for neuroscience research, simulating large cortical circuits to test theories of brain function, but the inflexibility of analogue circuits, which cannot easily be reconfigured to implement different neuron models, limits its applicability outside research.

The Sandia PDE result, published in Nature Machine Intelligence, is the most conceptually significant recent finding. Since at least the 1990s, the consensus was that neuromorphic hardware’s natural domain was pattern recognition and neural network inference, tasks that share obvious structural similarity with the spike-based computation the hardware implements. Partial differential equations, which describe how physical quantities like temperature, pressure, velocity, and electromagnetic fields change over time and space, seemed to belong to a completely different computational domain: dense, high-precision, well-suited to conventional supercomputers.

What the Sandia researchers discovered is that PDEs can be reformulated as statistical sampling problems, random walks, in a way that maps naturally onto neuromorphic hardware. Each random walk step becomes a spike event. The accumulated statistics of many random walks, tracked across the artificial neurons, converge to the solution of the PDE. The approach is not applicable to every PDE in every context, the tradeoffs between accuracy, speed, and energy vary by problem, but it demonstrated that neuromorphic hardware can be useful in scientific computing contexts that had not previously been considered. Brad Aimone, the lead researcher, described the finding as opening up applications ranging from radiation transport and molecular simulations to biology modelling and particle physics.

A separate milestone arrived at the SCOPE workshop at the International Conference on Learning Representations in April 2025, where Jason Eshraghian and colleagues at the University of California, Santa Cruz, presented the first large language model implemented on neuromorphic hardware, specifically, a model adapted to run on a Loihi 2 chip. The model matched the accuracy of a comparable GPU-based LLM while using half the energy. This is a proof of concept, not a deployment-ready system: current neuromorphic hardware lacks the scale to run the very large models that dominate commercial AI. But it established that the inference phase of language models, the computation required to generate a response, is within the reach of neuromorphic approaches, and that the energy savings are real.

What changes when intelligence becomes cheap to run

The energy implications of neuromorphic computing are the most straightforwardly quantifiable consequence of its adoption. According to the IEA’s January 2025 Energy and AI report, data centres consumed around 415 terawatt-hours of electricity globally in 2024 — roughly equivalent to France’s total annual electricity consumption — and AI workloads are driving nearly half of the net growth, with AI-focused server consumption projected to grow at approximately 30% annually. Data centres are a meaningful fraction of global electricity demand, and the trajectory has attracted regulatory attention: the EU AI Act requires providers of general-purpose AI models to document energy consumption in their technical disclosure obligations, while California’s SB 253 requires large companies to report Scope 3 emissions including from computing.

Neuromorphic hardware does not address the training cost problem, which is where the largest energy expenditure currently occurs. What it addresses is inference, running trained models at the edge, in devices, in real time, which is where AI energy consumption is distributed most broadly across the economy. A neuromorphic processor handling sensory data in a robot, a vehicle, a medical monitoring device, or an industrial sensor consumes a fraction of the energy of a GPU doing the same work. At the scale of millions or hundreds of millions of deployed AI systems, those fractions compound into a very large number.

Data centres consumed around 415 terawatt-hours of electricity globally in 2024 — roughly equivalent to France’s annual consumption — and AI is driving the fastest growth. The training cost is concentrated; the inference cost is distributed across every deployed device. Neuromorphic hardware addresses the distributed cost, which is where the long-term growth lies.

The robotics implications are the most tangible near-term consequence. Alex’s robot, running visual processing on a neuromorphic chip rather than a GPU, does not drain its battery in eight hours. ANYbotics’ ANYmal D Neuro, a quadruped inspection robot for industrial environments, pairs Loihi 3 with Prophesee event cameras. Scheduled for commercial release in Q3 2026, it is reported to achieve 72 hours of continuous operation on a single charge, compared to approximately eight hours for the previous GPU-powered version.

That ninefold improvement in operating time is not a marginal engineering gain. It changes what kinds of inspection tasks are operationally feasible, what environments can be reached, and whether robots of this kind are economically viable for the organisations that need them.

Event-based cameras paired with neuromorphic processors represent a specific and significant application. Unlike conventional cameras, which capture full frames at fixed intervals regardless of what is happening, event cameras fire at each pixel only when the light level at that pixel changes. The result is a stream of sparse, precise events, equivalent to a neuromorphic spike, rather than a dense stream of frames.

The latency between a change in the visual field and the camera’s output is measured in microseconds rather than the milliseconds of frame-based cameras. Paired with neuromorphic processors that handle spike-based data natively, this architecture enables real-time visual processing with sub-millisecond latency and minimal energy consumption. Prosthetic limb control requires exactly this capability: responding to visual and sensory information faster than conventional frame-rate processing allows.

The longer-term consequence that is harder to quantify is what becomes possible when the energy cost of intelligent processing falls by orders of magnitude. The constraint on how many intelligent devices can be deployed is currently, in significant part, the energy required to run them. A sensor network that monitors agricultural fields for disease and drought stress cannot scale to the numbers that would change how farming works because the energy cost per device is too high. Neuromorphic hardware addresses that constraint directly.

What the energy gap actually means

The energy case for neuromorphic hardware is already demonstrated in hardware that exists today. Alex’s robot that can run for 72 hours rather than eight is a more useful tool — not marginally, but categorically. The inspection task that required three battery swaps and a crew of technicians can now be completed in a single continuous deployment. The medical monitoring device that previously required a wired connection or a large battery pack can now operate on coin cells for months. These are not theoretical capabilities.

The more interesting question is whether neuromorphic hardware expands the class of problems that intelligent systems can address, or whether it primarily makes existing AI applications more energy-efficient. The Sandia PDE result suggests the former: a problem domain, physics simulation, that was not previously considered within neuromorphic hardware’s reach has been demonstrated to be tractable. The neuromorphic LLM result points in the same direction: the inference phase of language models, which was considered a GPU-native workload, can run on neuromorphic hardware at half the energy cost. If the range of applicable problems continues to expand as researchers find new mappings between computational domains and spike-based processing, the technology’s significance extends well beyond edge AI efficiency.

The constraint that most limits wider deployment is not the hardware, it is the software. Conventional software runs on von Neumann processors, and the entire software ecosystem has been built around that architecture since the 1940s. Writing code for neuromorphic hardware requires thinking in terms of spiking neural networks and event-driven computation, rather than the loops and function calls that software engineers learn first. Intel’s Lava framework and SpiNNaker 2’s hybrid deep-learning support are steps toward narrowing that gap. Until they succeed, neuromorphic computing will remain primarily a research and specialist application domain rather than the broadly deployed alternative to GPU-based edge AI that its energy advantages suggest it could become.

Where the demonstration ends

This is not the most dramatic AI story being told in 2026. It is not the largest model, or the most capable system, or the technology that attracts the most press attention. Neuromorphic computing is the hidden innovation for a reason: it works differently from everything else, it has been advancing steadily since around 2017 in university laboratories and government research programmes, and its significance is not yet visible in the mainstream conversation about what AI is becoming.

AI’s energy consumption is growing at rates that have attracted regulatory attention and genuine concern from energy system planners. Neuromorphic hardware addresses the inference cost, not the training cost. Whether that is enough to make a material difference to the AI energy problem, and over what timescale, is a question the deployment trajectory will answer, not the laboratory results.

The Sandia PDE result showed that a problem domain assumed to be outside neuromorphic computing’s reach turned out to be tractable. Partial differential equations underpin climate modelling, drug discovery, structural engineering, and financial simulation. What it would mean for those fields if the energy cost of running large-scale simulations fell by orders of magnitude is not yet clear but the assumption that the problem was out of reach has now been tested.

The programming difficulty of neuromorphic hardware is one of the main barriers to wider adoption. The GPU dominated AI not only because of its processing power but because CUDA, NVIDIA’s programming framework, made GPU computing accessible to software engineers without hardware expertise. What the equivalent for neuromorphic computing looks like, and who is best placed to build it, are open questions with commercial as well as technical consequences.

The robotics application, a quadruped inspection robot running for 72 hours rather than eight, is a demonstration in a commercial industrial context. Other sectors have AI applications whose deployment has been constrained primarily by the energy cost of the processing. The question is which ones, and whether neuromorphic hardware removes the constraint in time to matter.

Alex’s robot runs for 72 hours on a single charge. The visual processing happens on a chip that fits in the palm of a hand and uses less power than a bicycle lamp. The images it sees are not processed as frames — they are processed as events, as changes, the way a biological retina has always processed them.

The brain uses 20 watts. Everything we have built to approximate its capabilities uses enormously more. Neuromorphic computing is the most direct attempt to close it.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. This is part of the Emerging Science & Technology series.

If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

A surgeon is treating a blocked artery. The traditional approach involves threading a catheter to the blockage and expanding a stent using a balloon, an elegantly simple procedure, but one that requires precision inflating of the balloon to exactly the right pressure in a precise anatomical location. Too much force and the artery wall is damaged. Too little and the stent does not open properly. The clinician is managing a mechanical problem with a mechanical tool, and the margin for error is narrow.

In the operating theatre next door, a different kind of stent is being placed. This one is made from nitinol - a nickel-titanium alloy with an unusual property. It has two stable crystal structures, one that exists at low temperatures and one that exists at higher temperatures, and it can be programmed to remember a specific shape in its high-temperature form. The stent is manufactured in its intended open configuration, then cooled and compressed into a delivery catheter narrower than a matchstick.

Once positioned at the blockage site, it is released. As it warms to body temperature, its crystalline structure undergoes a phase transition. It opens into the shape it was programmed to remember. No balloon. No precise mechanical inflation. The material does the work.

This is shape memory, and it is one of the most counterintuitive properties in materials science: a material that can be deformed, held in a different configuration, and then recover its original form when a trigger is applied. The trigger is usually temperature, but it can also be light, a magnetic field, moisture, or an electrical current. The range of materials that exhibit some version of this property — alloys, polymers, hydrogels, liquid crystal elastomers — is broad. The range of things they can do, at a moment when additive manufacturing is giving engineers the tools to program shape changes in three dimensions, is expanding rapidly.

Why materials that change shape are hard to build

Shape memory alloys like nitinol work through a reversible phase transition in their crystalline structure. Below a transition temperature, the alloy exists in a phase called martensite, a more flexible, deformable structure. Above it, it exists in austenite, a more ordered, rigid structure with a different geometry. The trick is that you can program the austenite geometry: heat the material to a high temperature, hold it in the desired shape, and the crystal structure locks that geometry in as its remembered form. Cool it, deform it, and when you heat it again, back it comes. The recovery can be almost instantaneous, and the recoverable strain, the amount of deformation the material can undergo and still return from, is around 8%, far beyond what conventional metals can manage without permanent damage.

Nitinol has been in clinical use since the 1980s. It is the standard material for self-expanding vascular stents, the type used in peripheral arterial disease, carotid artery stenosis, and neurovascular applications. Orthodontic arch wires made from nitinol apply gentle, continuous corrective force as they try to return to their programmed shape, a force that moves teeth without the patient needing to return for frequent tightening. Bone anchors, spinal correction devices, and neurovascular occluders for treating aneurysms all exploit the same phase transition. It is a mature commercial technology that millions of patients encounter every year, in most cases without knowing the material in their bodies is doing something physically unusual.

Shape memory polymers, a broader and more chemically diverse class of materials, work on different principles, and their design space is considerably wider. A shape memory polymer has two distinct phases: a soft segment that can be deformed and a hard segment that stores the programmed shape. Deform the material above its transition temperature, cool it while deformed, and it holds the temporary shape. Heat it again and the stored shape is recovered. The transition temperature can be tuned across a wide range by adjusting the polymer chemistry, which makes shape memory polymers adaptable to applications where the trigger needs to be body temperature, or a slightly elevated temperature, or a specific heat source.

The third major class is the soft, fluid-like materials: hydrogels that swell or contract in response to moisture or pH, liquid crystal elastomers whose molecular ordering changes under heat or light, and magnetically responsive composites that deform when a magnetic field is applied. These materials are the substrate of most research into soft robotics: mechanisms that move and grasp and bend without rigid components, actuated by the properties of the material itself rather than by motors or pneumatics. A soft robotic gripper made from a shape-responsive hydrogel does not require a power source to grip it requires the right chemical environment. A structure made from liquid crystal elastomers can fold and unfold in response to light, performing actuation that would otherwise require electronics and motors with no moving parts at all.

The gap between what the laboratory shows and what reaches patients

The most mature commercial applications of shape memory materials — nitinol stents, orthodontic wires, endoscopic instruments — are all in medicine, and they share a common property: the transformation happens once, or a small number of times, in a controlled environment where the trigger is predictable and the required recovery force is modest. The stent opens once. The orthodontic wire applies a known corrective force. The deployment is the function.

The applications that attract the most research attention — soft robots that repeatedly cycle through configurations, implants that actively adapt to changing anatomical conditions over years, aerospace structures that morph their shape mid-flight — require something more demanding: reliable, repeatable transformation over thousands or millions of cycles, in environments where the trigger must be controlled precisely, and where the forces involved may be substantially larger than what current materials can generate reliably. The gap between single-deployment devices and actively programmable adaptive structures is the central engineering challenge of the field.

4D printing, 3D printing with stimuli-responsive materials, where the fourth dimension is the shape change that happens after fabrication, is the technology most likely to bridge this gap. The concept is straightforward: instead of printing a static object, you print a structure whose geometry will change when a trigger is applied. The structure folds from flat to three-dimensional when heated, or a scaffold for tissue engineering gradually dissolves while cells grow around it, or an airfoil changes its cross-section profile in response to airspeed. The research literature on 4D printing has grown rapidly since 2013, when the term was coined, and functional demonstrations cover origami folding, self-assembling robotic grippers, drug delivery capsules that open in response to specific pH, and textile structures that change their insulating properties with temperature.

The honest assessment of 4D printing in 2026 is that it is a legitimate and growing research area producing functional demonstrations, while many of the more ambitious applications — adaptive aircraft wings, self-assembling structures at the scale of buildings, fully programmable soft robots — remain laboratory achievements whose path to manufacturing at useful scale is not yet clear. The cost of 4D-printable materials is high. The precision required to encode specific shape-change programmes into printed structures is demanding. Regulatory validation for medical applications is lengthy. And the fatigue behaviour of printed shape-memory structures, how they perform after being cycled through their transformation thousands of times, is not yet fully characterised for most material and geometry combinations.

What changes when materials can be programmed to move

The most immediate consequence of mature programmable materials technology is in minimally invasive medicine. Nitinol has been in clinical use for forty years, but the direction of travel is toward more complex, more adaptive devices. Shape memory polymer stents that gradually soften and dissolve after their work is done, avoiding the need for a second procedure to remove a device that is no longer needed, are moving through preclinical research. Endoscopic instruments that change shape to navigate complex anatomy, scaffolds for regenerative medicine that match the stiffness of the tissue they support and then degrade as new tissue grows, implants that respond to changing load conditions over years rather than being static once placed: these are the next generation of applications, and several are in early clinical development.

The access implications of less invasive medicine are substantial. Procedures that currently require open surgery, with its associated risk, recovery time, and specialist facility requirements, become safer and more widely deliverable when the device can reach the target through a catheter and deploy itself. A self-expanding stent placed without general anaesthesia and a balloon catheter can be delivered in a setting that could not safely perform the balloon-inflation procedure. The further the technology moves toward single-deployment devices that work autonomously, the more broadly accessible the treatment becomes.

In aerospace and defence, the application is deployable structures. Satellites are launched folded and must unfold reliably once in orbit, a mechanical process that has historically been a significant source of mission failures. Shape memory alloy actuators for deploying solar panels and antenna structures are already in use because they are lighter, simpler, and more reliable than the spring-loaded or motorised mechanisms they replace. The same logic extends to morphing aircraft surfaces: a wing that changes its cross-section profile in response to flight conditions without requiring a hydraulic actuation system is lighter, less mechanically complex, and harder to damage. The weight savings in aerospace translate directly to fuel efficiency or payload capacity.

The soft robotics implications are harder to characterise because the field is still in early demonstration. A gripper made from a responsive hydrogel, a soft robot that moves through a pipe by swelling and contracting in response to chemical signals, these are functional laboratory demonstrations. The path from demonstration to deployment depends on whether the materials can be manufactured consistently, whether they can survive the operational cycles required, and whether the design tools exist to encode specific behaviours reliably. Those are engineering problems rather than scientific ones, and they are tractable, but they have not yet been solved at scale.

The gap between remembering and adapting

The clearest value of programmable and shape-memory materials is the human intervention they displace. The stent that deploys itself, the implant that adapts to changing anatomy, the satellite structure that unfolds without a motorised mechanism, in each case, the material carries some of the work that a human action or a mechanical system previously had to perform. That displacement is concrete and verifiable, and in the most mature applications it has already been demonstrated in millions of clinical procedures and commercial deployments.

Whether programmable materials genuinely extend what is possible in medicine and engineering, or whether they primarily offer a more elegant route to outcomes that conventional approaches already achieve, depends on the application. For the self-expanding stent, the answer is clear: the procedure is safer, simpler, and more widely deliverable than balloon-expansion because the material does what the balloon previously had to do. For 4D-printed soft robots, the answer is more speculative: the demonstrations are impressive, but the engineering pathway to reliable deployment in demanding environments is not yet complete. The technology is at the threshold where the scientific imagination of what might be possible has outrun the engineering demonstration of what reliably works.

Programmability is the entire point of this technology, which makes the next question the most interesting: do the design tools and manufacturing processes exist to specify adaptation reliably? A material that adapts its shape or stiffness to the specific conditions it encounters — the anatomy of a particular patient, the load profile of a specific flight condition, the chemical environment of a specific tissue — is doing something different from a material that merely remembers. 4D printing is moving in that direction, but the precision of shape-change encoding and the consistency of material behaviour across production batches remain active engineering challenges. Until they are addressed, the adaptation is programmed at design time rather than responsive at deployment time.

Where the demonstration ends and the engineering begins

The stent in the artery opens and holds its shape without anyone having to do anything after deployment. The satellite solar panel unfolds reliably in a vacuum, with no motor and no human hand. These are the quiet applications of a technology whose more dramatic demonstrations are still being developed. The research literature is full of soft robots and morphing wings and self-assembling structures. Most of them are not yet ready for the world outside the laboratory. The applications that are ready, and that are already in widespread use, work because the underlying property is real: a material that has been taught what it should be and returns to it when conditions are right.

• Nitinol is the standard material for self-expanding vascular stents, and millions of patients carry one without knowing the metal in their artery is doing something physically unusual. What else in the built and medical environment uses materials whose interesting properties are invisible to the people they serve?

• Shape memory polymer implants that dissolve after their work is done could eliminate the second procedure currently required to remove devices that are no longer needed. What other medical interventions involve permanent implants whose permanence is a limitation rather than a feature?

• The most ambitious 4D printing demonstrations — adaptive aircraft surfaces, self-assembling structures, soft robots that move without motors — are genuine laboratory achievements whose path to manufacturing scale is not yet clear. What is the engineering bottleneck most likely to limit how quickly this technology moves from research to deployment?

• Minimally invasive procedures that rely on shape memory devices are more widely deliverable than open-surgery equivalents. If the next generation of programmable medical implants reduces the specialist skills and facility requirements for a procedure, what does that mean for where complex medical interventions can be performed and for which patients they become accessible?

The question the field is working toward is whether materials can be taught something more complex: not just one remembered shape, but a response to changing conditions. A material that does not just remember — but adapts.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. This is part of the Emerging Science & Technology series.

If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

A small town in northern India has solar panels, enough of them to generate more power than the community uses on a summer afternoon. What it does not have is a way to store that power so it can be used in the evening, when the panels are dark and demand is highest. Until 2025, the economics of battery storage made this unaffordable. The batteries that could do the job are made from lithium, which is mined primarily in South America and Australia, refined in China, and priced at a level that makes grid-scale storage out of reach for communities that are not already well-capitalised.

In a factory in Fujian province, China, a different kind of battery is now in production. CATL — the world’s largest battery manufacturer, holding 39.2% of the global electric vehicle battery market in 2025 — launched its Naxtra sodium-ion brand in April 2025 and confirmed large-scale deployment across passenger vehicles, commercial vehicles and grid storage for 2026. The chemistry is different from lithium-ion. The cathode uses sodium iron hexacyanoferrate, a form of Prussian white — a material based on iron, carbon and nitrogen rather than the lithium, cobalt and nickel that dominate conventional batteries. The electrolyte carries sodium ions rather than lithium ions. The operation is the same: charge the battery, ions move in one direction; discharge it, they move back. The material doing the moving is what changes.

Sodium is not a rare element. It makes up roughly 2.4% of the Earth’s crust by mass — about 23,600 parts per million, against 20 parts per million for lithium. IRENA puts the ratio at around 1,000 times more abundant. It is found on every continent, in seawater, in salt deposits, in rock formations, distributed across the world with none of the geographic concentration that makes lithium a geopolitical instrument. A battery built from sodium is, in principle, a battery that any country with access to basic industrial chemistry can manufacture domestically. Whether that principle translates into practice — into affordable energy storage for the communities that most need it — is the story of sodium-ion batteries in 2026.

The case for sodium, and its honest limits

Sodium-ion batteries work through the same basic electrochemical principles as lithium-ion. Ions shuttle between cathode and anode through an electrolyte, storing energy on the way in and releasing it on the way out. The physics is essentially identical. Sodium ions are larger than lithium ions. They require electrode materials with different crystal structures to accommodate them. But the engineering challenge is not different in kind, only in the specific materials required.

The practical advantages over lithium-ion are concrete. The first is cold-weather performance. Lithium-ion batteries can lose 30 to 40% of their effective capacity at minus 20 degrees Celsius. CATL’s Naxtra sodium-ion cells retain around 90% of their power at minus 40 degrees Celsius and operate across the full range from minus 40 to plus 70 degrees Celsius. For communities in cold climates — northern India in winter, Siberia, the Canadian prairies, Nordic countries — this is not a marginal improvement. It is the difference between a battery that works and one that does not.

The second is safety. Sodium-ion batteries have a lower risk of thermal runaway, the chain reaction that causes lithium-ion batteries to catch fire, because they can be transported and stored at zero charge state without degradation, which lithium-ion batteries cannot safely do. They are also less susceptible to the internal short circuits that arise from lithium dendrite formation during rapid charging. These safety properties matter for stationary grid storage, where fire risk is a serious operational and regulatory concern.

The third advantage is the raw material geography. Lithium reserves are concentrated primarily in Australia, Chile and Argentina — the so-called lithium triangle. Cobalt, used in many high-energy-density lithium-ion chemistries, comes overwhelmingly from the Democratic Republic of Congo, with around 72% of global mine output and serious concerns about the conditions under which it is extracted. Graphite, used in lithium-ion anodes, is refined predominantly in China, which also processes most of the world’s cobalt. Sodium-ion cathode materials based on iron compounds and sodium have no comparable concentration. The supply chain vulnerability that runs through every lithium-ion battery manufactured today does not, in principle, run through a sodium-ion equivalent.

The limitations are real and significant. Energy density is lower: lithium-ion cells at the high end reach 300 Wh/kg or more; CATL’s Naxtra achieves 175 Wh/kg, comparable to lithium iron phosphate batteries but below the high-nickel chemistries used in premium electric vehicles. For a passenger car where range and weight compete directly, this gap matters. For a city runabout, a two-wheeler, a commercial van, or a stationary grid storage installation where weight is not the primary constraint, it does not. The applications best suited to sodium-ion are those where cost, safety, cold-weather performance and supply chain resilience matter more than maximum energy density per kilogram.

The other complication is the lithium price cycle. Lithium carbonate prices spiked dramatically in 2022 as EV demand surged, making sodium-ion’s cost advantage appear decisive. Since then, lithium prices have fallen more than 70%, narrowing the gap significantly. IRENA’s November 2025 technology brief projects that sodium-ion cell costs could fall to around $40 per kilowatt-hour at scale, against LFP lithium-ion at roughly $70 per kilowatt-hour in 2025. That cost gap, if it materialises in commercial production, is real. But it is not yet demonstrated at scale, and if lithium prices remain low or fall further, the economic case for sodium-ion rests more on supply chain resilience and specific performance characteristics than on direct cost comparison.

Who is building it, and who is not

China controls around 96% of current global sodium-ion production capacity. This is not an accident of geography — it is a consequence of the same industrial policy and manufacturing scale that gave China dominance in lithium-ion. CATL launched the Naxtra brand in April 2025 and confirmed large-scale deployment for 2026. BYD broke ground on a 30 GWh sodium-ion facility in Xuzhou, Jiangsu, with a third-generation platform reportedly capable of up to 10,000 cycles. Global announced production capacity across sodium-ion projects had reached around 370 GWh by early 2026, with over $20 billion in committed investment more than 90% of it in China.

For countries outside China, the picture is more complicated. Faradion, the British sodium-ion company that was among the pioneers of the chemistry, was acquired by Reliance New Energy in stages from 2022 — initial majority stake at an enterprise value of around £100 million plus £25 million in growth capital, with full 100% acquisition completed in October 2024. The technology now sits in the Indian conglomerate’s energy portfolio rather than in a domestic European supply chain. Germany’s Federal Ministry of Education and Research committed €14 million to the SIB:DE FORSCHUNG programme in February 2025, a 21-partner consortium coordinated by BASF to industrialise sodium-ion technology. Peak Energy in the United States is deploying sodium-ion grid storage at substations and industrial sites. These are real commitments, but at different scales from China’s industrial bet. The pattern is familiar from the lithium-ion story: China invested early, moved fast, and built manufacturing scale that competitors are now trying to match from a standing start.

The question this raises is whether sodium-ion batteries will replicate the lithium-ion geopolitical structure in a different material, or whether the abundant and distributed nature of sodium changes who can participate in the supply chain. The encouraging sign is that sodium-ion cathode materials based on Prussian white and layered oxide compounds do not require the same refined critical minerals as lithium-ion cathodes; the dependence on cobalt refining capacity concentrated in China is absent. A country with basic industrial chemical manufacturing capacity and a sodium source — which describes most of the world — can in principle build sodium-ion cells domestically. Whether the technology transfer and investment happen before China locks in the same kind of manufacturing dominance it achieved in lithium-ion is the race that the next five years will decide.

What cheaper, more abundant storage changes

The most direct consequence of viable sodium-ion grid storage is for communities like the small village in India, places with renewable generation capacity and no affordable way to store it. Stationary energy storage is the dominant near-term application for sodium-ion, with industry estimates putting it at roughly three-quarters of current deployment, and this proportion reflects where the technology’s cost and performance characteristics are most immediately compelling. A grid storage installation does not care about weight. It cares about cost per kilowatt-hour, cycle life, safety and operating temperature range. Sodium-ion is competitive on all four. IRENA suggests that at projected costs of $40 per kilowatt-hour, sodium-ion grid storage becomes economically viable for small-scale community storage in settings where it was not before.

The two- and three-wheeler electric vehicle market is the other near-term application with direct access implications. In India, Indonesia, Vietnam and across South and Southeast Asia, electric two- and three-wheelers are the primary personal transport technology for hundreds of millions of people. The batteries in these vehicles are typically small, the range requirements are modest, and the total cost of ownership is the overriding purchase criterion. A sodium-ion battery pack that is meaningfully cheaper than its lithium-ion equivalent, safer to operate in high ambient temperatures, and not dependent on a supply chain that must pass through constrained lithium processing capacity, is precisely the product these markets need. HiNa Battery, the Chinese sodium-ion specialist, is already deploying in low-speed EV and electric two-wheeler markets. CATL’s Naxtra is targeted explicitly at affordable entry-level vehicles.

The grid storage consequence for renewable energy integration is the larger systemic argument. Solar and wind are now the cheapest sources of new electricity generation in most of the world, but they generate power when the sun shines and the wind blows rather than when demand is highest. Affordable grid-scale storage is the component that allows renewable generation to displace fossil fuels in the baseload role. Every additional gigawatt-hour of affordable storage installed moves that substitution further along. Sodium-ion batteries do not solve this problem alone — long-duration storage at grid scale requires other technologies — but they address the four-to-eight-hour storage window that is the most commercially immediate requirement.

There is also a strategic security argument that runs alongside the energy access argument. The dependence of the current clean energy transition on lithium, cobalt, nickel and graphite supply chains concentrated in a small number of countries, with processing and manufacturing concentrated in China, is a geopolitical risk that the US Inflation Reduction Act, the EU Battery Directive and equivalent policies in India and Japan are all partly designed to reduce. A battery chemistry whose core materials are distributed globally, and whose cathode compounds do not require refined critical minerals, addresses that dependency at the source rather than by relocating it.

What the abundant material cannot guarantee

The test for sodium-ion batteries is not whether the technology works — it does — but whether the economics reach the small town, and the thousands of places like it that have renewable generation and no affordable way to store it. The CATL Naxtra is in passenger cars, commercial vehicles and grid storage in 2026. At IRENA’s projected cost of $40 per kilowatt-hour, the answer to that access question becomes different from the answer at $70 per kilowatt-hour. The question is when, not whether.

The cold-weather performance genuinely expands where electric vehicles and grid storage are viable: Finland’s electric bus fleet and northern India’s community storage system are real use cases that lithium-ion serves poorly. The supply chain geography expands who can build battery manufacturing capacity domestically. But the core function, storing electricity, is identical to what lithium-ion already does. The difference is in the access, not in a new capability.

The most important question for the energy transition’s long-term equity is whether sodium-ion batteries enable countries with no lithium reserves and no existing position in the battery supply chain to participate in manufacturing the storage component of their own energy systems. If production concentrates in China in the same way lithium-ion production has, the chemistry changes but the geopolitical structure does not. The technology makes the more equitable outcome possible. Whether it happens depends on investment decisions, industrial policy and technology transfer arrangements that are being made right now.

What the supply chain question comes down to

Salt is everywhere. The clean energy transition requires storage, and the storage currently required runs through supply chains that are among the most geopolitically concentrated in the global economy. A battery whose core material is found across every continent, whose cathode does not require cobalt or nickel, and whose manufacturing technology is accessible to any country with basic industrial chemistry, that battery addresses a structural problem in the energy transition that the technology everyone has been investing in cannot solve.

It is also not a guarantee. CATL controls 39.2% of the global EV battery market and around 96% of current sodium-ion production capacity. The abundant material is in the ground everywhere. The factory that processes it is in Fujian.

• Sodium is around 1,000 times more abundant than lithium and found everywhere. The supply chain for sodium-ion cathode materials does not pass through the same chokepoints as lithium-ion. Yet China currently controls 96% of sodium-ion production capacity. Does a more abundant material guarantee a more distributed supply chain, or does industrial scale reproduce concentration regardless of the underlying geology?

• Lithium prices fell more than 70% between 2022 and 2025, narrowing sodium-ion’s cost advantage at precisely the moment it was supposed to become decisive. What does that price history suggest about whether sodium-ion’s commercial case rests on a durable structural advantage or on the volatility of a competing material’s price?

• Grid-scale storage is the component that allows renewable generation to displace fossil fuels in the baseload role. Sodium-ion batteries are most immediately competitive in stationary storage. If sodium-ion costs reach $40 per kilowatt-hour at commercial scale, IRENA’s projection, what does that change for the economics of community-scale renewable energy storage in the places that have been priced out of lithium-ion solutions?

• The town in northern India has solar panels and no affordable storage. If sodium-ion batteries reach the projected cost target and if the manufacturing capacity is distributed rather than concentrated, their situation changes. What other technologies in this series have a similar relationship to the energy transition — where the science is not the constraint, but the economics, the supply chain and the industrial policy are?

The technology can change who participates in the energy transition. Whether it does depends on decisions being made now about investment, industrial policy and who gets access to the manufacturing knowledge. Those decisions are not technical. They are political. And they are being made while the solar panels generate more power than the town can use, with nowhere to put it until evening.

If this raised something worth thinking through further, The Next Evolution at neilcatton.substack.com is where this conversation continues. My books The Next Evolution, The Cognitive Crucible and The Shadow System are available on Amazon.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. This is part of the Emerging Science & Technology series.

If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.

Jordan has worked the same assembly line at an automotive plant for eleven years. The work is repetitive — reach, pick, place, repeat — and over time it has taken a toll. A shoulder injury two years ago put Jordan on light duties for three months. A wrist problem before that. The plant’s occupational health team know Jordan well.

Earlier this year, a robot started working the same line. It is roughly human-shaped, two arms, an upright torso, roughly Jordan’s height. It picks up parts. It places them. It does not call in sick. It does not get injured. And it does not, so far, do anything Jordan cannot do.

Jordan’s reaction, when asked, is measured. “It’s doing the stuff nobody wants to do anyway,” Jordan says. “The boring bits. The heavy bits. I’m doing more of the interesting work now.” Then a pause. “For now.”

That pause carries a lot of weight. We are at the earliest stage of a shift in the physical workplace that could be as significant as the arrival of the personal computer in the office, and the questions it raises are not really about the technology. The machines are arriving. The question is whether the systems around them — the training programmes, the wage structures, the safety nets — are arriving at the same pace.

From Demo to Deployment

For most of the past decade, humanoid robots existed primarily as demonstrations. A company would release a video. The robot would walk across a stage, pour a glass of water, fold a t-shirt. Audiences would applaud. Experts would point out, sometimes in public, that the footage was carefully controlled, that half the tasks were teleoperated, that the gap between a choreographed showcase and a functioning workplace colleague was enormous.

That scepticism was largely justified. And then, quietly, things changed.

In late 2025 and into 2026, the demos gave way to deployments. Tesla’s Optimus Gen 3 robots began working autonomously inside the company’s Fremont and Austin facilities — moving boxes, sorting parts, handling repetitive material tasks. Figure AI’s robots started shifts at a BMW manufacturing plant in South Carolina. Agility Robotics deployed its Digit robot inside Amazon fulfilment centres. Apptronik’s Apollo went into a Mercedes-Benz facility. These are not proof-of-concept trials. They are operational.

The numbers are still modest. Across all deployments globally, the humanoid robot population in real workplaces is measured in the hundreds to low thousands, not the hundreds of thousands. But the trajectory matters as much as the current state. The International Federation of Robotics reported that the global market value of industrial robot installations hit an all-time high in 2026. Goldman Sachs revised its forecast for the humanoid robot market to reach $38 billion by 2035 — six times its earlier estimate.

The demos gave way to deployments. The gap between a choreographed showcase and a functioning workplace colleague has closed faster than most experts predicted.

The acceleration has several causes. Battery life improved, modern units can operate 16 to 20 hours before recharging, making two-shift factory schedules viable. Actuators got cheaper and more precise. And crucially, the AI underneath the robots matured. Large language models gave robots a way to understand plain-language instructions without bespoke programming. What was once a specialist task, deploying an industrial robot, is beginning to look more like managing a member of staff.

The World Is Built for Humans — and That Is the Point

The question that has always haunted industrial automation is: why humanoid? For decades, the answer from robotics engineers was: there is no good reason. A robotic arm bolted to a single point on an assembly line is faster, more precise and far cheaper than any machine trying to walk on two legs. If you know exactly what task needs doing and where, purpose-built automation wins every time.

The humanoid form only makes sense when you don’t know exactly what task needs doing, or where, when the environment is unpredictable and the work is varied. Which describes most of the world beyond the carefully designed production line. Warehouses where the layout changes. Construction sites. Care homes. Kitchens. The human body evolved for generalism, and humanoid robots are a bet that the best way to navigate a world built for humans is to be shaped like one.

This is why the most significant humanoid deployments in 2026 are not in purpose-built robotics facilities. They are in places already designed around human workers: factory floors with standard-width aisles, shelving at human reach height, tools sized for human hands. Boston Dynamics’ Atlas, developed in collaboration with Toyota Research Institute, began testing in Hyundai facilities for exactly this reason — not because the facility was built for robots, but because it was built for people and the robot could work within it without expensive retrofitting.

Who Is Building Humanoid Robots in 2026

• Tesla (Optimus Gen 3) — Deploying in own factories; targeting ~$20,000–$30,000 at scale

• Figure AI (Figure 03) — Pilots at BMW; Helix AI model for full-body autonomy

• Agility Robotics (Digit) — Amazon fulfilment centres; owned by Hyundai

• Apptronik (Apollo) — Mercedes-Benz manufacturing; focus on hazardous tasks

• Boston Dynamics (Atlas) — Hyundai pilots; NVIDIA collaboration for Large

• Behaviour Models 1X Technologies (NEO) — Consumer-facing; $20,000 pre-orders open; 2026 delivery target Chinese Big 5 — Unitree, Agibot, Leju, Fourier, Huawei.

China holds ~90% of 2025 global shipments

China deserves particular attention here. Chinese manufacturers shipped roughly 13,000 humanoid robot units in 2025, representing approximately 90% of global shipments. Agibot, largely unknown outside specialist circles, shipped its 10,000th cumulative unit in early 2026, with half of those deliveries in the three months prior. The Chinese government’s industrial policy has provided both demand, state-owned enterprises buying early, and supply chain advantage, with domestic manufacturers controlling components that every other country’s robotics industry depends on. The race being run is not purely commercial. It is geopolitical.

The Pause in Jordan’s Answer

The labour economics of humanoid robots sit in uncomfortable territory, and the honest answer is that no one fully knows how they will play out.

The optimistic case is straightforward: there are not enough workers. Bain & Company projects a global shortage of nearly eight million manufacturing workers by 2030, driven by ageing populations and declining birth rates in major industrial economies. The United States currently has around 8.1 million job openings but only 6.8 million unemployed workers to fill them. Robots that can handle the dull, dirty and dangerous work — the work that most injures bodies and most struggles to find takers — are not taking jobs. They are filling a gap that the workforce cannot close on its own.

Humanoid robots could also change the character of the work that remains. If a robot handles the repetitive, high-injury tasks on an assembly line, the human workers alongside it shift toward quality oversight, maintenance, exception handling, work that draws on judgement rather than endurance. Some early evidence from automotive deployments points in this direction. Workers in facilities with robot colleagues report fewer repetitive strain injuries and more varied daily responsibilities.

The pessimistic case is also straightforward: McKinsey’s Global Institute estimates that automation could displace between 400 and 800 million jobs worldwide by 2030, forcing roughly 375 million workers — 14% of the global workforce — to change occupations entirely. Cost reductions in humanoid hardware have exceeded projections: Goldman Sachs reported a 40% year-on-year drop in manufacturing costs between 2022 and 2024, against earlier predictions of 15 to 20%. At $16,000 to $20,000 per unit, a humanoid robot costs less than a year of minimum wage in the United States. The economics of replacement are beginning to work.

The honest position is somewhere between these two narratives, and it depends enormously on what happens in the transition period, the years between initial deployment and mass adoption. History offers mixed lessons. When ATMs arrived in banks, most analysts predicted the end of the bank teller. The opposite happened, at least initially: more ATMs meant more branches, more branches meant more tellers. But that transition unfolded over decades and in a sector with relatively strong institutional structures. The manufacturing workers most at risk from humanoid robots are often in regions and industries where those structures are thin.

The workers who will bear the cost of a poorly managed transition are not abstract. They are people like Jordan, who has built a life and a household budget around a role that may look quite different in five years, and who has limited access to the retraining infrastructure that a smooth transition would require.

Three Tests the Technology Has Not Yet Passed

Humanoid robots, at their best, are assistive — they help people do physically demanding or hazardous work without injury. In the deployments that work well, that is precisely what they do. But assistive does not mean the technology has been designed with the human it affects most clearly in view.

The augmentive test is harder to satisfy. A technology augments when it genuinely adds something — capability, safety, time — rather than simply substituting. The question worth asking of any humanoid deployment is whether the workers alongside the robots end up with more valuable, more interesting, more sustainable work. In some cases, the answer appears to be yes. In others, the robots are doing simpler tasks while the workforce shrinks, the augmentive promise has been made, but the evidence is not yet there.

The adaptive test is the one the industry has barely begun to address. A robot that is deployed the same way in every facility, regardless of the workforce composition, the local labour market, the age profile of the existing workers, or the availability of retraining programmes, is not adaptive. It is standardised. The most honest assessments of current humanoid deployments acknowledge that the machines are advancing much faster than the governance frameworks that should accompany them, the retraining provision, the safety standards, the liability frameworks, the worker involvement in deployment decisions.

The International Federation of Robotics noted in its 2026 trend report that the close cooperation with employees in implementing robots plays a crucial role in ensuring acceptance. That observation sounds obvious. It is, in practice, frequently ignored.

Now What?

We know robots are here and it is evident they can help us do more, but think about a few things:

1. If you work in or lead a business that uses manual labour, what would the arrival of humanoid robots actually mean for the people in your workforce, and how far in advance would they need to know?

2. If you or someone you know has spent years developing physical skills in manufacturing, logistics or care work, what does that expertise become when the physical tasks change?

3. The companies deploying these robots are largely making decisions about workforce composition that used to be made through negotiation with workers and unions. Does the current governance framework feel adequate to that task?

4. Jordan’s “for now” is carrying a lot. What would a transition that genuinely served Jordan, and the millions of workers in the same position, actually look like? Who is responsible for designing it?

The technology is arriving faster than most of the systems that should accompany it. That is not a reason to resist the machines. It is a reason to move faster on the things that matter more.

Leave a comment

You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.

Share

Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.