Twenty Watts

The world's largest neuromorphic computer, 1.15 billion artificial neurons, consuming 2,600 watts of power. This is technology learns the way the brain learns, Most people have never heard of it.

May 11, 2026

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.

Neuromorphic computing — state of the technology in 2026

Intel Loihi 2 / Hala Point: 1M neurons/chip; 1.15B neurons in Hala Point (1,152 chips); 20 petaops; 2,600W max
Intel Loihi 3 (Jan 2026): 8M neurons/chip; 64B synapses/chip; 4nm; 32-bit graded spikes; first commercial neuromorphic chip; general availability Q3 2026
IBM TrueNorth: ~46B synaptic ops/sec/watt; optimised for inference; fixed architecture
IBM NorthPole: Production-ready; vision and enterprise inference; neuromorphic-influenced design
SpiNNaker 2 (Manchester / SpiNNcloud): ARM-core mesh; 175M neurons at Sandia (early 2025)
BrainScaleS-2 (Heidelberg): Analogue circuits; 10,000× faster than biological real time; neuroscience focus
Milestone: Sandia PDE result - neuromorphic hardware solving physics simulation (Nature Machine Intelligence, Jan 2026)
Milestone: First neuromorphic LLM - Loihi 2 LLM at SCOPE workshop, ICLR April 2025; half the energy of GPU equivalent
Edge applications: Event cameras + neuromorphic chips for robotics; sub-millisecond latency; battery-efficient
Market: ~$920M in 2024; estimated $8.76B by 2033 at 30.4% CAGR (DataM Intelligence; note: estimates vary widely across analysts)
AI energy context: Data centres consumed ~415 TWh globally in 2024 (IEA, Energy and AI, Jan 2025); AI workloads projected to grow at ~30%/year; total consumption projected to nearly double by 2030

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.

The Next Evolution

Discussion about this post

Ready for more?