Twenty watts. That's roughly what your brain draws while it does everything you experience — seeing, remembering, reasoning, staying balanced on two legs. Less power than the bulb in a cheap desk lamp.
Now set that against where artificial intelligence is heading. Training a single frontier model can eat as much electricity as a small town. Running inference at global scale is on track to make data centers one of the largest new loads on power grids. And the industry's standard answer has been more of the same: bigger GPUs, more of them, larger buildings to house them.
That answer is running into walls. Physical ones. Economic ones. So a different camp of engineers stopped asking how to feed the machine and started asking why the machine is hungry. Their answer is to copy the organ that does intelligence best: chips that compute the way neurons do. It's called neuromorphic computing, and after decades of being perpetually five years away, it has started producing measured results. In 2025 and 2026, the field crossed a line it had never crossed before — from promising research to deployable systems.
The signals arrived in the same short window: Intel's largest neuromorphic deployment ever built, IBM's startling inference benchmarks, and a startup's silicon headed for space-grade applications. Three organizations, three approaches, one shared conclusion: the long "neuromorphic winter" appears to be ending.
The commute costs more than the math
A conventional processor is built on the von Neumann architecture, whether it's a CPU or a GPU. Memory lives in one place. Computation happens in another. Every operation begins with a journey — data shuttling back and forth — and here's the part the industry spent years not saying out loud: the shuttling consumes most of the energy in AI workloads. Moving data costs far more than the arithmetic itself. The math is cheap. The commute is ruinous.
Your brain simply doesn't commute. Its roughly 86 billion neurons keep memory and computation in the same place, communicating in sparse electrical spikes that fire only when something actually happens. Nothing happens, nothing fires, nothing is spent. The most sophisticated intelligence known, running on about 20 watts.
Neuromorphic chips copy that playbook. Instead of dense matrix multiplications running continuously, whether or not there's anything worth computing, spiking neural networks communicate through discrete events, and processing happens only when a spike arrives. That suits the naturally sparse workloads: always-on sensing, which doesn't need full power while it listens for nothing; real-time perception, which can't wait for a round trip to a distant data center; anomaly detection, long stretches of nothing punctuated by the one moment that matters. For them, the savings aren't marginal. They're measured in orders of magnitude, not percentage points.
A conventional chip earns its keep by being busy; idleness is waste. A spiking chip earns its keep the opposite way — doing nothing at nearly no cost, waking only for the event it exists to catch. For a sensor watching a quiet room, "nearly no cost for nearly all of the time" isn't a rounding error. It's the whole power budget, transformed. And the claim is narrower than it sounds: nobody serious says these chips beat GPUs at everything — only that for the right workloads, brain-inspired hardware uses a small fraction of the power conventional silicon needs.
Intel built a brain in a rack
Intel has been the field's most persistent industrial backer, through its Loihi line of research chips. The Loihi 2 processor implements asynchronous spiking neural networks, and Intel's large-scale system built from it — Hala Point — is now the world's largest neuromorphic deployment. It packs 1,152 Loihi 2 processors offering 1.15 billion neurons and 128 billion synapses, organized across 140,544 neuromorphic cores. All of it runs in a compact rack setup consuming up to 2,600 watts, at peak throughput equivalent to 20 peta-operations per second, with greater than 15 TOPS/W efficiency on 8-bit workloads.
Throughput on its own is a vanity metric; anyone can go faster by burning more. Hala Point's point is the envelope: computational muscle that would normally demand a far larger power budget, delivered inside 2,600 watts. Hala Point was built to prove that axis is real at scale. On the published numbers, it did.
Hala Point is a research testbed, not a product. Fair enough. But watch what researchers are doing with it. Scientists mapped the entire fruit fly whole-brain connectome — Drosophila melanogaster, 140,000 neurons and 50 million synapses — onto Loihi 2 across just 12 chips, and it ran far faster than traditional simulators. A whole brain. Mapped, running, on twelve chips. Whatever else this hardware is, it is no longer slideware.
The algorithmic work is moving too. New conversion methods using graded spikes turn conventional neural networks into spiking equivalents without excessive time steps, with reported results up to 3× throughput improvement and 2× energy savings relative to edge GPU baselines for small transformer-style models. Event-driven vision pipelines mapped to Loihi 2 have run at as little as 15 percent of the energy of equivalent CPU implementations. And reporting around Intel's roadmap describes the Loihi 3 generation as marking the end of the "neuromorphic winter," with the hardware beginning to compete with, and in specialized edge applications surpass, conventional architectures.
IBM skipped the spikes
IBM looked at the same enemy and chose a different weapon. Where Loihi implements spiking networks, IBM's NorthPole attacks the von Neumann bottleneck directly: it co-locates memory and compute on a single chip, eliminating the energy-hungry data movement that dominates conventional AI inference. No spikes required. Just no commute. The chip is fabricated at 12 nanometers, carries 22 billion transistors, and occupies about 795 square millimeters.
The measured results are striking. In experiments with a 3-billion-parameter large language model, a NorthPole system attained latency under 1 millisecond per token and throughput of 28,356 tokens per second, while delivering approximately 72.7× better energy efficiency than the next most efficient GPU in the comparison. On the ResNet-50 image-recognition benchmark, it demonstrated 25× better energy efficiency than comparable-process GPU/CPU systems. Read the 72.7× figure the way engineers read it — not as proof GPUs are obsolete, but as evidence the bottleneck was never the arithmetic. It was the journey the data took to reach it. Reporting on IBM's roadmap describes NorthPole moving into production for 2026.
Follow the implication out to the edge. If inference can run at a fraction of today's energy cost, workloads that need cloud round-trips — real-time translation, on-device assistants, autonomous perception — become feasible on battery-powered devices. The phone stops phoning home. The sensor stops waiting. The same pressure is reshaping everything around it: inference spreading onto decentralized networks that challenge the big clouds, and multi-agent systems cutting enterprise latency in production settings.
Intel and IBM aren't the whole story. BrainChip Holdings launched its Akida 2.0 architecture in late 2025, built around what it calls Temporal Event Neural Networks, aimed at always-on sensing: anomalies in industrial machinery vibrations, EEG signals in healthcare, keyword spotting. Most tellingly, BrainChip licensed the technology to NASA for space-grade AI, where power is the most limited resource of all. There is no wall socket to fall back on. Mercedes-Benz is integrating neuromorphic vision for collision avoidance, where event-based cameras feeding neuromorphic processors can achieve pedestrian-detection latency around 0.1 milliseconds, versus 30–50 milliseconds for conventional frame-based systems. A pedestrian detected in a tenth of a millisecond is a different safety proposition from one detected in fifty — measured in meters of braking distance.
The frontier keeps widening. The Chinese Academy of Sciences' "Speck" chip, with its SpikingBrain-1.0 model, idles at 0.42 milliwatts. KAIST's "Frequency Switching Neuristor" showed a 27.7% energy reduction in simulations. Analog in-memory computing has demonstrated up to 12.4 TOPS/W. Different labs, different continents, different techniques — all attacking the same enemy, the energy cost of moving data.
Read the benchmarks like an engineer
Now the fine print — this field has earned some skepticism, and the honest version beats the hype version. The efficiency figures here are real, measured and published. They are also workload-specific, and that qualifier is not a footnote — it is the story. Neuromorphic chips excel at sparse, event-driven, inference-heavy workloads: always-on sensing, real-time perception, anomaly detection, and increasingly small transformer models converted to spiking form. That is a genuine and enormous territory. It is not the whole map.
What these chips do not do is train frontier models. Training is dense, continuous, large-scale mathematics — exactly what GPUs were built for, and exactly what spiking architectures are not. Nor do they run dense large-scale matrix math as efficiently as GPUs. And then there is the quiet limiter: software. The compilers, frameworks and developer tools around neuromorphic hardware remain years behind CUDA's maturity. But an ecosystem has to be lived in by thousands of developers before it compounds, and that takes time no benchmark can shortcut.
So the right mental model is complement, not replacement. Data centers will keep their GPUs for training and dense inference. But the edge — billions of devices where power budgets are measured in milliwatts and latency in milliseconds — is where neuromorphic architectures hold a structural advantage conventional silicon cannot match by scaling alone. It comes from where the memory sits. Which means it doesn't evaporate the next time a GPU generation ships.
As for the "10×" framing: it understates the measured results on favorable workloads and overstates them as a general claim, both at once. The truth is in the benchmarks — all of them. For product builders, the point is narrow and clear: if your AI workload is always-on, event-driven and power-constrained, the neuromorphic ecosystem of 2026 deserves a serious evaluation. Not a press release. An evaluation. The winter is over; what comes next is the harder, more interesting work of finding out exactly where this hardware earns its place.
Efficiency claims get tested against the silicon actually shipping. On the device side that now includes Huawei's Mate 90 with the Kirin 9050 Pro and its ten-year AI life pitch. At the opposite pole of the power budget sit NVIDIA's Blackwell B200 installations and their liquid-cooling demands — the two ends of the AI hardware market are diverging, and neuromorphic designs are aimed squarely at the gap.
Sources
- OpenPR — "Global High-Speed Neuromorphic Computing Chips Market" (Hala Point specs, NorthPole LLM results, Loihi 2 conversions): openpr.com
- FinancialContent/TokenRing — "The Brain-Like Revolution: Intel's Loihi 3 and the Dawn of Real-Time Neuromorphic Edge AI" (NorthPole production 2026, BrainChip Akida 2.0/NASA, Mercedes-Benz): markets.financialcontent.com
- GitHub / knowledgeekza3224 — "SC Totality Energy Convergence Map 2026-09-10" (NorthPole 25× ResNet-50, analog in-memory 12.4 TOPS/W, IBM research links): github.com
- FinancialContent — "Brain-Inspired Breakthroughs: Neuromorphic Computing Poised to Reshape AI's Future" (TrueNorth, Loihi milliwatt operation, Akida Pulsar, Speck, KAIST): markets.financialcontent.com
