In 2017, Intel Labs engineer Mike Davies stood on stage in Silicon Valley and held up a chip the size of a fingernail. It consumed less power than a hearing aid. It could learn from data in real time. And it processed information the way a brain does: through spikes, not clock cycles. Nobody in the audience, not even the semiconductor veterans in the front row, knew what to make of it. Loihi, the company's first neuromorphic processor, was a research project without a clear commercial path. Nine years later, that path has arrived.
Intel's Loihi 3 and BrainChip's Akida 2.0 deliver brain-inspired computing that consumes 1/1000th the power of GPUs, enabling autonomous robots to run continuous AI on battery power for the first time.
The developer shipped the first production-ready neuromorphic chip — the AKD1500 — in June 2026, fabricated on GlobalFoundries 22nm FD-SOI, targeting industrial and defense applications.
The Lockheed Martin and Raytheon collaborations signal that neuromorphic is moving from academic research into deployed systems, with defense primes embedding Akida directly into custom silicon.
A decade ago, running a neural network on a drone meant the drone carried a GPU, a cooling fan, and a battery so large the flight time collapsed. Neuromorphic computing reverses the equation. Instead of pulling data through separate memory and processing units (the von Neumann bottleneck that has dominated computing since 1945), these chips integrate both in the same circuit, mimicking biological neurons. Only the data that changes gets processed. Everything else stays dark, drawing no power at all.
How brain-inspired silicon actually works
Conventional AI accelerators — GPUs, TPUs, NPUs — process data in dense, synchronous batches. Every pixel in every frame gets multiplied, whether it changed or not. A security camera running object detection on a GPU might consume 15 watts while watching an empty street. A neuromorphic chip running the same model draws milliwatts, because nothing happens until a person walks into the frame.
The difference is architectural. In a von Neumann machine, memory and compute are separate; data shuttles between them through a bus that creates both a bottleneck and a power drain. In a neuromorphic chip, memory and compute are co-located in each artificial neuron. Information moves only when a neuron fires, an event-driven sparse computation that mirrors biology.
Neuromorphic efficiency vs conventional AI hardware
The chipmaker's Loihi 3 and the manufacturer's Akida achieve 1000× lower power consumption for spiking neural network inference compared to GPU-based edge AI. Event-driven processing means chips draw power only when data changes. · the company Labs / the firm, 2026
Intel's third-generation design, Loihi 3, announced in early 2026, pushes this architecture further. It integrates a 1.15-billion-neuron system called Hala Point at Sandia National Laboratories — the largest neuromorphic system ever built, achieving 15 trillion operations per second per watt. The more immediate story for robotics engineers is at the other end of the scale: sub-watt chips that fit inside a drone arm or a security camera housing.
The company's commercial breakthrough
The chipmaker dominates the research headlines. BrainChip, an Australian-headquartered firm listed on the ASX, dominates the commercial reality. On June 30, 2026, the company announced production shipments of its AKD1500 chip, the first commercially available neuromorphic processor built for industrial and military environmental tiers.
The AKD1500 is fabricated on GlobalFoundries' 22nm FD-SOI process and draws under 300 milliwatts in PCIe mode and under 200 milliwatts in serial mode. It supports PCIe for complex edge architectures alongside x86, ARM, or RISC-V application processors, and serial interfaces for battery-constrained embedded systems. "We are providing industrial and defense partners with the rugged hardware flexibility they need to deploy intelligence Anywhere, completely free from cloud dependence," said Sean Hehir, CEO of the manufacturer.
The entire edge AI market has been built on a compromise: run inference locally, but retrain in the cloud. the firm's Akida architecture supports on-chip learning: the chip adapts to new data without sending anything upstream. For a defense application, that is not a feature. It is a requirement. A drone operating in a communications-denied environment cannot phone home for a model update. It has to learn on the fly.
Defense primes are already embedding neuromorphic silicon
Two defense primes have publicly committed to neuromorphic integration in 2026. Raytheon, through an AFRL Phase II SBIR contract, is validating the company's Akida for radar processing on power-constrained platforms: missiles, small drones, and drone-defense interceptors where the computing budget is measured in watts. Lockheed Martin, through its ForwardEdge ASIC subsidiary, is embedding Akida directly into custom silicon for fighter aircraft sensor fusion, space-based surveillance payloads, and counter-drone systems.
The Lockheed partnership, announced in March 2026, goes further. ForwardEdge builds application-specific integrated circuits for the defense microelectronics supply chain. Embedding a neuromorphic core into a custom ASIC means the brain-inspired architecture becomes a permanent part of the hardware, not an evaluation board that might be removed in the next revision.
"BrainChip's Akida architecture is a strong complement to our ASIC and RF platforms," said Bill Jenkins, ForwardEdge's Chief Revenue Officer.
Neuromorphic vs GPU: when each wins
GPU still wins: large transformer models, dense image classification on stationary hardware, training workloads at data-center scale.
The two architectures are complementary, not competitive. A humanoid robot might run its vision pipeline on a Jetson Thor GPU and its reflex loops on an Akida AKD1500, the same way a human brain delegates survival reflexes to the spinal cord while the cortex handles deliberation.
Where neuromorphic breaks down
Three limitations keep neuromorphic from replacing GPUs anywhere soon. First, software tooling remains fragmented. Intel's Lava framework and BrainChip's MetaTF are incompatible with each other and with the PyTorch/TensorFlow ecosystem that the entire AI industry is built on. A team that adopts neuromorphic today is committing to a parallel software stack with a smaller community and fewer pre-trained models.
Second, the performance advantage is workload-specific. Neuromorphic chips excel at sparse, event-driven tasks — object detection in a mostly empty scene, keyword spotting, anomaly detection in vibration data. They underperform on dense transformer inference, which is what the most capable AI models currently require. the firm's Akida 2.0 added support for vision transformers, a step toward closing the gap, but the architecture is still not competitive with an H100 for training.
Third, the supply chain is narrow. The developer's Loihi remains a research chip with no announced commercial production date. the company's AKD1500 is shipping in volume, but it depends on a single foundry partner (GlobalFoundries) and a single process node (22nm FD-SOI). A disruption at that node halts the entire commercial neuromorphic channel.
The chipmaker announces commercial Loihi 3 production date or foundry partnership
Defense SBIR Phase III awards for neuromorphic-equipped platforms
the developer or the company publish a standardized benchmark for neuromorphic vs GPU at edge
A major robotics OEM (Boston Dynamics, ANYbotics, Agility) announces neuromorphic integration
The market is already accelerating
The neuromorphic computing market is projected to reach $20.3 billion by 2030, according to industry estimates. the manufacturer alone raised $25 million in January 2026 to scale its Akida product line, and the AKD1500's production milestone in June validated that commercial demand is no longer theoretical. The developer's Hala Point — a research system deployed at Sandia — suggests the architecture scales to datacenter levels, but the real revenue opportunity is at the edge: security cameras, industrial sensors, drones, and wearable devices where running AI on milliwatts unlocks applications that GPUs simply cannot serve.
The defense angle accelerates this timeline. Traditional military hardware procurement cycles run five to seven years, but both the Raytheon SBIR and the Lockheed ForwardEdge partnership are structured to push neuromorphic into fielded systems by 2028 at the latest. A drone that can identify a target, learn new acoustic signatures on station, and operate for 12 hours on a single battery charge does not compete with GPU-equipped alternatives. It competes with nothing, because GPU-equipped alternatives do not last 12 hours in the air.
The AKD1500 is shipping. Loihi 3 exists in working silicon. Whether the software ecosystem can mature fast enough to make them programmable by ordinary robotics teams, not just the specialized researchers who currently write Lava and MetaTF code — that is the open question. That gap between working hardware and accessible software is the last bottleneck between brain-inspired computing and the robots that will use it.
IBM NorthPole and the competition
The chipmaker and the firm are the most visible players, but they are not the only ones. IBM's NorthPole chip, now in production in 2026, takes a different approach: 256 cores with 256KB SRAM each, eliminating external DRAM entirely. It achieves 25 times better energy efficiency than an NVIDIA H100 for ResNet-50 inference, though it is optimized specifically for computer vision at the edge rather than spiking neural networks.
NorthPole takes a digital, non-spiking path to brain-inspired efficiency: it keeps the von Neumann architecture but compresses memory and compute onto a single die to minimize data movement. Akida and Loihi, by contrast, are spiking architectures that exploit temporal sparsity. Both approaches achieve dramatic power savings, but they are not interchangeable. A robotics team evaluating neuromorphic hardware today must choose which flavor of brain inspiration fits its workload: dense vision inference (NorthPole) or event-driven sensor fusion (Akida, Loihi).
The emerging ecosystem also includes the well-funded Femtosense, a Silicon Valley startup that shipped over 100,000 units of its Sparse Processing Unit in 2025, and a new entrant called Aspirare Semi from Ottawa. Neither has the defense relationships that the company has built through its Lockheed and Raytheon partnerships, but they widen the commercial base and create alternative supply chains. The market is still small enough that no single architecture has locked in the standard — and that window of competition is where the most important design decisions for next-generation robotics will be made.
What changes when robots learn on the edge
The most underappreciated consequence of neuromorphic computing is autonomy, not power efficiency. A robot running a GPU needs cloud connectivity for retraining: the model on the device is frozen at deployment time, and updating it requires shipping data to a datacenter, retraining, and deploying a new binary. That cycle introduces latency, security risk, and operational complexity. A robot running Akida or Loihi learns continuously, adapting to new environments without phoning home.
For a warehouse robot, that means recognizing a new package shape on the first encounter and updating its grip model before the next pick. For a defense drone, it means identifying a never-before-seen radar signature and adjusting its classification within seconds. For a surgical robot, it means adapting to tissue variability across patients without uploading intraoperative data to a cloud server. The architecture does not merely save power. It rewrites the relationship between the device and the network that controls it. In an era where connectivity cannot be guaranteed and the geopolitical trend is toward communications denial, that rewiring may be the most strategically significant output of the entire neuromorphic industry.
The bottleneck is software
Hardware is shipping. The software stack is not keeping pace. Intel's Lava framework, BrainChip's MetaTF, and IBM's toolkit for NorthPole are each proprietary and mutually incompatible. A robotics team that standardizes on Akida today cannot switch to Loihi without rewriting its inference pipeline from scratch. There is no PyTorch for neuromorphic — no common intermediate representation that abstracts away the underlying architecture. The field is roughly where deep learning was in 2013, when Caffe and Theano competed for mindshare before TensorFlow and PyTorch consolidated the ecosystem.
That analogy cuts both ways. In 2013, the hardware was not ready either — GPUs existed, but the software that made them programmable for neural networks was still being written. The next five years will determine whether neuromorphic follows the same trajectory. The chips work. What remains to be built is the bridge between working silicon and the thousands of robotics engineers who have never heard of a spiking neural network but need to deploy one next quarter.
The first chips that think like brains are no longer a laboratory curiosity. They are shipping in production volumes, certified for military environmental tiers, and embedded in custom ASICs for fighter aircraft. The market is small — BrainChip's revenue is still measured in millions, not billions — but the trajectory is clear. Neuromorphic computing will not replace GPUs for training foundation models. It does not need to. The opportunity is in the vast middle of the computing spectrum: devices that must see, hear, and decide without exhausting their battery or phoning a datacenter. That is most of robotics. And for the first time, there is silicon designed specifically for it.
The race to deploy neuromorphic silicon at the edge is no longer about whether the hardware works. It works. The race is about engineering teams: the tools, the abstractions, and the trained engineers who know how to program spiking neural networks. The first company that solves that software bottleneck will own the robotic reflex loops of the next decade. Intel has research depth. BrainChip has production silicon and defense contracts. The difference between winning and losing this market will come down to which one makes its chips programmable by a Python engineer who has never seen a spike.