Meet the chip that learns without a teacher, without a cloud, without asking permission.
Two defence primes are already embedding it: Raytheon via an AFRL Phase II SBIR for radar processing on power-constrained platforms, and Lockheed Martin through its ForwardEdge ASIC subsidiary for fighter sensor fusion and counter-drone systems.
The chip draws under 300 mW in PCIe mode, less than a Bluetooth earbud, and supports on-chip learning: it adapts to new data without ever sending anything upstream.
The Chip That Broke the Cloud Dependency
The standard edge AI playbook runs on a compromise: infer locally, retrain in the cloud. A drone operating in a communications-denied environment cannot phone home for a model update. It has to learn on the fly or fail.
The AKD1500 is built on its Akida architecture, a fully digital, event-based neuromorphic engine fabricated on GlobalFoundries' 22 nm FD‑SOI process. It delivers 800 effective GOPS at under 300 mW in PCIe mode and under 200 mW in serial mode. The chip carries 1 MB of on-chip SRAM and runs models entirely self-contained. No external DRAM, no cloud round-trip, no data leakage.
AKD1500 Edge AI Co-Processor
800 GOPS peak performance · 22 nm FD‑SOI · On-chip learning · PCIe Gen2 / SPI · 7×7 mm package · $39.99 per unit · BrainChip, June 2026
A conventional GPU processes every pixel of every frame, burning watts on data that has not changed. The Akida architecture only fires when an event occurs: a pixel crossing a contrast threshold, a sensor detecting a change. In a static scene, the chip is asleep. In a radar or lidar stream, where relevant signals are sparse, the power savings compound with every empty frame.
Every autonomous platform today: drone, missile, UGV, submarine — operates under a hard power budget. The battery is not a convenience; it is the limiting factor on mission duration, sensor payload, and processing capacity. A chip that consumes milliwatts instead of watts, and that adapts its behaviour without network access, removes two constraints simultaneously.
Two Defence Primes, One Silicon Bet
The AKD1500 is not a research curiosity waiting for adoption. Two of the largest defence contractors in the world have already committed to embedding neuromorphic silicon into their next-generation platforms.
Raytheon, through an AFRL Phase II SBIR contract, is validating the Akida architecture for radar processing on power-constrained platforms: missiles, small drones, and drone-defense interceptors where the computing budget is measured in watts, not kilowatts. The contract builds on earlier independently validated demonstrations by RTX and the Fraunhofer Institute that showed measurable performance gains for radar discrimination algorithms running on neuromorphic hardware.
Lockheed Martin, through its ForwardEdge ASIC subsidiary, announced in March 2026 that it is embedding its Akida architecture directly into custom silicon for fighter aircraft sensor fusion, space-based surveillance payloads, and counter-drone systems. As we wrote in July, neuromorphic processors were approaching production readiness for edge robotics. That readiness has now arrived. The first customers are not consumer gadget makers. They are defence primes.
What a neuromorphic core in a custom ASIC means
The third signal comes from Neuromorphyx, a deep-tech hardware startup that selected the AKD1500 as the compute core for its Vision NeuroNode, a ruggedised always-on edge AI device for defence, robotics, and industrial sensing. Neuromorphyx's modular NeuroBlocks architecture pairs event-based vision sensors (DVS) with the Akida co-processor in a package designed for multi-year field operation on a single battery. The company's NeuroHive platform provides fleet orchestration and over-the-air model updates, a combination of high energy efficiency and remote manageability.
The Ecosystem Beyond BrainChip
It is the most visible player in the defence-neuromorphic space, but not the only one. Neurobus, a French startup founded by a former Tesla engineer, is designing neuromorphic computing platforms for autonomous defence and aerospace systems with customers including Airbus Defence & Space and Safran. VectorWave, which emerged from stealth in March 2026 with a $2.5 M seed round, has built a neuromorphic analog compute platform that performs AI inference directly on raw RF signals, enabling nanosecond-scale spectrum awareness for contested environments. Femtosense shipped over 100,000 units of its Sparse Processing Unit in 2025, though its focus remains on consumer audio and medical devices rather than defence.
The market is small. Its revenue is still measured in millions, not billions, but the direction is clear. The neuromorphic computing for defence market is projected to grow from $1.8 B in 2025 to $9.7 B by 2034, a compound annual growth rate of 20.5%, per Market Intelo.
What happens to the market a year from now?
Probability: 65% — The AKD1500 production milestone and the Lockheed/Raytheon commitments create a validation cascade. Northrop Grumman and BAE Systems are the most likely next movers, given their sensor-fusion and electronic warfare roadmaps. The constraint is not technology; it is the 18‑24 month qualification cycle for defence silicon.
✅ Arguments for
+ Two primes already integrating at silicon level (Lockheed ASIC, Raytheon SBIR)
+ Power advantage (sub-300 mW) is structurally unbeatable by GPUs for sparse sensor workloads
+ Patent filings in neuromorphic computing surged 401% in 2025 (PatSnap)
Confirmation criteria: A third prime (Northrop, BAE) announces a neuromorphic partnership by mid-2027.
❌ Arguments against
− Defence silicon qualification cycles run 18‑24 months. Adoption will lag commercial availability by at least two years
− Intel's Loihi 2 and IBM's NorthPole are competitive alternatives with deeper R&D budgets
− Neuromorphic toolchains (MetaTF, Lava) are immature compared to CUDA, creating a software moat for GPU incumbents
Disconfirmation criteria: No new prime commits by 2027, or it runs into production or supply issues.
Key signals to track
Its quarterly revenue crossing $10 M run rate — indicates production scale has moved beyond sampling
A third defence prime announcing a neuromorphic partnership — validates the Lockheed/Raytheon pattern
Intel or IBM securing a defence contract for Loihi 2 or NorthPole — introduces credible competition
The AKD1500 passing military-grade qualification (MIL‑STD‑883) — removes the last technical risk for prime adoption
Development scenarios
🟢 Optimistic scenario (30%)
Implications: Early investors in the neuromorphic defence supply chain capture a multi-year lead over incumbents still optimising GPUs for sparse workloads.
🟡 Base-case scenario (50%)
Implications: The market develops along dual tracks. Defence platforms with tight SWaP constraints adopt neuromorphic; platforms with available power and cooling stay on GPU. The total addressable market for neuromorphic defence silicon reaches $3–4 B by 2030.
🔴 Pessimistic scenario (20%)
Implications: The neuromorphic defence market consolidates around a larger incumbent. The specialised startup model fails to capture value in a procurement system optimised for prime-led integration.
