Larry Steel

Good to see more people building Akida into LLM infrastructure. Selective execution, neuromorphic routing in front of LLMs, and multi-chip Akida clusters are where this field is heading, among other places. They’re also what I’ve been running and publishing all year, across far more than one box. A few dated markers from the public record: - May 8, 2026 (Demo 34). Akida-routed LLM fan-out. A phone message went into a Hermes orchestrator, which fanned out to 8 parallel sub-agents across 8 Symphony nodes. Each sub-agent was routed in about 5 ms by AKD1000 chips to the right model tier, either Granite 4.1 on IBM Cloud or Nemotron on-prem. kevindjohnson.org/#demo34 - June 14 (Demo 52). Three synthetic cortexes on Akida in Washington DC, Dallas, and Pittsburgh. They learned online, with no GPU, and reached consensus across the continent through a mixture-of-experts council. kevindjohnson.org/#demo52 - June 26 (Demo 55). An AKD1000 in the M.2 slot of an NVIDIA Jetson Orin Nano, running real inference. With LSF or Symphony on board, the same card coordinates across many Jetsons. kevindjohnson.org/#demo55 - July 30 (Demo 70). One unmodified vLLM endpoint served by Akida across 56 nodes at three sites, with placement done live. kevindjohnson.org/#demo70 - August (IEEE NAECON 2026). Three peer-reviewed papers with Steven Harbour and colleagues: -- A spiking-network workload manager on an Akida cluster that dispatches in 5 ms and cuts GPU cost 30–50%. -- A 46-node fleet holding 8,832 concurrent inference contexts at 99.4% of ideal scaling. ieeexplore.ieee.org/document… ieeexplore.ieee.org/document… -- Decision-Optimized Compute Theory, which formalizes selective execution: suppress, answer on neuromorphic, or escalate to the GPU. ieeexplore.ieee.org/document… - September 29 (Demo 82). Akida orchestrated across 131 nodes on six continents. kevindjohnson.org/#demo82 The point isn’t a benchmark on any one platform. Jetson is one target among many: Intel N100 fleets, Qualcomm RUBIK Pi, FPGAs, IBM Cloud, and federated multi-site clusters across multiple clouds and on-prem efforts. Akida's relevance is anywhere and everywhere. The substrate is IBM Spectrum Symphony for placement and GPFS for shared state, held constant across 82 published demonstrations in domains as far apart as electronic warfare, credit risk, sleep staging, and Renaissance counterpoint. The heterogeneous compute ontology of the neuromorphic hive mind. The proven claim is architectural. Intelligence doesn’t have to live in one ever-larger model and it's not limited to 1950/1980s brute force LLM work. True brain-inspired intelligence can live in many small neuromorphic units that perceive locally, learn continuously, and coordinate. A coordination idea that only works on one box or platform is interesting. One that holds across 12 silicon architectures and six continents is an architecture. Welcome to the river. The bridges are already here. Run Akida anywhere. Run Akida everywhere. Full record, papers, and every demo: kevindjohnson.org

BOOM. Now people can start to see what's possible.

Normally, I'd save this for the weekend. But, I just can't stand by and watch people claim things like neuromorphic development is broken. So, this is a bit long and somewhat complicated, but still worth your time if you care about the field.

The better question for 2027 isn't how big one chip can get or more research that never commercializes. It's how well many chips can be orchestrated. Neuromorphic computing is a new paradigm you can build now, not years in the future and not merely something to watch.