Lamb Labs

Lamb Labs (YC S26) is a London-based chip startup building Model Processing Units (MPUs): inference chips that hardcode a specific language model, weights included, directly into silicon, so the model never moves between memory and compute. The founders argue GPUs spend most of their inference energy fetching weights rather than computing, and target 20,000+ tokens per second at 63x higher intelligence per watt than today's best AI accelerators — company targets, not measured results. Its roadmap has four stages: Woolly, a post-trained Qwen model shrunk to fit on-chip (live demo against Qwen3-8B); an FPGA accelerator running an 8B-parameter model under 10 W on a $250 AMD Kria KV260 board; Little Lamb, a tiny self-trained model generating text entirely inside FPGA fabric with no DRAM reads per token; and, as the endgame, a custom ASIC per model designed with a reinforcement-learning environment that searches chip architectures for speed and energy per token. It pitches fully local, private inference to power-bound neocloud and data-center operators, air-gapped and sovereign deployments, latency-sensitive uses such as high-frequency trading, and thermally constrained robotics and wearables. Founded in 2026 by CEO Niki Kotecha (AI PhD at Imperial College London, engineering at Cambridge) and CTO Thomas Lanning (mathematical and theoretical physics at Oxford). Featured by Forbes among YC startups in September 2026. No funding amount or valuation has been disclosed beyond YC backing.