UK chip startup Fractile raises $220M Series B for inference silicon; Anthropic in talks to buy when chips ship 2027

Fractile, a London-based chip startup founded in 2022, has raised $220 million (£165m) in Series B at a $1 billion+ post-money valuation, led by Accel, Factorial Funds, and Founders Fund (Peter Thiel’s firm). The company is building inference chips using an SRAM-based in-memory compute architecture that avoids traditional off-chip DRAM shuttling, targeting token-generation latency as the central bottleneck for frontier AI workloads. Existing hardware runs advanced models at ~40 tokens/second; a 100-million-token output takes roughly a month. Fractile aims for 1,200+ tokens/second, unlocking new applications in software engineering automation, drug discovery, and materials simulation.

According to founder Walter Goodwin (Oxford Robotics PhD, age ~28), the architecture co-locates memory and compute on the same die using SRAM, reducing memory-bandwidth bottlenecks that are the primary constraint on inference speed. Fractile claims simulations show 100x faster, 10x cheaper inference vs. NVIDIA GPUs for inference workloads—though production test chips have not shipped. The team includes engineers from Graphcore, NVIDIA, and Imagination Technologies, with investments from Kindred Capital and the NATO Innovation Fund. Earlier in 2026, Fractile announced a £100M three-year UK expansion including new Bristol hardware-engineering facility.

Anthropic has reportedly held early-stage discussions to purchase Fractile chips upon commercial availability (expected ~2027), according to The Information. This would make Fractile Anthropic’s fourth named silicon supplier alongside NVIDIA, Google TPUs, and Amazon Trainium/Inferentia. Anthropic has deliberately avoided single-vendor lock-in, even as it explores in-house chip design. Google separately is assembling a four-partner inference stack (Broadcom, MediaTek, Marvell) to challenge NVIDIA at the inference layer, signaling broader vendor appetite for inference specialization.

Fractile exemplifies a growing thesis that inference—the runtime layer where deployed models spend most of capex dollars—is structurally distinct from training and winnable by specialists. Competing inference-chip startups (Groq, Cerebras) are following similar near-memory-compute strategies. For architects evaluating supplier diversification and latency budgets: Fractile’s 2027 silicon is too far out for near-term procurement, but the inference-vendor ecosystem is fracturing, creating procurement optionality beyond NVIDIA for the inference-dominant tail of most workloads.