Anthropic Commits $250M to Fractile's Unbuilt Chips — and the UK Startup's Valuation Jumps 6.5x in Three Months

Anthropic has agreed to buy roughly $250 million of inference chips from Fractile, a London startup whose silicon won’t be production-ready until 2027. The deal has instantly repriced the company: Fractile is now raising around $600 million at a $6.5 billion pre-money valuation — up from roughly $1 billion in May, when it raised $220 million from Accel, Founders Fund, and Factorial Funds. A 6.5x markup in three months, driven by a single customer commitment for chips that don’t exist yet.

Founded in 2022 by Oxford roboticist Walter Goodwin, Fractile is attacking the most expensive line item in frontier AI: inference. Its architecture is a genuine departure from how every mainstream accelerator works. Instead of shuttling data between compute cores and stacks of High Bandwidth Memory — the HBM that’s currently rationed, marked up, and fought over across the entire industry — Fractile integrates compute directly into SRAM on the same die. No HBM stacks. No off-chip DRAM round-trips. The data largely stays where the math happens.

That design choice is perfectly timed. Memory pricing is in a genuine crunch, with HBM allocations sold out and DRAM costs spiking. A chip that sidesteps the memory hierarchy entirely isn’t just an architectural curiosity — it’s a hedge against the exact supply chain everyone else is trapped in. Both sides have signaled intent to expand the contract beyond the initial $250 million.

For Anthropic, the logic mirrors what every frontier lab is doing: diversify away from Nvidia before inference costs eat the margin. Anthropic already runs on Google TPUs and Amazon’s Trainium alongside Nvidia GPUs. A $250 million pre-purchase buys Fractile the capital to finish its tape-out and buys Anthropic a guaranteed place in the delivery queue — with an option on much cheaper inference if the architecture delivers.

The risk is equally clear. Fractile has never shipped a chip. Cerebras and Etched, its closest analogs in the “novel architecture” category, both took years longer than promised to reach volume production. Betting $250 million on unshipped silicon is what an inference cost problem looks like when it gets severe enough — Anthropic’s Q2 revenue hit $11.5 billion, and serving that demand on GPU economics is precisely the pain this deal is designed to escape.

Neither company commented on the announcement. The 2027 delivery date will do the talking.

Sources

• Yahoo Finance