Fractile seeks a $6.5bn valuation on an Anthropic chip deal and no shipping product
The British startup's valuation would rise more than sixfold since May on a $250m supply agreement for inference chips that are not expected to ship until 2027.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Fractile, a British AI chip startup, is in talks to raise about $600 million at a $6.5 billion valuation, Bloomberg reported on August 19.
- The valuation would be more than six times the roughly $1 billion it raised at in May, when it took $220 million from Accel, Founders Fund and Factorial Funds.
- The jump follows an initial agreement to sell Anthropic roughly $250 million of inference chips, with the intention of expanding the contract later.
- Fractile's chips are not expected to ship until 2027; its claims of running large models far faster at lower cost are company figures with no deployed hardware behind them yet.
— The Tom's Hardware citation pointed at a truncated URL that returned 404; the link now carries the full slug ending in "-fractile" and resolves to the article as cited. The desk's error in transcription, caught by the weekly link check. No claim in the story changes.
Fractile, founded in 2022 by Oxford PhD Walter Goodwin, is in talks to raise roughly $600 million at a $6.5 billion valuation, Bloomberg reported on August 19. That would value the British chip startup at more than six times the roughly $1 billion it commanded in May, when it raised $220 million from Accel, Founders Fund and Factorial Funds.
What changed in between is a customer. Fractile reached an initial agreement to supply Anthropic with about $250 million of its inference chips, with both sides intending to expand the contract. The company's design places compute and memory on the same die using SRAM, rather than fetching weights from separate off-chip memory — an approach it calls memory-compute fusion, and one that sidesteps the DRAM the rest of the industry is currently competing for amid a severe memory squeeze.
The caveat is the timeline. Fractile's chips are not expected to ship until 2027, and its headline claims — running large language models up to 100 times faster than existing hardware at 90% lower operating cost — are company figures for hardware that does not yet exist in deployable form. Neither the performance nor the economics has been independently verified in production.
A sixfold valuation step on one supply agreement and no shipping silicon is a bet on the memory bottleneck rather than on Fractile's execution, and it tells you how much buyers will now pay to design around DRAM scarcity. For anyone planning inference capacity for 2027, the useful signal is that a major lab is willing to underwrite an unproven architecture to get off the standard memory supply chain — but a contract for chips that do not exist is a purchase option, not a roadmap you can build against.