Fractile, a British chip startup that has yet to ship a product, is in talks to raise roughly $600 million at a $6.5 billion pre-money valuation on the strength of an initial agreement to supply Anthropic with about $250 million of inference silicon, Bloomberg reported Wednesday. If the round closes at that price, it would value the four-year-old company at more than six times what investors paid for it in May.
The compression is the story. Fractile announced a $220 million Series B on May 13 at a valuation of roughly $1 billion, co-led by Factorial Funds, Accel and Founders Fund, with Gigascale, Conviction, Buckley Ventures, 8VC and Felicis joining. Three months and one anchor customer later, it is negotiating at $6.5 billion pre-money. Its first chips are not expected to reach customers until 2027, which means essentially the entire markup rests on a contract for hardware that does not yet exist and a thesis about what inference will cost when it does.
A note on sourcing. The $6.5 billion figure, the $600 million target and the $250 million Anthropic contract all trace to a single Bloomberg report published Wednesday and aggregated by Techmeme the same day. The Next Web, Yahoo Finance and Investing.com carried the numbers within hours, but each attributes them to Bloomberg, and neither Fractile nor Anthropic has commented publicly. Tom's Hardware reported in early May that Anthropic was in early talks to buy Fractile accelerators, which corroborates the relationship but not the size of the deal or the round. Treat the valuation as reported, not confirmed.
What is on the record is the trajectory. Fractile was founded in 2022 by Walter Goodwin, then a PhD student at the University of Oxford's robotics institute, and emerged from stealth in July 2024 with a $15 million seed round co-led by Kindred Capital, the NATO Innovation Fund and Oxford Science Enterprises. Pat Gelsinger, the former Intel chief executive, invested as an angel in January 2025, a signal that carried outsized weight for a company then measured in dozens of employees. Fractile now runs out of London and Bristol and has been hiring in the US and Taiwan.
The technical bet is in-memory compute: fusing storage and arithmetic onto the same die so model weights never have to be shuttled across an external bus. Fractile's architecture leans on SRAM rather than the HBM stacks that dominate GPU accelerators, and the company claims a hundred-fold increase in effective bandwidth alongside sharply lower energy per token. Its public materials claim LLM inference up to 100 times faster than current hardware at roughly a tenth the operating cost, all unvalidated on silicon that has not run a customer workload.
"The technical and economic limits on inference speed, above all from memory bandwidth that has failed to scale on current architectures, are what is constraining progress," Goodwin wrote when the Series B was announced in May. Most of the industry now shares the diagnosis. The disagreement is over whether a startup can fix it before Nvidia, Google and Amazon do.
Why It Matters
For Anthropic, $250 million is small change against a compute budget now measured in gigawatts. What makes it interesting is the pattern it fits. Anthropic's stack is already quad-vendor: Nvidia GPUs, Google TPUs under a Broadcom-built expansion bringing more than a gigawatt online during 2026, AWS Trainium 2 behind Project Rainier, and up to two gigawatts of AMD Instinct MI450 accelerators under a July deal that also had AMD committing as much as $5 billion into the lab. Adding a pre-revenue British startup to that roster is a different category of bet, an option on an architecture rather than a purchase of capacity.
The memory angle sharpens it. All three major DRAM manufacturers have reportedly sold out their 2027 DRAM and HBM output, booked months ahead of the normal planning cycle by hyperscalers on long-term agreements. In that market, a design that sidesteps HBM entirely is a hedge against the one input no amount of capital reliably buys. If Fractile's SRAM approach works, Anthropic gets serving capacity that is not competing for the same constrained supply as everyone else's. If it does not, the company is out a fraction of what it spends on GPUs in a quarter.
The price also looks less absurd in context than in isolation. Nvidia paid roughly $20 billion in December 2025 to license Groq's inference technology and absorb its core team. Cerebras went public on May 14 and closed its first day near a $56 billion fully diluted valuation. Etched raised a $300 million Series C led by Sequoia in July at $10.3 billion, backed by more than $1 billion in signed contracts. Against those marks, $6.5 billion pre-money for a company with a named frontier-lab customer starts to read as a market rate rather than an outlier.
What To Watch
First, whether the round closes at all, at what price, and who leads it. A step-up this size usually needs a crossover or sovereign fund willing to underwrite a 2027 product. Second, whether Anthropic confirms the agreement and whether the reported room to expand hardens into a multi-year commitment with volume attached, which is the difference between a design partnership and a supply deal. Third, tape-out and first-silicon timing, plus any independent benchmarks, because in-memory compute has a long history of promising order-of-magnitude gains that shrink under real workloads. Fourth, whether the DRAM squeeze eases, since much of Fractile's strategic value is priced off scarcity that may not last to 2027. And fifth, the British angle: a UK-headquartered accelerator company worth $6.5 billion with a frontier-lab anchor becomes a national industrial asset, and Whitehall has not been shy about sovereign-compute framing.
“The technical and economic limits on inference speed, above all from memory bandwidth that has failed to scale on current architectures, are what is constraining progress.”— Walter Goodwin, Founder and CEO, Fractile