Fractile Raises $720M Series C to Break the AI Inference Memory Bottleneck

As AI agents work longer and reasoning models think longer, the bottleneck shifts from how smart a model is to how fast its answers come out. A UK startup that wants to solve this by redesigning the chip from scratch, not the software, has raised a $720 million Series C.

Fractile announced the round on Nov. 11. Thrive Capital, Founders Fund, Lightspeed and Redpoint co-led. According to Reuters, the company was valued at $7 billion, making it one of Britain’s most valuable tech startups. At the Series B in May its valuation was reportedly about $1 billion, so it has grown roughly sevenfold in six months.

CapitalG, DST Global, Elad Gil, BlackRock, Bain Capital Ventures, T. Rowe Price, SV Angel, Shopify Ventures, Felicis, Gigascale Capital, 20VC, Georgian and Balderton also participated, along with existing backers. From the UK government side, the Sovereign AI fund and the British Business Bank invested. Reuters reported that UK AI minister Kanishka Narayan said Britain has produced some of the world’s best scientific ideas without turning them into industry, and that the government is determined to change that. The round is read as an effort to build domestic UK AI infrastructure rather than depend on Nvidia and a handful of Asian chipmakers.

Fractile targets the memory that limits inference speed. Running a giant model quickly takes both enough capacity to hold the whole model and enough bandwidth to feed data to the compute chip nonstop, and until now the two have been a trade-off. On-chip SRAM is fast but small, while the HBM used with GPUs and TPUs has capacity but less bandwidth. Fractile says it co-designs DRAM stacked in 3D with logic to deliver far higher bandwidth and speed than HBM while keeping the capacity needed to run the largest models at long context. It also stresses being a full-stack company, with one team spanning machine learning research, circuit design and rack-scale systems. It has offices in London, Bristol, Palo Alto and Taipei.

No chip has shipped yet. According to Reuters, Fractile plans to launch its first commercial chip next year. The company once claimed it could run large language models up to 100 times faster and 10 times cheaper than existing hardware, but TNW reported that those results came from simulation and small-silicon tests, and no benchmark against large deployed GPU clusters has been published. In August, Bloomberg reported that Anthropic had struck an initial agreement to buy about $250 million of Fractile chips, which would not be usable until 2027. Fractile and Anthropic declined to comment. The $7 billion valuation therefore rests on an unreleased chip and, effectively, the prospect of a single large customer.

Fractile was founded in 2022 by CEO Walter Goodwin, who holds a PhD in robotics from the University of Oxford. It emerged with a $15 million seed in July 2024 and in February announced a £100 million, three-year UK expansion plan with a hardware engineering site in Bristol. In May it closed a $220 million Series B led by Accel, Factorial Funds and Founders Fund, with former Intel CEO Pat Gelsinger joining as an angel investor and adviser. Simply adding up the rounds, the company has raised about $955 million. In its announcement, Fractile said it bet everything on the premise that the impact of the most capable AI would be limited by the time it takes to produce useful outputs, and that the only way to make speed viable at scale is to radically reinvent the hardware.

For memory makers, this is another challenge aimed at HBM, which now helps set AI chip performance, and a growing number of inference chip startups are trying to avoid or replace it. How well Fractile’s stacked DRAM holds up in production will only become clear when its first chip arrives next year.

Inference chips are one of the most heavily funded areas this year, and each company attacks the memory bottleneck differently. Cerebras turns an entire wafer into one chip, listed on Nasdaq raising $5.5 billion, and jumped 108% on its first day. Positron AI takes a “memory-first” design that uses commodity LPDDR5X instead of HBM, and raised $875 million at a $5 billion valuation. d-Matrix is closest to Fractile, using digital in-memory computing that does the math inside memory cells; it raised $275 million at a $2 billion valuation and says it is preparing 3D memory stacking. Etched etches transformer computation into silicon and raised $700 million at a $21 billion valuation, with SK hynix among investors in its Series C. Volantis wants to break the memory wall with optical interconnects and raised an $88 million Series A. Groq, by contrast, struck a roughly $20 billion licensing deal with Nvidia and stepped back from chips. While they tackle the bottleneck with SRAM, wafers, commodity memory and optical links, Fractile has chosen to stack DRAM on logic.