Fractile raises US$720 million Series C to build AI inference chips

Fractile calls itself a chip-first frontier AI company. It says it was founded in the summer of 2022 on the bet that the most capable AI systems would eventually be limited by how long they take to produce useful outputs, and that the remedy was new hardware for those models[1]. It says it has worked full-stack since, on chips and systems meant to generate thousands of tokens per second for large models at very long contexts[1].

Founder and chief executive Walter Goodwin, in a 13 May 2026 essay for an earlier round, wrote that the hardest runs already reach 100 million tokens, and that about 40 tokens per second on existing chips makes one such output a month of work. About 1,200 tokens per second would fit that into a day[2]. That is his sketch of the gap, not a Fractile benchmark. Georgian says he founded the company in 2022 as a PhD student at Oxford's Robotics Institute[6].

Balderton, Georgian and Lightspeed describe the same chip. Fractile packages DRAM, the same type of memory used in high-bandwidth memory, on top of the logic, aiming for far higher bandwidth than that and much more capacity than on-chip static RAM[5][6][7]. Lightspeed says the accelerators and rack-scale systems are meant to serve advanced models more than an order of magnitude faster, at a fraction of the cost[7]. Georgian states that aim and says no production chip has shipped[6].

The government says the aim is faster, more energy efficient and less costly hardware, across chips, software and machine-learning research[4]. Reuters reported the same claim from Fractile, on speed, energy and cost[8]. No source here gives a benchmark from a commercial chip.

Georgian dates the design to late 2023, when it says Fractile decided that longer contexts would cap any chip built for speed alone. Over the next two years the company worked with memory suppliers and manufacturing partners on stacked DRAM and advanced packaging[6]. TechCrunch called that shift a pivot to DRAM stacked on each chip[9].

Fractile is hiring a senior machine-learning compiler engineer and an engineering manager for compilers, beside silicon and manufacturing roles, in London, Bristol, Palo Alto and Taipei[3]. Balderton and Lightspeed say it designs its own chips, systems and software, including its own machine-learning research, and owns the stack rather than handing backend design to an outside partner[5][7]. TechCrunch says that stack includes a machine-learning compiler to speed models up on the chips[9]. The ads do not show that compiler in a customer's hands.

Georgian also says the four led the round, and adds new investors Dragoneer, DST Global, BlackRock, T. Rowe Price and CapitalG, plus returning investors Accel, Conviction and Felicis[6]. That list does not include the Sovereign AI fund, the British Business Bank, Balderton, Elad Gil or SV Angel.

Reuters, citing the company, also says the four led. Other investors in its account include CapitalG, which it called Google's venture arm, plus BlackRock, DST Global, Elad Gil, Balderton, SV Angel, Britain's Sovereign AI fund and the British Business Bank[8]. It attributes US$7 billion to a person involved in the deal, and does not call that figure post-money[8].

The government does not give a dollar total or the size of its own cheque. It says it does not comment on specific figures, and that a typical Sovereign AI equity investment is around £1 million to £10 million, though every company differs[4].

Georgian says the money is meant to bring the first chips and systems to customers, with a platform ramp targeted for the second half of 2027[6]. Reuters reported a plan to launch the first commercial chip next year, with no customer named[8].

Business Secretary Jonathan Reynolds said Fractile plans to expand in Bristol. AI Minister Kanishka Narayan tied the backing to faster, cheaper and more energy-efficient British hardware[4]. Georgian says Fractile has separately committed £100 million over three years to UK operations, including a larger London presence and a hardware engineering and testing site in Bristol[6].

Georgian says no production chip has shipped, so the valuation it states comes before a product is on the market[6]. Fractile and its backers say the binding limit on inference is how fast memory can feed the chip[1][5][7]. The company is still hiring compiler and silicon staff for a first chip, while the government has joined the round without disclosing its cheque[3][4].