Fractile raises about £540m Series C at a £5bn-plus valuation, a year before its first inference chips ship

Fractile, the London and Bristol company designing chips to run AI models, has closed a Series C of more than £500m at a valuation above £5bn. The company puts the total at roughly £540m, a figure that several outlets have reported in dollars as $720m at a $7bn valuation. The money arrives before Fractile has shipped a single chip.

Thrive Capital, Founders Fund, Lightspeed and Redpoint led the deal. Alongside them are Alphabet's growth fund CapitalG, BlackRock, Dragoneer, Balderton and Shopify, plus two public investors: the British Business Bank and the government's Sovereign AI fund. The UK government announced the round in a press release on 11 November, though parts of it had been reported over the summer.

Six times the price in six months

The pace is the striking part. In May, Fractile raised $220m in a Series B led by Accel, Factorial Funds and Founders Fund, with Conviction, Felicis and 8VC also investing. That round valued the company at around £820m, or roughly $1bn after the money came in.

By August, reports said the company was in advanced talks to raise about $600m at a pre-money valuation near $6.5bn, with Lightspeed and Redpoint co-leading and some of the money coming in at a lower price. Those same reports said Fractile had agreed to supply chips to Anthropic. The final round came in larger than those early numbers, and the valuation has risen around sixfold in half a year.

Founders Fund is the only lead investor to appear on both the Series B and Series C. Its first disclosed round was much smaller: a $15m seed round in July 2024, led by Oxford Science Enterprises.

Walter Goodwin founded Fractile in 2022 after finishing a PhD in AI and robotics at the University of Oxford, and he runs it as chief executive. The startup sells itself as a chip-first company aimed at frontier AI labs, and it is pitched openly as a British alternative to Nvidia in the market for inference, the work of running a trained model rather than training it.

Stacking memory on the chip

Fractile's argument is about data movement. When a large language model generates text, a conventional accelerator repeatedly sends intermediate results out to separate high-bandwidth memory and then reads them back for the next step. Those round trips cost time and electricity, and they get worse as models grow and context windows stretch.

The company's design keeps the data on the chip. Rather than packing in large amounts of SRAM, the route taken by US rival Cerebras with its wafer-scale processors, Fractile places ordinary DRAM directly on top of the logic. With the memory sitting right above the computing circuits, data has a much shorter trip and needs less power to move. SRAM is quick, but it is costly and holds relatively little, which is why stacking denser DRAM is attractive.

The claims are large. Fractile says its hardware can run language models up to 100 times faster than current systems and cut running costs by around 90%, and it is aiming for several thousand tokens per second over long contexts. None of this has been tested in public, since the chips are not yet in customers' hands.

It is not the only company working on the idea. Samsung showed a memory product in August, called zHBM, that is meant to be stacked on top of AI accelerators. Cerebras, which listed on the stock market in May, already sells systems built around keeping data close to the processor.

Fractile plans to sell complete rack-scale systems that bundle its chips with its own software. That puts it in competition with GPU clusters at the level of whole data-centre installations, not single components.

Why London is backing it

The British state has a stake in this round through two vehicles: the Sovereign AI fund and the British Business Bank. The government framed the deal as part of an effort to own more of the AI stack at home and depend less on fragile foreign supply chains.

AI minister Kanishka Narayan said Britain had too often seen its research turned into companies and jobs abroad, and that the country needed to build more of the hardware behind AI itself.

"Fractile represents the very best of British ambition, bringing together world-class chip design, engineering and AI expertise to develop faster, cheaper and more energy-efficient hardware," Narayan said.

For European founders, the investor list is the other signal. Most of the money comes from US funds, with Thrive, Founders Fund, Lightspeed, Redpoint, CapitalG and Dragoneer all on board, while Balderton is the main European venture firm named. Deep-tech hardware companies in Europe have long struggled to raise this kind of capital locally, and Fractile's cap table suggests that gap has not closed, even as governments add public money on top.

The next test is concrete. Fractile says its first chips will be ready for deployment in data centres in 2027. Until then, a valuation of more than £5bn rests on design claims and a supply agreement, and the company will need to show that chips designed in Britain can deliver the speed it has promised to the AI labs it wants as customers.