Key Takeaways
- • UK-based AI chip start-up Fractile has reached a $7bn valuation in a new funding round, up from about $1.1bn in a May round, according to people familiar with the matter cited by the Financial Times.
- • Fractile raised $720mn from Thrive Capital, Founders Fund, Conviction, Alphabet's CapitalG, Lightspeed Venture Partners and others, making the four-year-old, London-headquartered company one of the UK's most valuable start-ups.
- • The company is developing AI inference chips with custom memory stacked directly on the chip, which founder Walter Goodwin says can speed response times 10-fold versus current designs; the approach targets the memory bottleneck rather than raw compute.
- • Anthropic has shown interest in becoming an early customer, according to people familiar with the negotiations.
- • Fractile has not yet released a production chip and plans to launch its first product next year, so the valuation rests on prototypes and hardware breakthroughs rather than revenue.
Summary
UK chip start-up Fractile has reached a $7bn valuation following a new funding round that drew several prominent Silicon Valley investors, according to the Financial Times. The London-headquartered company, founded in 2022, has raised $720mn from Thrive Capital, Founders Fund, Conviction, Alphabet's independent venture arm CapitalG, Lightspeed Venture Partners and others. The deal values the company at $7bn including the new investment, up from about $1.1bn in a May funding round, according to people familiar with the matter. That May round raised about $220mn in a deal led by Accel, Founders Fund and Factorial Funds. The UK government's sovereign AI fund also participated in the latest round.
Fractile is developing its own AI chips with custom memory. Founder Walter Goodwin argues that chipmakers have concentrated on raw computing power while memory bandwidth has lagged; he noted that compute has grown a million-fold over 25 years while memory bandwidth has improved about 40-fold. Fractile's design layers custom DRAM directly on top of its compute silicon in a single chip, rather than relying on separate parallel pieces of silicon as competitors do. Competitors, including Nvidia, largely rely on high-bandwidth memory (HBM), which is expensive and increasingly scarce as suppliers such as SK Hynix and Micron struggle to meet AI demand. Goodwin claims the approach could speed response times 10-fold for AI inference, the process of generating chatbot responses.
The company operates in a crowded field of custom AI chip developers, including Etched, MatX and UK-based Olix, all seeking to offer faster and cheaper alternatives to Nvidia's GPUs. Nvidia acquired chip start-up Groq last year and launched its own inference chip in March. Anthropic has shown interest in becoming an early Fractile customer, according to people familiar with the negotiations.
The valuation is forward-looking. Fractile has not yet released a production chip and plans to launch its first product next year, so the valuation is based on prototypes and hardware breakthroughs. Thrive Capital investor Philip Clark, whose firm co-led the round, said the inference market Fractile targets will soon be "at least as large as the market Nvidia is addressing today" and that the chips suit AI agents that run autonomously for hours or days. Sarah Guo, who runs Conviction, said the industry is converging on the view that memory bandwidth and capacity are the central constraints on AI inference, adding that the firm backed Fractile for its rare expertise in that area. Analysts and investors note that AI growth projections have been tempered in recent months amid rising costs and public backlash.
Fractile is developing its own AI chips with custom memory. Founder Walter Goodwin argues that chipmakers have concentrated on raw computing power while memory bandwidth has lagged; he noted that compute has grown a million-fold over 25 years while memory bandwidth has improved about 40-fold. Fractile's design layers custom DRAM directly on top of its compute silicon in a single chip, rather than relying on separate parallel pieces of silicon as competitors do. Competitors, including Nvidia, largely rely on high-bandwidth memory (HBM), which is expensive and increasingly scarce as suppliers such as SK Hynix and Micron struggle to meet AI demand. Goodwin claims the approach could speed response times 10-fold for AI inference, the process of generating chatbot responses.
The company operates in a crowded field of custom AI chip developers, including Etched, MatX and UK-based Olix, all seeking to offer faster and cheaper alternatives to Nvidia's GPUs. Nvidia acquired chip start-up Groq last year and launched its own inference chip in March. Anthropic has shown interest in becoming an early Fractile customer, according to people familiar with the negotiations.
The valuation is forward-looking. Fractile has not yet released a production chip and plans to launch its first product next year, so the valuation is based on prototypes and hardware breakthroughs. Thrive Capital investor Philip Clark, whose firm co-led the round, said the inference market Fractile targets will soon be "at least as large as the market Nvidia is addressing today" and that the chips suit AI agents that run autonomously for hours or days. Sarah Guo, who runs Conviction, said the industry is converging on the view that memory bandwidth and capacity are the central constraints on AI inference, adding that the firm backed Fractile for its rare expertise in that area. Analysts and investors note that AI growth projections have been tempered in recent months amid rising costs and public backlash.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.