London AI Chip Startup Fractile Raises $720 Million at a $7 Billion Valuation, Up From $1.1 Billion in May

Fractile, a four-year-old chip startup headquartered in London, has raised $720 million in a round that values it at $7 billion post-money, up from roughly $1.1 billion in May, according to the Financial Times. The investor list includes Thrive Capital and Founders Fund as co-leads alongside Conviction, Google's independent venture arm CapitalG, Lightspeed Venture Partners and the UK government's sovereign AI fund — a haul that makes Fractile one of the most valuable startups in Britain.

Founder Walter Goodwin's thesis is that the industry has optimized the wrong variable. Compute, he argues, improved roughly a million-fold over the past 25 years while memory bandwidth improved only about 40-fold, leaving compute 25,000 times ahead of the pipe that feeds it. The most advanced AI models, in his analogy, are drinking through a straw: the chip can think far faster than it can read the model and its context from memory.

Fractile's answer is to stop treating memory as a separate component. The company is running what it calls a secret project to develop its own custom DRAM and stack it directly on top of the compute die, co-designing compute and memory in a single chip instead of bolting separate silicon together. Because the model and its conversational context stay on-chip, Goodwin claims the design can make responses up to 10 times faster — a company figure that has not been independently verified.

The commercial validation is still prospective but notable: people involved in the negotiations told the FT that Anthropic is interested in becoming one of the first customers for Fractile's chips. Philip Clark, the Thrive investor on the deal, argued the inference market Fractile targets will soon be at least as large as the market Nvidia serves today, and that agentic workloads — running unattended for hours and eventually days — are exactly what memory-centric hardware is built for.

The competitive set is crowded and increasingly well-funded. Nvidia acquired inference-chip startup Groq last year and launched its own inference silicon in March; Etched, MatX and the British company Olix are chasing the same prize. They are all, in different ways, attacking the same scarcity: high-bandwidth memory is expensive and supply is tight, with SK Hynix and Micron struggling to keep pace with AI demand.

The caveats are as concrete as the numbers. Fractile has no shipping chip; the valuation rests on prototypes and the promise of a hardware breakthrough, with the first product planned for next year. The May round — about $220 million at roughly $1 billion, led by Accel, Founders Fund and Factorial Funds — is now a rounding error against the new valuation, a six-fold markup in six months. If inference really is becoming the bulk of AI compute, the premium may look cheap in hindsight; if agentic workloads stay on GPUs, it will be remembered as the top of a very hot market.