Walter Goodwin likes a particular number. Over 25 years, he says, chip performance has improved about a millionfold while memory bandwidth has grown roughly fortyfold, leaving a gap of some 25,000 times between what a processor can calculate and how quickly it can be fed. His phrase for the result is that frontier models are “drinking through a straw.”
Goodwin is the founder of Fractile, a London company that raised $720 million in a Series C round this week at a valuation of $7 billion. The Financial Times first reported the terms, citing people familiar with the deal. Four years ago the company did not exist. It has not shipped a commercial chip.
The round was led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with participation from Alphabet’s CapitalG, DST Global, BlackRock, Bain Capital Ventures, T. Rowe Price, Shopify Ventures, Balderton, the UK’s Sovereign AI fund and the British Business Bank.
The valuation is the striking part. In May, Fractile raised about $220 million at a valuation of roughly $1 billion. Six months later it is worth seven times as much, a figure that places it among the most valuable private companies in Britain and ahead of most listed semiconductor firms of its age.
What investors are buying is an argument about where AI performance now binds. Training large models is limited by raw computation, which is why accelerators from Nvidia sell at a premium. Inference, the process of answering a query, is limited by memory: a model with a long context window must move enormous quantities of data between memory and logic for every token it emits.
The industry’s current answer is high-bandwidth memory, stacked beside the processor and bought from SK Hynix, Micron and Samsung. It is expensive, it is short in supply, and it ties the price of an inference chip to the price of a component that every other chipmaker also needs.
Fractile’s design stacks custom DRAM directly on top of logic in three dimensions, which shortens the distance data travels and, the company says, delivers several times the bandwidth of high-bandwidth memory while holding enough capacity to serve long contexts. Goodwin claims inference up to ten times faster than Nvidia’s approach.
Those claims have not been demonstrated in a product. The company’s first commercial chip is planned for 2027, and the $7 billion valuation rests on prototypes, simulation and the judgment of investors who have examined both.
Thrive’s Philip Clark has described the opportunity in terms that explain the price. Inference, he has said, could soon be a market at least as large as the one Nvidia addresses today, with particular demand from agents that run continuously for hours rather than answering a single question.
The competition is not standing still. Nvidia bought Groq last year and launched its own inference chip in March. Etched, MatX and Olix, another British company, are pursuing parts of the same market, and all of them are chasing a customer list that is short and famous.
Anthropic has expressed interest in becoming an early customer, according to people familiar with the discussions. No purchase agreement has been signed. For a startup, an expression of interest from a frontier laboratory is worth a great deal in a fundraising and nothing at all in a purchase order.
The comparison that matters is not to other startups but to the incumbent. Nvidia’s inference products carry margins that reflect a near-monopoly on the software that runs on them, and customers pay those margins because the alternative is rewriting their code. A challenger wins business by being faster per dollar on the workloads buyers actually run, and by being available in volume. Neither has been demonstrated outside a laboratory.
The British government has reason to cheer. Fractile is the most prominent beneficiary of its Sovereign AI fund, set up to keep a domestic stake in technology the country’s universities helped invent. Goodwin has credited the fund with helping the company build at the frontier from London. The company also has offices in the U.S. and Taiwan, where its chips will be built and its customers managed.
The financing says as much about memory as about Fractile. Three suppliers control most of the world’s high-bandwidth memory, and every AI lab now forecasts its own demand for it years in advance. A chip architecture that removes the requirement would change more than one company’s prospects.
The risk is the ordinary one for hardware startups with large valuations. The money has to become a product before the market moves. A design taped out in 2027 arrives into a bottleneck that may have shifted from memory bandwidth to power, packaging or something nobody is measuring yet, and the customers who expressed interest will be buying whatever is shipping.
Fractile now has enough capital to find out. It also has a valuation that assumes it will.
