An Oxford Spinout Sold Anthropic 250 Million Dollars of Chips That Do Not Exist Yet

Fractile, a four-year-old chip company spun out of the University of Oxford, is reportedly in talks to raise around 600 million dollars at a 6.5 billion dollar pre-money valuation. In May the company raised 220 million at roughly 1 billion, in a round led by Accel, Founders Fund and Factorial. That is a more than sixfold jump in three months. The reason is an initial agreement to supply Anthropic with about 250 million dollars of inference chips, with both sides signalling they want to expand it. Bloomberg first reported the raise; treat the figures as reported rather than confirmed.

The technical pitch is worth understanding because it explains the timing. Chips designed for AI normally keep the model’s weights, meaning the numbers that make up the trained model, in separate memory chips and fetch them across a connection every time they are needed. That fetching is slow and it is where a great deal of the energy goes. Fractile, founded in 2022 by Oxford PhD Walter Goodwin, puts compute and memory on the same piece of silicon using SRAM, a faster memory type that sits directly on the die. The company claims this can run large language models up to 100 times faster than existing hardware at 90 percent lower operating cost. Those are the company’s own numbers and no independent benchmark exists, so hold them loosely.

Here is why a 250 million dollar order for chips that will not ship until 2027 is worth 5 billion dollars of valuation. Two things are squeezing the industry at once. Inference, the cost of actually running a model for every user request, has overtaken training as the dominant expense for anyone with real customers. And memory has become brutally expensive: DDR5 prices are up several hundred percent in a year because AI accelerators consume the wafers that would otherwise make ordinary DRAM. A chip architecture that needs far less external memory is therefore attacking both problems at the same moment. It also explains why Anthropic, which has been reported committing tens of billions to compute and is heading toward a public listing where margins will be examined, is willing to place a bet this early.

The caveats are large. Nothing here has been demonstrated at scale outside the company, and building a genuinely new chip architecture is where most silicon startups die. Anthropic’s order is an initial agreement, not a locked contract, and a sixfold valuation increase in three months on the back of one customer reads differently depending on how the next two years go.

What this means for you: nothing you can buy, and no price you pay changes this year. What is worth noticing is the direction. Almost every AI cost story right now, from memory prices to model discounting to Anthropic reportedly buying an efficiency software company, comes back to the same thing: running models is expensive and everyone is hunting for a way to make each chip serve more requests. If that hunt succeeds, the result is eventually cheaper access for the rest of us. There is a smaller European point too: a UK company built serious hardware, and hardware is the layer where Europe has been almost entirely absent so far.

Sources

• Bloomberg: AI chip startup Fractile in talks for 6.5 billion dollar value after Anthropic deal
• The Next Web: A British AI chip startup is raising at 6.5bn after an Anthropic deal
• Tom’s Hardware: Anthropic in early talks to buy DRAM-less AI inference chips from UK startup