Fractile raises $720 million at a $7 billion valuation for its inference chips
Fractile raised about £577 million to build inference chips that keep model weights in SRAM, right next to the compute. The company did not announce it: the numbers come from share allotment filings at Companies House.
What happened: the round the company did not announce
The filings, lodged on October 15 and October 28, record three new classes of preferred shares worth roughly £577 million. At the exchange rate most coverage uses that is about $720 million; other outlets, at a higher rate, say $774 million. It is the same round converted, not two deals.
The valuation in circulation is around $7 billion, per Reuters on November 11, citing a person involved in the transaction. In August the talk was a $600 million round at a $6.5 billion pre-money valuation. The final figure is higher, and it comes from filings and leaks, not from a company statement.
The loudest detail is the share price: the main tranche sold at about £1,452 a share, more than five times what Series B investors paid in May, when the company raised $220 million.
Read also
What Fractile does, in one sentence and in detail
It designs chips that run models rather than train them, and puts fast memory on the same silicon as the arithmetic instead of relying on external memory.
Why inference is a memory problem
Every token a model generates means reading its weights. The dominant cost is not multiplying numbers, it is moving data to the unit doing the math. On a GPU that means pulling weights from HBM, the stacked memory that sits outside the chip: plenty of capacity, but slower. That is why memory bandwidth rather than the FLOPS on the spec sheet decides production throughput.
SRAM is the opposite: memory inside the silicon itself, with tiny latencies and far less energy per access. The bet: keeping weights close is cheaper than shuttling them back and forth.
The price of that bet: capacity
SRAM is very expensive per byte. A large model does not fit on one chip, so it has to be split across many and they have to talk to each other. That turns the product into a system with its own interconnect, not a card you slide into a server, and explains why software weighs so heavily.
Why inference is where the money is
Training is a handful of customers with enormous budgets and one-off buys. Inference is the cost that repeats with every query and every agent, every day: anyone selling tokens lives or dies by that cost, and that is where different silicon can move the needle without winning on everything else.
Sovereign money joins the cap table
The investors named across coverage include the UK Sovereign AI Fund, backed by public money, alongside Lightspeed, Redpoint and Thrive Capital. Reading it as industrial policy is the right frame: the state is buying equity in the AI supply chain, not just subsidising it. In February the company had already pledged £100 million over three years to expand in London and Bristol.
The Anthropic deal: demand before product
In August, Bloomberg reported that Fractile had reached an initial agreement to sell about $250 million of chips to Anthropic, delivered once the silicon is ready. A frontier customer signing before seeing the product explains the valuation; it is also the risk, because if the schedule slips, the revenue slips with it.
The three open questions
2027
First sales are planned for 2027. Today there is no chip to buy, no published price and no independent benchmark. The performance figures the company has used are its own claims.
Manufacturing
A design is not a product. Tape-out, volume production and yields that survive a real customer are still ahead, and that stage sinks more silicon startups than physics does.
Software
Nobody beats Nvidia on silicon alone; they lose on CUDA: libraries, compilers and a decade of code already written. The same problem shows up from the other side in OpenAI's inference ASIC, and comparing it with Google's TPUs explains why memory became the battleground: Google doubled the SRAM on its latest long-context chip.
Conclusion
Memory is the new frontier of AI and capital is already paying for it in advance, sovereign money included. The real exam is 2027: whether the chip exists, performs and ships with software anyone wants to use.
