Fractile raises $720M at $7B, six times its May valuation

Fractile, the Bristol- and London-based chip designer, has closed a $720 million Series C at a $7 billion post-money valuation, according to its investor Balderton. That is roughly six times the $1.1 billion it was worth in May, according to the Financial Times, when it raised a $220 million Series B — and it makes Fractile one of the most valuable technology companies in Britain, with a first commercial chip still a year from data centres.

The round was co-led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with CapitalG, BlackRock, DST Global, Dragoneer, Shopify, Balderton, SV Angel and Elad Gil alongside. The UK government put in money too: the Sovereign AI Venture Fund and the British Business Bank joined a round it described as worth more than £500 million, valuing Fractile above £5 billion, per the government's own announcement. Fractile put the total at about £540 million.

The bet is memory, not flops

Fractile's argument, and the reason investors are paying up before silicon exists, is that inference has stopped being a compute problem. Every accelerator in the field — Nvidia's GPUs, Google's TPUs, Meta's MTIA, Microsoft's Maia, OpenAI's own design — pairs tensor cores with HBM stacked beside the logic die and buys advanced packaging from TSMC. Fractile stacks DRAM on top of its logic instead, which its backers say delivers far more memory bandwidth than HBM and far more capacity than SRAM, enough to run large models at thousands of tokens per second at long context. The company's public claim, repeated in Tech.eu's write-up of the round, is models running up to 100 times faster at 90% lower operating cost. Those are the company's numbers, not a third-party benchmark.

The architecture has already been rebuilt once. For its first two years Fractile was an SRAM shop in the Groq and Cerebras tradition — memory on the same piece of silicon as the compute, enormous bandwidth, almost no capacity. Founder Walter Goodwin has said the team concluded in late 2023 that growing parameter counts and, more importantly, growing context windows made that unscalable, and spent two years working with memory vendors to get SRAM-class bandwidth out of cheap DRAM. The pivot is a useful tell: the same shift from "compute inside memory" to "capacity first" is playing out across the silicon field, as we wrote when Houmo moved compute inside memory and put its A20 into production.

The $250 million that moved the valuation

Valuations this size need an anchor customer, and Fractile has one on paper. In August, Bloomberg reported that Fractile had reached an initial agreement to sell roughly $250 million of chips to Anthropic and was raising at a $6.5 billion pre-money mark. The order was three years early by design; nothing shipped, and nothing ships until 2027.

That single commitment explains most of the step-up. A frontier lab willing to sign for a chip that does not exist yet is the strongest possible signal to the next investor, and Anthropic has been shopping for inference capacity all year while its own compute bill compounds. It does not, however, prove the chip works. A purchase agreement is a claim on future supply, not a deployment, and the gap between an early commitment and a live supply line is exactly where chip startups have died before.

The state money is worth reading carefully for the same reason. The government's own notes say Sovereign AI writes cheques of roughly £1 million to £10 million and does not disclose individual positions. Against a £540 million round, that is an endorsement rather than financing — useful for a Bristol expansion and a photo with the Business Secretary, not the reason the round happened.

Who wins and who pays

The winners are obvious enough. Fractile gets a war chest large enough to fund a full-stack silicon effort — chip design, systems, software and its own ML research — without another raise, and the UK gets a flagship it can point to as sovereign capability. CapitalG's presence is a small irony: Alphabet's growth fund is backing an Nvidia alternative while Alphabet builds TPUs of its own.

The losers are harder to name, because nothing about this round takes a dollar from Nvidia today. But the timing is hard to ignore. Nvidia has shed about $430 billion in market value since the Huang spy allegations surfaced, according to The Guardian, with the stock down more than 20% in ten days. Capital that was content to hold the incumbent is now looking for the second source, and the second source is being priced at private-market multiples that assume it works.

There is a ceiling on this trade. The last British company to be called an Nvidia challenger, Graphcore, raised about $700 million and was sold to SoftBank in 2024 for a reported sum far below its peak private valuation. The difference this time is a named frontier-lab customer and a memory thesis that the whole industry is converging on anyway. The similarity is that neither company had volume silicon when the money arrived.

What the skeptics say

Three objections are worth taking seriously. First, stacked DRAM on logic is a manufacturing problem before it is a performance one: yield on a bonded stack and getting heat out of a die with memory sitting on top of it are the parts that have broken previous attempts, and Fractile has not published a yield or a power figure. Second, the 100x claim is a bandwidth argument, not an end-to-end number — real inference throughput depends on batching, KV-cache management and the memory capacity you can actually buy, and none of that has been measured outside the company. Third, the input is getting more expensive, not less: DRAM is the tightest component in the industry right now, which is the same shortage that has Apple's hardware chief calling it absolutely a crisis. A company whose whole thesis is "buy cheap DRAM, put it on the logic" is exposed to the price of DRAM in a way Nvidia, with multi-year HBM contracts, is not.

The counter to all three is that the market Fractile is attacking is real. Memory bandwidth has lagged compute for two decades, and the marginal AI workload — long-running agents holding hours of context — is the one that punishes exactly that gap. We made the case in Jensen Huang says no 2029 memory glut. The HBM numbers say 2027 that the memory wall, not the flop count, sets the pace; Fractile is the purest private bet on that view.

What to watch

First silicon, in 2027. Then the Anthropic order converting from an agreement into racks in a data centre. A second lab signing would matter more than either, because it would turn a single customer into a category. And the next round's price: a flat or down mark would say the market decided the memory thesis was right and the timing was wrong.

Would you back a chip company at $7 billion before it has shipped a part? Write to us via our contact page.