On the evening of November 11, Walter Goodwin posted four sentences to LinkedIn. The chief executive of Fractile, a British chip company few had heard of, thanked his team, acknowledged that "there is still a lot of work ahead," and promised the money would let it "ship [its] first chips and systems next year." What he did not say was the number that made the post travel: $720 million, at a reported $7 billion valuation.
The Series C was led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint. CapitalG, BlackRock, DST Global, Dragoneer, Balderton and SV Angel joined, alongside two public bodies, the British Business Bank and the Sovereign AI Fund. The UK government described the valuation as "over 5 billion pounds," roughly $6.6 billion. The Financial Times and a Reuters source put it nearer $7 billion.
Six months ago the company was worth about $1 billion. Its $220 million Series B closed in May, led by Accel, Founders Fund and Factorial, on top of a $15 million seed in 2024. That is a more than sixfold markup in half a year, for a company with no production silicon, no disclosed revenue and no public benchmark.
The bet is not on a product but on a thesis about where the cost of AI actually lives. Inference, where a trained model answers a prompt, is mostly a memory problem. Every token generated requires streaming the model's weights from memory to the processor, so the chip that moves weights fastest produces tokens fastest. Agentic workloads make that wall harder to climb every quarter.
The Chip That Isn't There Yet
Chip designers have three ways to put memory next to arithmetic, and each is a compromise. SRAM sits on the die itself: astonishingly fast and hopelessly small. Groq's LPU holds 500 megabytes at 150 terabytes per second, while an Nvidia Rubin GPU carries 288 gigabytes of HBM4 at 22 terabytes per second. A 256-chip Groq rack holds 128 gigabytes in total, not enough for a frontier model.
HBM solves capacity by stacking DRAM dies beside the processor and wiring them through an interposer, reaching 64 gigabytes and 2 terabytes per second. DRAM is denser and far cheaper per bit than SRAM: one transistor and one capacitor, against four to six. Fractile's idea is to put that cheap DRAM directly on top of the logic, close enough to behave like on-die memory: more bandwidth than HBM, more capacity than SRAM.
That was not the original plan. Fractile launched in 2022 out of Goodwin's PhD work on early foundation models, with a design that put compute and memory on one die using SRAM. By late 2023 the team concluded SRAM could not scale to the models customers wanted. The company pivoted and spent years working with memory vendors on pulling bandwidth out of commodity DRAM. Coverage as recently as August still described an SRAM chip.
The physics is promising. Academic work on 3D-stacked DRAM puts the pitch of copper-to-copper hybrid bonding at roughly one micrometre, against about ten for the through-silicon vias HBM relies on, enabling an order of magnitude more connections between memory and logic. A 2026 evaluation framework cites access energy of 0.66 to 0.88 picojoules per bit for stacked DRAM, 77 to 83 percent below HBM. Those are estimates for the technology class, not measurements of a Fractile part, and the company has not said which bonding process it uses.
Stacking has costs. Heat builds inside a 3D stack and hotspots are hard to manage, and in wafer-to-wafer bonding a single defective die can ruin the whole sandwich. Samsung's zHBM concept, which also stacks DRAM on a processor, targets about four layers rather than the twelve or sixteen of conventional HBM because of thermal limits. A 2026 modelling study found effective bandwidth peaking below roughly ten stacked layers at 7nm, then declining. Fractile has published no layer count, no bandwidth figure, no power number and no yield data.
The demand side is less mysterious. Bloomberg reported in August that Anthropic agreed an initial deal to buy about $250 million of Fractile chips, with both sides intending to expand it. Deliveries are not expected before 2027. That single order equals roughly a third of the round: either remarkable validation or a warning about how thin the pipeline is outside one customer.
The performance claims have moved too. At its 2024 debut Fractile claimed large models could run up to 100 times faster and 10 times cheaper than existing hardware. Computer Weekly reported the figure as 50 times faster at about 10 percent of the cost. Recent investor materials settle on 25 times faster at one tenth the cost. May's announcement targeted throughput rising from 40 tokens per second to more than 1,000. There is still no public benchmark.
Public money is now part of the story. The Sovereign AI Fund launched in April with 500 million pounds to back British AI companies, entering at a slightly lower valuation than the final price. In June the government's 1.1 billion pound AI hardware plan named Fractile and Olix among its domestic champions, and the round's political framing was explicit: ensuring the UK has a greater stake in the hardware underpinning its future economy.
None of this makes Fractile foolish, and none of it makes it safe. It is one of the best-funded challengers in a field whose leaders have already shipped. Etched raised $700 million at $21 billion in August and sent its first rack to Jane Street in July. Cerebras went public in May at a $56.4 billion fully diluted valuation. Nvidia licensed Groq's technology outright. Fractile's $7 billion prices a design, a team, a Taiwanese packaging supply chain and one very large customer's intent. In 2027, when the first systems are supposed to leave the building, the market finds out whether stacked DRAM was the obvious answer everyone missed or a heat problem with a term sheet. Goodwin's four sentences were the easy part.
Comments