Fractile, a London-based AI chip startup, has raised $720 million in a Series C round that values the company at $7 billion, up sharply from roughly $1 billion in May 2026. The round was led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint, with new backers including Dragoneer, DST Global, BlackRock, CapitalG, T. Rowe Price, Shopify and British Business Bank joining returning investors Accel, Felicis and Conviction. Fractile announced the round on November 11, 2026, and says the capital will fund production of its first commercial chip, due to ship next year.
Fractile was founded in 2022 by Walter Goodwin, an Oxford researcher, on the bet that memory bandwidth, not raw compute, is the real bottleneck limiting how fast large AI models can run. Its core product is a stacked-DRAM chip architecture designed to keep more of a model's weights close to the processing logic, cutting the data-shuttling delay that slows inference on conventional GPU setups. The company has offices in London, Bristol, Palo Alto and Taipei, and is hiring across all four.
Why inference speed matters for physical AI
Fractile's pitch is narrowly about inference — running an already-trained model quickly and cheaply, as distinct from training it. That distinction reaches well beyond chatbots. Every humanoid robot, delivery drone or self-driving vehicle running a vision-language-action model, whether onboard or on a nearby edge server, depends on inference speed to react to its surroundings without lag. Slow inference means a robot that hesitates before catching a falling object or a car that brakes a beat too late.
Nvidia still dominates AI chips largely on the strength of its training hardware, but a crowded field of inference specialists is chasing the same gap Fractile is targeting, including Groq, Cerebras, SambaNova, Etched and UK rival Graphcore. Fractile's bet is that its full-stack, memory-first design will let it serve large models more cheaply per token than general-purpose GPUs, a claim it has yet to prove at commercial volume since no chip has shipped. The jump from a roughly $1 billion valuation in May to $7 billion now, before any product has reached a customer, signals how much investor appetite exists for an alternative to Nvidia, but it also raises the bar for what Fractile has to deliver once chips actually leave the fab.
What happens next
The test for Fractile is simple: ship the chip. A 2027 launch window gives the company roughly a year to prove its stacked-DRAM approach works at the performance and cost levels it has promised backers who collectively now have more than $700 million riding on the bet. Until then, the valuation rests on architecture claims rather than shipped silicon, and rivals including Groq and Cerebras aren't waiting around. BusinessCloud reported the investor list first.