Anthropic is negotiating a partnership with British semiconductor startup Fractile to secure custom chips for AI inference, aiming to cut costs and reduce dependence on Nvidia.
Founded in 2022 by Oxford PhD Walter Goodwin, Fractile uses a novel "memory-compute fusion" design that keeps data on the chip using SRAM. The startup claims this can run large language models up to 100 times faster and cut operational costs by 90%.
The company, which counts Amazon and Google as major investors, currently relies on Nvidia's H100 units and cloud partners' processors. Sky-high prices and limited availability have squeezed margins, pushing firms like Anthropic to seek alternatives.
A deal with Fractile-which is reportedly raising $200 million at a $1 billion valuation-could make it Anthropic's fourth chip supplier, alongside Nvidia, Google, and Amazon. However, talks are early, and Fractile's chips aren't expected for data center deployment until 2027.