Anthropic in talks to buy inference chips from UK AI chip startup Fractile
The AMW Read
Novelty 2: updates Anthropic's silicon strategy with a new supplier candidate not previously in corpus. Significance 2: segment-level impact on inference chip competition and compute cost dynamics for frontier labs.
Anthropic in talks to buy inference chips from UK AI chip startup Fractile
Anthropic is in early-stage discussions to purchase inference chips from UK-based startup Fractile, according to a report from The Information. The talks, which involve chips that could ship as early as next year, aim to add Fractile as a supplier alongside existing partners Google, Amazon, and Nvidia. Anthropic expects annual server and chip spending to reach tens of billions of dollars and is seeking to diversify suppliers to improve negotiating leverage, while reducing reliance on any single vendor — a strategy that contrasts with rivals OpenAI and xAI, which depend heavily on Nvidia.
This move fits a recurring pattern we track closely: AI labs pursuing chip supply optionality to escape the hyperscaler distribution moat and Nvidia's pricing power. Anthropic already rents a range of server chips and recently signed a deal to use Google TPUs outside Google Cloud data centers. The push for Fractile, Cerebras, and Groq — startups using SRAM-based architectures to avoid high-bandwidth memory costs — reflects the structural pressure of inference economics. The article explicitly notes that Anthropic's gross profit margin for AI product operations fell short of targets last year due to higher-than-expected inference costs, the same problem OpenAI faced.
If completed, the Fractile deal would represent a meaningful step in the capital-compression arc gripping frontier labs: the race to bend inference cost curves while sustaining model quality. SRAM-based inference chips offer a potential escape from the HBM cost structure that currently constrains margins. However, the talks remain early and no deal is guaranteed. The outcome will signal whether inference chip startups can break into the hyperscaler-dominated AI infrastructure stack, or whether Nvidia's ecosystem lock-in will absorb yet another alternative.
