Anthropic in talks for chips

AI chips are splitting into two businesses now — training models and serving them. Training still gets the glamour, but serving is where the meter keeps running every time someone asks Claude a question. That is why Anthropic’s reported talks with Fractile matter. The company is apparently exploring whether a small UK chip startup could help run Claude more cheaply and efficiently, instead of relying only on the usual giants. (theinformation.com) ### Why is “inference” the important word? Inference is the part after the model is built. A user types a prompt, the model generates tokens, and the hardware has to keep doing that at scale, fast enough to feel instant. For a product like Claude, that workload never stops. The catch is that inference economics can wreck margins even when demand is boomin(theinformation.com)n goals last year partly because inference costs came in higher than expected. (digitaltoday.co.kr) ### What exactly is Anthropic discussing? The reported news is narrow but important: Anthropic has recently held talks with Fractile about buying inference chips. Fractile is a London-based startup, and the discussions are described as early-stage. No contract size is public, and there may be no deal at all. But even this much tells you Anthropic is looking beyond the standard menu of Nvidia GPUs and cloud-provider hardware as Claude’s usage grows. (theinformation.com) ### Who is Fractile? Fractile is a UK chip startup founded in 2022 by Walter Goodwin. It came out of stealth in 2024 with $15 million in seed funding, and it is building inference hardware around in-memory compute and SRAM-heavy design choices. Basically, the company’s pitch is that moving data around is the expensive, slow part — so put more of the work (theinformation.com)ce. (fractile.ai) ### Why would Anthropic want another chip supplier? Because dependence is expensive. Anthropic already works across Google, Amazon, and Nvidia-backed infrastructure, and it has expanded its Google TPU relationship into a deal worth tens of billions of dollars for 2026 capacity. But serving demand is rising so fast that supplier diversity becomes a product decision, not just a procurement one. More suppliers can (fractile.ai)ns, and more flexibility in where different Claude workloads run. (anthropic.com) ### Why not just keep using Nvidia? Nvidia is still the default, but default is not the same as optimal. GPUs are great general-purpose AI engines, especially for training, but inference at huge scale has a different pain point: memory bandwidth and the cost of moving model weights back and forth. That is why startups like Fractile, Groq, and Cerebras keep a(anthropic.com) money shifts to serving them efficiently. (ft.com) ### Is this about speed or cost? Both, but cost is the deeper story. Faster responses help product quality, sure. But cheaper tokens change everything — pricing, margins, and which model gets routed to which user request. If a lab can make one class of Claude queries dramatically cheaper to serve, it can offer more generous plans, push usage higher, or reserve premium hardware for harder tasks. That is where hardware starts shaping product design. (digitaltoday.co.kr) ### Why now? Because Anthropic’s business is scaling into infrastructure math. The company’s sales have been climbing fast enough to strain server capacity, and outside estimates put its annualized revenue run rate around $30 billion by March 2026. Whether that exact figure is right or not, the direction is obvious: once usage gets big enough, shaving cost per token matters almost as much as making the model smarter. (theinformation.com) ### What’s the bottom line? This is still only a reported early conversation. But it points to the next phase of the AI race. The big labs are no longer just buying compute — they are shopping for the exact kind of compute that fits their model economics. If that keeps happening, “best model” and “best chip stack” stop being separate questions. (theinformation.com)