Anthropic negotiating Fractile chip deal to diversify inference supply

AI chips are becoming the real bottleneck in the model race. Not training chips, even — inference chips, the hardware that serves answers after a model is built. That matters because once a chatbot gets popular, the expensive part is often not making the model smarter. It is paying to run it every minute of every day. That is why Anthropic has reportedly opened talks with London startup Fractile about buying future inference chips instead of relying only on Nvidia, Google, and Amazon. (theinformation.com) ### What is Anthropic actually trying to buy? Anthropic is not shopping for generic servers. It is looking at a more specialized kind of processor meant for inference — the step where Claude generates tokens for users. The talks appear to be early, and there may never be a deal. But the point is clear: Anthropic wants another source of supply as usage grows and server spending heads toward the tens of billions of dollars annually. (theinformation.com) ### Why does inference matter more now? Training gets headlines because it sounds dramatic — giant clusters, giant runs, giant budgets. But inference is the meter that never stops running. Every user prompt hits the bill. If your model gets widely adopted, inference becomes the operating business. That is why companies are suddenly obsessed with hardware that can answer t(theinformation.com)ft. (fractile.ai) ### So what is Fractile’s angle? Fractile says it is building systems for frontier-model inference that can run up to 25x faster at 1/10th the cost. The company’s basic claim is that today’s AI hardware wastes too much time moving model weights back and forth between compute and memory. Fractile is trying to collapse more of that traffic into the chip itself with a(fractile.ai) the part of the workload that turns into latency and power bills. (fractile.ai) ### Why not just keep buying Nvidia? Because Nvidia is powerful, but dependence is expensive. Anthropic already mixes suppliers more than some rivals do, using Nvidia GPUs while also working with Google TPUs and Amazon infrastructure. Adding Fractile would not replace those relationships. It would give Anthropic one more negotiating lever and one more path if supply stays tight or pricing stays painful. For a lab serving a fast-growing product, optionality is a strategy. (digitaltoday.co.kr) ### Is Fractile big enough to matter? Not yet in shipments — that is the catch. Fractile was founded in 2022 and is still pre-scale, though it has raised seed funding, expanded in the UK, and has been in talks to raise more than $200 million at around a $1 billion valuation. So Anthropic is not buying certainty here. It is buying a future option on a startup that might have the right architecture if inference economics get harsher. (fractile.ai) ### Why negotiate this early? Because chip supply is not something you solve when the servers are already full. If Fractile’s products arrive around 2027, Anthropic has to line up access well in advance. That is normal in semiconductors and especially normal in AI, where capacity, packaging, networking, and datacenter power all create long lead times. Waiting until launch would mean competing for whatever is left. (theinformatio([fractile.ai)s/anthropic-talks-buy-ai-chips-u-k-startup)) ### What does this say about the AI race? It says the contest is getting more industrial. Better models still matter. But once several labs are close enough in quality, the next edge is who can serve answers cheaply, quickly, and at huge scale without wrecking margins. That turns AI into a supply-chain game as much as a research game. Anthropic talking to Fractile is one (theinformation.com)before everyone else does. (theinformation.com)