Anthropic is evaluating whether to design its own AI chips, according to Reuters, which cited three people familiar with the matter. The effort is still at a very early stage, with no dedicated engineering team in place, and the company could still decide to keep buying off-the-shelf chips instead of pursuing a custom design path.
Revenue growth is pushing compute needs higher
Compute has become one of the company’s most critical constraints. The source material says Anthropic’s annualized revenue has exceeded $30 billion, while the number of enterprise customers paying more than $1 million a year has passed 1,000. As training and inference workloads expand, any increase in chip pricing or delay in delivery can affect the pace of expansion. The pressure is immediate.
Current chip mix centers on Google TPU and Amazon silicon
Anthropic’s existing setup follows a split model. Training workloads mainly rely on Google’s TPU chips, while inference uses chips designed by Amazon. Reuters also said Google and Broadcom finalized a multi-year supply framework this week covering compute capacity at multiple gigawatt scale levels, with deliveries expected to start in 2027 and the agreement running through 2031.
Why discuss custom chips after securing long-term supply
The timing may look contradictory at first. Anthropic has just locked in a long-term compute arrangement with major suppliers, yet it is still studying an internal chip option. The article points to the other side of such contracts: deeper dependence on a small number of vendors, including their pricing, delivery schedules, and technical direction. A custom chip program would not necessarily replace outside procurement soon, but it could give Anthropic another lever.
High development costs meet a broader shift away from Nvidia dependence
Industry estimates in the source put the cost of developing an advanced AI chip at around $500 million, before accounting for elite engineering talent and long-term yield optimization. That makes the question less about one-time affordability and more about whether Anthropic can support the ongoing cost structure tied to a custom chip supply chain. The same source notes that Meta has fully deployed its MTIA chip for inference workloads, while OpenAI is also advancing its own chip effort, showing how major AI firms are seeking more control over compute infrastructure.