Anthropic Stands Up Custom-Chip Team, Keeps Multi-Silicon Mix

The interesting part of Anthropic quietly building out its own silicon effort is not the chip itself, which does not yet exist in any concrete form, but the shape of the strategy around it. Business Insider frames the effort as a multi-chip approach rather than a bet-the-house move to a single custom part, and that framing matches everything else the company has been telegraphing this year.

Anthropic already runs Claude across a mix of Amazon's Trainium accelerators, Google TPUs designed with Broadcom, and Nvidia GPUs. Rather than trying to collapse that stack, the plan reads as adding an in-house design as one more option. The Information's reporting, republished via Yahoo Finance, said the early exploratory talks with Samsung target the foundry's 2nm process and advanced packaging, though the conversations are preliminary and Anthropic has not decided what the chip will do, how it will fit into a server, or how powerful it needs to be.

The personnel signal is the more concrete one. Anthropic hired Clive Chan from OpenAI's chip team in early June 2026; Chan spent about two and a half years on the Broadcom-designed inference accelerator OpenAI later unveiled as 'Jalapeño'. Bringing in someone who has already lived through one hyperscaler chip program is a different kind of commitment from the earlier exploring-only posture Anthropic had this spring.

The reason any of this matters is scale. Claude's run-rate revenue surpassed $30 billion in 2026, and Anthropic has separately committed more than $100 billion over ten years to AWS while signing multi-gigawatt expansions with Google and Broadcom for TPU capacity coming online starting in 2027. At that size, even single-digit efficiency wins from silicon designed around the workload compound into billions, and being locked to one vendor's pricing starts to look like a strategic liability rather than a procurement detail.

The honest caveat is that a lot of this reporting is still tentative, and Anthropic itself has said no final design has been selected and it may still decide to keep buying from outside vendors. What the reporting does not give you is a chip name, a target workload, a tape-out date, or how many engineers sit behind Chan on this. The forward-looking read is that the winners here are the ones who can serve a customer with real optionality: Samsung foundry gets a frontier-lab reference win it badly needs, AWS and Google keep their multi-year contracts, and Nvidia's frontier-training customer list keeps quietly getting shorter.