Anthropic is hiring people who have shipped silicon
Job listings at $320,000 to $485,000 and a public confirmation. Anthropic wants to co-design its own inference chips alongside its own models.
No blog post, no timeline, no named lead. The evidence is a job listing and a confirmation to Reuters.
On August 5, 2026 Anthropic confirmed to Reuters and TechCrunch that it is building an in-house custom silicon team to design AI chips for running Claude, after Business Insider reported the effort. The confirmation is backed by live job listings looking for engineers who have, in the listings' own phrase, shipped silicon, advertised between $320,000 and $485,000 and describing a deliberately small team of people comfortable making consequential calls without a large organization behind them.
What is confirmed and what is not
Confirmed: the team exists as a hiring effort, the stated goal is co-designing hardware and models together so Claude runs faster and more cost-efficiently at scale, and this is additive rather than a replacement for Anthropic's existing arrangements with AWS, Google TPUs, Nvidia, and AMD. Not confirmed: any timeline, any named team lead, or any manufacturing partner, though The Information reported separately in July that Anthropic had been in talks with Samsung about a 2 nanometer process. There is no Anthropic blog post. The entire public record is statements to reporters plus job listings, which is worth knowing before anyone treats this as a shipping roadmap.
The third lab down the same road
This is a well-worn path now. Google has run TPUs for a decade, Meta has MTIA, and OpenAI moved to a Broadcom co-designed inference accelerator earlier this year. The logic in each case is the same: at sufficient scale, inference is a recurring cost that general-purpose GPUs price inefficiently, and a chip built for one company's model shapes claws some of that back. Anthropic's version is notable mainly for the co-design framing, designing the hardware and the model against each other rather than fitting the model to whatever silicon exists.
Why a build studio cares
We do not buy chips. We do build products whose unit economics are a function of somebody else's inference cost, and every one of those products inherits the cost curve of the lab it sits on. Three of the largest model providers now believe the way to bend that curve is to stop renting general-purpose hardware, which tells you where they expect the cost of serving a token to be a few years out. It is a slow signal rather than a change you can act on this quarter, but it is the kind that eventually shows up on an invoice.
Next step: read TechCrunch's report for the confirmation and the listing details. If you want a product costed properly against real inference pricing before you build it, write to us at hello@gattyworks.com.