Anthropic is building an in-house chip team to run Claude cheaper

Anthropic is staffing a custom silicon team to design chips for Claude. The company confirmed it on August 5, four months after Reuters first reported the idea was under consideration. What Anthropic won’t say is when a chip ships, who fabricates it, or which workload it targets.

What Anthropic confirmed

The statement landed at Business Insider and Reuters the same week a Silicon Engineer req went live on Anthropic’s own careers page. Anthropic said it’s hiring engineers “with experience across the hardware and software stack to help co-design custom chips and AI models” so that Claude runs “faster and more efficiently at the scale required by customers,” according to Reuters. The listing is blunter than most corporate reqs. It wants someone who “has shipped silicon, has a realistic relationship with schedules, and is comfortable making consequential calls without a large organization behind them,” in a band of $320,000 to $485,000. The disciplines run from front-end design and pre-silicon verification through physical design, analog and mixed-signal work, foundry engineering and packaging.

Nothing in the announcement points at training frontier models. Anthropic calls custom silicon the latest step in a “multi-chip approach” that keeps Amazon Web Services, Google, Nvidia and AMD hardware in the fleet, as TechCrunch reported. The distinction is where the money goes: inference serving, the job of answering every request against an already-trained model, is typically an AI provider’s largest infrastructure expense, SiliconANGLE notes. The same report says Anthropic wants Claude to help verify its own chip designs in simulation.

The precedent and the money

Anthropic is late to this. It already trains on three other companies’ parts, AWS Trainium through Project Rainier, Nvidia GPUs and Google TPUs, and it locked in roughly one million TPUs plus more than a gigawatt of capacity for 2026 in a deal Data Center Dynamics put in the tens of billions of dollars. OpenAI got to a named part first with Jalapeño, an inference ASIC co-designed with Broadcom, announced in June. Nvidia meanwhile still takes the overwhelming majority of data-center AI spend, and that share is the number every one of these silicon programs exists to dent.

The financing context makes the timing legible. Bloomberg reported on August 4 that Blackstone has sounded out investors on a second mega debt package to fund Anthropic’s use of Google chips, with one early proposal at $36 billion or more, according to Bloomberg. That would follow a $35 billion structure in which a special-purpose vehicle buys Google’s TPUs and leases them back, keeping the debt off Anthropic’s own balance sheet while Broadcom backstops the senior tranches. A company borrowing at that scale to rent accelerators has a plain reason to want silicon it owns.

Samsung keeps surfacing, still unconfirmed. The Information reported talks with Samsung Electronics about manufacturing, and TrendForce pointed at Samsung’s 2nm process node plus advanced packaging as the likely pairing. Anthropic hasn’t named a foundry. Samsung’s foundry arm has been signing AI customers all year, so the fit is credible without being confirmed.

What nobody has said yet

Four questions stay open, and each one changes what this program actually is. Anthropic answered none of them on August 5.

• Timeline. No tape-out date, no ship date, and no word on whether Anthropic fabricates anything itself.
• Workload. Inference is the strong read from the posting, but nobody has ruled out a training part.
• Design model. A Broadcom-style co-design with an ASIC vendor and a fully in-house RTL effort are very different programs.
• Foundry. Samsung 2nm is a report, not a signed contract.

Cost is the fifth unknown. Designing an advanced AI chip runs around half a billion dollars once you count engineering talent and manufacturing validation, industry figures cited by Reuters suggest. Anthropic has filed confidentially for a US listing, per the same Bloomberg reporting, so that spending eventually surfaces in a public document.

What this means for you

Nothing changes for anyone calling the Claude API this year, and probably not next year either. Custom silicon runs on a multi-year clock from team formation to production serving, and Anthropic hasn’t even named a foundry yet. The near-term signal worth tracking is price: if per-token costs on Claude start falling faster than the rest of the market late this decade, this team is the reason. Until then, the multi-chip promise is the part that touches your capacity planning, because Anthropic’s serving fleet keeps running on Google TPUs, AWS Trainium and Nvidia GPUs. Plan against those roadmaps. And read “custom silicon team” as a hiring stage. The reqs are open. The chip doesn’t exist yet.