Anthropic builds an in-house silicon team for Claude, not an Nvidia exit

Anthropic confirmed it is building an internal custom silicon team to design chips for running Claude. The confirmation followed a Business Insider find of a senior engineer listing for people who have shipped semiconductor designs; a spokesperson confirmed the plans to Business Insider and TechCrunch. Ars Technica (August 6, 2026) and Reuters placed the news in the broader race for capacity and less Nvidia dependence.

Multi-chip stays; own silicon is capacity insurance

The spokesperson said Anthropic will keep a multi-chip approach, using other companies hardware alongside its own designs. Public framing still includes Nvidia, AMD, AWS, and Google Cloud. The Information earlier reported Anthropic was weighing Samsung as a manufacturing partner; that remains a reported trail, not an announced fab deal.

As the company frames it, the goal is co-designing chips and models so Claude runs faster and more efficiently at the scale customers now demand. Hiring targets people across the hardware and software stack who have shipped silicon and can make architectural calls without a giant org behind them. Secondary reports put silicon engineer pay roughly in the $320,000 to $485,000 annual band.

OpenAI has Jalapeño, Google has TPUs, Meta has chips. Claude was missing a stack piece

Anthropic is not alone. OpenAI recently announced a custom inference chip called Jalapeño with Broadcom. Google has long run on its own hardware, Meta has designed and deployed its own chips, and Mistral is reportedly looking at the same path. Custom silicon is now part of the story about who controls inference cost and availability, not only who wins benchmarks.

For Claude API buyers and enterprise contracts, the signal is strategic. Near term, capacity still flows through existing partners. Medium term, Anthropic buys room to tune silicon to its own model architecture and stop depending only on Nvidias roadmap. That is insurance against GPU queues and supplier pricing power.

A team and job posts are not a tape-out. Wafer is years away

Confirming a team is a start, not a tape-out. From senior hires through architecture, tape-out, yield, and datacenter bring-up runs years and hundreds of millions of dollars. The multi-chip promise also means the first generation of own silicon will likely sit as a specialized accelerator beside H100/B200 class fleets, not as a full fleet replacement.

The risk is twofold. First is the talent war: people who can ship AI accelerators are already hunted by OpenAI, Google, and hyperscalers. Second is opportunity cost: a silicon program eats management attention while Claude competes for developer adoption against Muse Code, Codex, and the Claude Code stack. Without a clear inference cost or latency win, the announcement stays a recruiting PR wave.

First tape-out and non-Nvidia capacity share will show if this is more than a slide

Watch whether Anthropic names a manufacturing partner after hiring (Samsung or another foundry) and a horizon for first silicon. A second signal is how fast non-Nvidia capacity grows in training and inference mix. A third is customer price: whether co-design shows up in Claude unit economics, or only in internal supply resilience.

For European teams on Claude, this does not change SLA today. It means the vendor under their coding agent and enterprise chat is playing the same vertical game as OpenAI and Google. Whoever controls silicon controls the queue and the margin.

Lilith's verdict

Anthropic did not buy a fab. It hired people who can stamp silicon and the model onto one inference bill, while Nvidia still owns the delivery truck.