Anthropic Starts a Custom Silicon Team for Claude

Anthropic Is Hiring Chip Designers. What It Confirmed Is Co-Design, Not a Chip.

Job listings for a custom-silicon team are real and Anthropic acknowledges the plan. The Samsung foundry story around it comes from The Information, and no manufacturing agreement exists.

Aug 6, 2026Source: TechCrunch

Anthropic confirmed on 5 August that it is building an internal chip effort, described in its own words as a plan to co-design hardware and models so Claude runs faster and more cheaply. Job listings seek engineers with silicon-design experience for a custom silicon team. Business Insider reported it first; TechCrunch published at 07:13 PDT.

What is confirmed and what is sourced

Confirmed: the team, the hiring, the intent. Reported by The Information, not confirmed: that Anthropic has held exploratory talks with Samsung about foundry capacity. Trade coverage has since attached a specific node — Samsung's 2nm-class SF2P — and a target of roughly halving per-token inference cost. Neither figure appears in anything Anthropic has published.

Everyone else got there first

OpenAI unveiled its Broadcom-built Jalapeño part in June 2026. Google has run TPUs for a decade; Meta has MTIA; Amazon has Trainium. Anthropic is the last of the frontier labs to start, and it starts with a hiring page rather than a design.

Why now

Anthropic buys capacity from AWS, Google, Nvidia and AMD, and separately committed to a $10bn, six-year deal for 121MW of Vera Rubin capacity at a Norwegian site — a number reported by Bloomberg that Anthropic itself declines to confirm. Every one of those is a rented margin. Silicon designed against your own model shapes is the standard answer, and the standard answer takes years.

Distance to a part

From a first hire to a working accelerator is typically three to four years and hundreds of millions of dollars, and OpenAI needed Broadcom to get there. Anthropic has announced the intention, not a schedule.

Why It Matters

The cost of serving a frontier model is now a bigger constraint than training it. A lab that does not control its inference silicon pays someone else's margin on every token — but announcing a team is several years short of paying less.