Anthropic Is Building an In-House Chip Design Team, Paying Up to $485K to Escape the Nvidia Tax
The Claude maker is hiring semiconductor engineers across architecture, design, and verification to co-design custom silicon with its models. The Information reports Anthropic has scouted Samsung as a potential manufacturing partner.
Anthropic is assembling an in-house AI chip design team, with job listings offering semiconductor engineers between $320,000 and $485,000 to build custom silicon for Claude. The company confirmed the effort this week: it wants to co-design hardware and models together so its systems “run faster and more efficiently.”
The hires span chip architecture, design, and verification — the full stack of roles you need to actually tape out a chip, not just spec one. And per earlier reporting from The Information, Anthropic has already scouted Samsung as a potential manufacturing partner.
Every frontier lab now designs silicon
With this move, custom silicon becomes table stakes at the frontier. Google has run Gemini on TPUs for years. OpenAI is co-developing accelerators with Broadcom. Amazon builds Trainium, Microsoft has Maia — and now the last major holdout among frontier labs is in the game.
The logic is brutal arithmetic. Anthropic’s compute obligations have ballooned alongside Claude demand: the company has stacked multi-year infrastructure deals across Google TPUs, Amazon’s Trainium clusters, and a 2GW arrangement with AMD. Every one of those contracts carries someone else’s margin. When inference is your dominant marginal cost, a chip co-designed around your own models’ attention patterns, memory access, and serving shapes is the only lever that bends the curve structurally rather than incrementally. Analysts covering the announcement estimate co-designed inference silicon could cut Claude’s serving costs roughly in half.
Co-design is the operative word. A general-purpose GPU has to run everything; a chip built alongside the model that will run on it can trade away flexibility for throughput. That’s the same thesis AMD just paid to acquire in Taalas, the startup that etches model weights directly into silicon. The industry is converging on the idea from both directions — chipmakers moving toward models, model labs moving toward chips.
Temper the timeline
Designing a competitive accelerator takes three to five years and hundreds of engineers; Google’s TPU needed several generations before it was a genuine Nvidia alternative internally. A team being hired in August 2026 won’t dent Anthropic’s Nvidia, AMD, or TPU spending before the end of the decade. The near-term value is negotiating leverage — every supplier now knows Anthropic has a credible exit — and optionality on an inference-dominated future.
There’s also a talent-market signal in the salary band. At up to $485K base for chip engineers, Anthropic is bidding against Nvidia, Apple, and Google’s silicon teams, not against other startups. The AI talent war has officially expanded from researchers to the people who design the substrate they run on.