Anthropic starts hiring a custom AI silicon design team
Anthropic posted roles for a custom silicon team, saying it intends to co-design hardware and models so Claude runs faster and more cheaply.
Anthropic is assembling an in-house chip design group, according to job listings for a "custom silicon team" published Wednesday. The company said it plans to co-design hardware alongside its models so that Claude runs faster and more efficiently.
What the listings describe
The roles seek engineers with chip design backgrounds. The stated goal is accelerators specialised for Anthropic's own workloads rather than general-purpose AI silicon, with hardware and model architecture developed in tandem. That framing matters: co-design implies Anthropic wants influence over memory hierarchy, interconnect and numerics decisions that today are set by its suppliers and then accommodated in software.
Anthropic currently runs on a mix of external hardware — Amazon's Trainium through AWS, Google TPUs, and Nvidia and AMD GPUs — and has layered on large compute commitments, including AMD's stated plan to invest up to $5 billion as part of a capacity deal. The Information has separately reported that Anthropic scouted Samsung as a possible manufacturing partner, though no fabrication agreement has been announced.
Context
The move follows OpenAI unveiling Jalapeño, its Broadcom-built accelerator, in June 2026. Google has shipped TPUs for a decade, Amazon has Trainium and Inferentia, and Meta has its MTIA line. Anthropic has until now been the largest frontier lab without a silicon programme of its own, relying instead on multi-vendor supply agreements.
Why it matters
Inference cost is the binding constraint on frontier-model economics, and it is increasingly a pricing weapon — the same day, Meta priced its new coding model well below Anthropic's flagship tier. Custom accelerators are the main lever a lab controls that is not simply buying more capacity: they can cut cost per token and reduce exposure to GPU allocation cycles. The caveat is timing. Even with an experienced team and a foundry partner, first silicon typically lands two to three years out, so this is a hiring signal about 2028-2029 unit economics rather than a change to what Claude costs today. It also nudges the industry further toward vertical integration, where the most capable models increasingly run only on hardware their creators helped specify.