Anthropic confirmed on August 5 that it is hiring an in-house semiconductor team to design custom chips for its Claude AI models, co-designing hardware and models so the service runs faster and more efficiently at the scale customers now demand. The effort targets inference workloads — the serving side of generative AI — rather than training.

A recent job posting for the custom silicon team listed a salary range of $320,000 to $485,000 and called for engineers who have taken semiconductor designs through development, verification, and production, with authority to make design and scheduling decisions within a small team. Industry sources estimate the design and verification of a high-end AI processor costs roughly $500 million before production begins.

Samsung has been reported as a potential manufacturing partner, while design partners could include Broadcom or Marvell, both major custom-chip design firms. Anthropic has not disclosed a timeline for a finished chip or confirmed whether it will manufacture the silicon itself.

What's new

  • Anthropic is recruiting a cross-disciplinary hardware and software team to build its first custom AI accelerator.
  • The project focuses on inference, joining a wave of custom silicon from Google, Amazon, Meta, Microsoft, and OpenAI aimed at the same bottleneck.
  • Custom silicon will complement, not replace, existing compute relationships.
  • Anthropic currently runs Claude on AWS Trainium2 chips through Project Rainier — a cluster that grew from roughly 500,000 Trainium2 processors at launch to a projected one million by the end of 2025 — as well as on Google Cloud TPUs, Nvidia GPUs, and AMD Instinct accelerators. AWS says the infrastructure is used for both developing and deploying Claude.

Why it matters

Inference has become the primary cost and power bottleneck for large-scale AI deployments. Microsoft's Maia 200 accelerator delivers a claimed 30 percent better performance per dollar than the prior generation in its fleet, while Google's seventh-generation Ironwood TPU roughly doubles performance per watt over its predecessor.

Custom designs also let companies tailor memory bandwidth and networking to their specific model architectures; Ironwood provides 192 GB of HBM per chip and 7.37 TB/s of memory bandwidth, and Maia 200 pairs 216 GB of HBM3e with a 750-watt system-on-chip envelope. AWS projects Trainium3 will deliver up to five times more output tokens per megawatt at similar latency.

Anthropic's explosive growth — spanning consumer adoption, government contracts, and agentic workloads that generate far more tokens per interaction than traditional prompts — makes the economics of inference increasingly consequential. Analysts estimate custom ASICs can lower total cost of ownership by up to 65 percent versus merchant Nvidia GPUs for inference-heavy workloads.