Anthropic builds in-house chip team for Claude

Anthropic has confirmed an in-house custom silicon team to design chips for Claude, marking its first public acknowledgment of a proprietary hardware program. The effort centers on software-hardware co-design, with Claude and its accelerator developed together so the chip can be tuned for transformer attention, matrix multiplication and key-value cache workloads.

The economic goal is lower inference spending at scale. OpenAI’s Jalapeño inference chip reportedly showed roughly 50% per-token cost savings versus standard GPU inference in early testing, and Anthropic is pursuing similar gains for Claude. The company is recruiting senior engineers with hardware and software experience, offering $320,000 to $485,000 annually for chip roles.

Clive Chan, who joined Anthropic in early June 2026 after work on OpenAI’s chip team and Tesla’s Dojo program, is anchoring the early technical effort. Anthropic has not disclosed a production date, architecture or manufacturing partner, though exploratory discussions with Samsung have included its 2nm SF2P process.

Proprietary silicon is being positioned as a future addition rather than a replacement for existing compute partners. AWS Trainium, Google TPU, Nvidia GPUs and AMD’s MI450 accelerators remain central to Claude today, while Anthropic’s own chip program points to tighter control over long-term cost, performance and supply chain planning.