Anthropic confirms custom AI chips for Claude

Anthropic has formally confirmed that it is building a Custom Silicon Team to develop AI chips for the Claude model family, moving the effort beyond earlier reports of exploratory talks.

The company is hiring semiconductor engineers who can take a chip from initial design through production. A posted role lists annual compensation of $320,000 to $485,000, suggesting Anthropic is recruiting senior specialists capable of making technical decisions without a large organizational structure.

Anthropic says it wants to design models and hardware together, with the goal of improving performance and efficiency at large scale. The company has not disclosed a chip architecture, manufacturing process, memory configuration, power target, performance figure, prototype, or production schedule.

Anthropic’s multichip strategy

The custom hardware is not intended to replace Anthropic’s existing infrastructure. The company describes its approach as “multichip”: AWS, Google, Nvidia, and AMD hardware will remain important for training and running Claude, while Anthropic’s own processors would supplement those platforms.

Recommended reading

Google reshuffles DeepMind leadership as Jeff Dean leaves

That makes the project more narrowly focused than a wholesale move away from cloud and accelerator suppliers. A purpose-built accelerator could be tuned for Anthropic’s models, but the company’s own description indicates it plans to add another option to its infrastructure rather than abandon current partners.

Reports previously said Anthropic was studying custom processors and discussing possible manufacturing with Samsung Electronics. Those discussions are still described as early, with no confirmed manufacturing agreement. Anthropic has not said whether any eventual chip would be used only in its own data centers or sold to external customers.

No Anthropic chip is available yet

There is currently no Anthropic-branded chip product for sale in the US or EU. The project is reported to be at an early, pre-design stage, and Anthropic has announced neither commercial pricing nor a release timeline.

For comparison, organizations needing high-end AI compute can already rent established alternatives:

• Google Cloud A3 Ultra, using eight NVIDIA H200 GPUs, is listed at $84.806908493 per hour for the on-demand eight-GPU node.
• AWS P5 p5.48xlarge, with eight NVIDIA H100 GPUs, is listed at $41.528 per instance per hour on an on-demand basis, with rates varying by region.
• OpenAI’s “Jalapeño” is an announced inference ASIC being co-developed with Broadcom, but it is intended for OpenAI infrastructure and has no published commercial price for third-party buyers.

Anthropic’s hiring confirms a strategic direction, not a finished product. Until the company discloses a design, manufacturing partner, timeline, or customer-access model, the practical result is a new recruiting push rather than a new accelerator buyers can evaluate.