Anthropic builds in-house chip team to power Claude AI systems

Anthropic said it is building an in-house team to design custom chips for Claude, moving the company deeper into the hardware layer that powers training and inference. The push is aimed at easing dependence on outside chip suppliers and responding to shortages of processors needed for advanced AI work.

The company is not cutting itself off from the market it already relies on. Anthropic will continue using hardware from Amazon Web Services, Google, Nvidia and AMD while it builds internal chip expertise, a sign that the company wants more control over cost, performance and availability without giving up access to the scale of the cloud and semiconductor ecosystem.

The hiring effort shows how the AI race has shifted from software alone to the infrastructure underneath it. Custom silicon can be tuned to the way a model actually runs, which can improve efficiency, lower inference costs and reduce bottlenecks when demand for compute spikes. That matters for Anthropic, whose Claude family competes directly with OpenAI’s ChatGPT and Google’s Gemini in a market where larger models can quickly turn compute bills into a strategic expense.

TechCrunch said Anthropic planned to co-design hardware and models so Claude could run faster and more efficiently at scale. The Next Web said the company was looking for engineers who have “shipped silicon,” a clue that Anthropic wants people who have already taken chips from design through production. A related job posting for a Research Engineer, Chip Design RL, or Reinforcement Learning, suggests the company is also exploring ways to use machine learning inside the chip-design process itself.

The move carries the same promise and risk seen across the AI industry’s most ambitious bets. Bringing chip design in-house could give Anthropic more leverage over its supply chain and a tighter economic grip on the compute layer beneath Claude. It could also make the company less exposed to tight capacity at Nvidia and to the pricing and allocation decisions of cloud providers. But chip development is expensive, slow and technically unforgiving, which makes the payoff far from guaranteed even as the pressure to secure more compute keeps rising.