Anthropic builds in-house chip team to power Claude models

Anthropic said it is building an in-house team to design custom chips for its Claude models, and it is hiring engineers with experience across both the hardware and software stack. The move gives the company more leverage over the silicon underneath Claude, where chip availability, energy efficiency and data-center economics can matter as much as model quality.

The push followed months of signaling. In April, Anthropic was already weighing whether to build its own AI chips, and the company now wants staff who can help co-design chips and models. A job listing sought engineers who had “shipped silicon,” a phrase that shows how far the company is reaching beyond pure software hiring. A social-media post citing the listing put compensation at about $320,000 to $485,000 a year, while a separate Anthropic hardware job-board listing showed pay of $405,000 to $850,000.

Anthropic is not abandoning outside compute at the same time. The company said, “We plan to expand our use of Google Cloud technologies, including up to one million TPUs.” Anthropic said the expansion is worth tens of billions of dollars and is expected to bring well over a gigawatt of capacity online in 2026, a reminder that frontier AI remains brutally power-hungry and tied to supply chains that can strain electricity grids and data-center budgets.

That strategy fits a wider vertical-integration race inside AI. Google has built its TPU line for years; Amazon Web Services has Trainium and Inferentia chips for training and inference at scale. Meta expanded its Broadcom custom-chip deal through 2029, and OpenAI tapped Broadcom and TSMC for its first chip. Each company is trying to control more of the stack rather than renting every layer from Nvidia and other outside suppliers.

For Anthropic, in-house chip design is about cost, performance and bargaining power. Owning more of the silicon layer can reduce dependence on vendors, improve control over serving speed, and make Claude cheaper to run as usage grows. In a market where inference costs can balloon as models scale to millions of users, the chip team is a strategic move as much as an engineering one.