Anthropic has officially confirmed that it is building its own team to design custom AI chips. According to the company, inference costs for the Claude language model are expected to fall by up to 50 percent as a result. The technical lead is Clive Chan, who previously worked on OpenAI’s chip program.
Chip and model take shape together
Anthropic is pursuing so-called co-design: the chip’s architecture and the structure of the Claude model are meant to be developed together, instead of adapting software to finished hardware afterward. As Data Center Dynamics reports, Anthropic is following the same approach as Apple with its M-series chips or Google with its own Tensor Processing Units. The industry views the method as a path to significant efficiency gains. The target of 50 percent lower inference costs is a figure stated by Anthropic itself, not independently verified.
The program is led by Clive Chan, who joined Anthropic in June 2026. Chan was previously the second hardware employee on OpenAI’s chip team and spent about two and a half years there on the Broadcom-codesigned inference chip “Jalapeño.” Before that, he gained experience with Tesla’s in-house AI accelerators on the Autopilot team.
To build out the team, Anthropic is specifically recruiting semiconductor engineers, with annual salaries of up to $485,000 - a sign of the scale of the planned investment. Early talks about manufacturing with South Korea’s Samsung have been underway according to Tom’s Hardware since July 2026, using a 2-nanometer process. The chip’s purpose, server configuration, and performance figures reportedly remain undecided.
Nvidia, Amazon, and Google stay on board
A switch to exclusively in-house hardware is not planned. An Anthropic spokesperson described the new team as the latest piece of the company’s existing multi-chip strategy and stressed that the mix of several suppliers remains central. Anthropic already trains and runs Claude on more than a million Trainium2 chips from Amazon Web Services as well as Google’s Tensor Processing Units.
Google and Broadcom have additionally committed several gigawatts of next-generation TPU capacity, expected to come online starting in 2027. In July 2026, a strategic partnership with AMD was added: up to two gigawatts of MI450-series graphics chips, with the first gigawatt due to go live in the first half of 2027.
The diversification gives Anthropic leverage with individual suppliers on supply bottlenecks, pricing, and development roadmaps. Just this week, the company secured additional computing power worth $10 billion through the Norwegian startup Volta. On top of that, Google linked roughly $200 billion in credit financing with partners for further Anthropic chip purchases.
Rivals are building their own AI chips too
Anthropic is not the only AI lab betting on its own compute chips. China’s DeepSeek is also developing its own inference chip to cut reliance on Nvidia and Huawei hardware, and Alibaba and Baidu are reportedly pursuing similar strategies. Chipmaker AMD, meanwhile, acquired Toronto startup Taalas on August 6, 2026 - a company that embeds entire AI models directly into silicon to speed up inference. Meta has also already started mass production of its own AI chips to double its global compute capacity.
The trend reflects industry-wide cost pressure: the more queries chatbots like Claude, ChatGPT, or Gemini answer every day, the more electricity costs and data-center depreciation add up. Specialized chips tailored to a single model promise bigger efficiency gains than general-purpose graphics processors from Nvidia. Nvidia still controls most of the market for AI accelerators, which raises the economic pressure on customers like Anthropic to develop alternatives.
It remains unclear when a first Anthropic-codesigned chip will actually reach production. The company has not given a timeline, and talks with Samsung are reportedly still at an early stage. For Claude’s customers, any effect would likely be felt only with a delay anyway, assuming Anthropic reaches its targeted cost cut at all: in the industry, several years typically separate a confirmed team launch from a shipped server chip.
