Anthropic plans to build in-house AI chips as compute costs hit $19B

Anthropic is no longer content renting other people’s hardware. On August 5, 2026, the AI company behind the Claude model series officially announced the formation of an in-house custom silicon team, tasked with designing proprietary chips for both training and inference workloads.

The move puts Anthropic in the same conversation as Google, Apple, and Microsoft, all of which have spent years building custom silicon to claw back control over their compute stacks. For Anthropic, the motivation is straightforward: the company reportedly spends around $19 billion on compute in 2026, a figure that makes even the most generous investor wince.

Why build your own chip

Developing advanced AI chips carries an estimated price tag of around $500 million, which is steep but increasingly standard for frontier AI labs trying to compete on cost efficiency. OpenAI has pursued a similar path, and Anthropic’s stated goal is to bring its inference costs down to a level that matches or beats its closest rival.

The company has tapped Clive Chan, a former OpenAI engineer who joined Anthropic in early June 2026, to help lead the effort. Engineering roles tied to the silicon program carry salary ranges between $320,000 and $485,000, signaling that Anthropic is recruiting aggressively at the senior end of the talent pool.

Anthropic has also begun exploratory conversations with Samsung Electronics about chip manufacturing, with those talks reportedly starting in July 2026. Samsung would serve as a potential fabrication partner, handling the physical production of whatever designs Anthropic’s team eventually produces.

Not an either/or play

Importantly, Anthropic is not abandoning its existing hardware relationships. The company is adopting what it describes as a multi-chip strategy, meaning it will continue working with Nvidia, AMD, AWS Trainium, and Google TPUs while simultaneously building toward proprietary silicon.

The Google TPU relationship is particularly notable. Google is expanding its next-generation TPU capacity by approximately 3.5 gigawatts starting in 2027, a buildout facilitated by Broadcom. Anthropic is set to tap into that expanded capacity, which means the company will likely lean on Google’s infrastructure heavily during the gap years before its own silicon reaches maturity.

That arrangement also reflects a layered dependency that Anthropic clearly wants to eventually reduce. Google is both a major investor in Anthropic and a supplier of the hardware Anthropic depends on to run its models.

What this means for the broader AI chip market

Inference, the continuous, high-volume process of actually running a model in production, is where custom silicon can deliver the most immediate cost savings. That’s precisely where Anthropic says it’s targeting its efforts first.

The Samsung manufacturing conversation adds another dimension. TSMC handles the vast majority of leading-edge AI chip production, making it a strategic bottleneck for the entire industry. If Anthropic builds a relationship with Samsung as an alternative fab, it gains negotiating leverage and supply chain resilience, two things that matter a great deal when you’re burning through compute at this scale.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.