Anthropic in early talks with Samsung to develop a custom AI chip

Anthropic has held early-stage discussions with Samsung Electronics about developing a custom AI chip, with Samsung's 2-nanometer manufacturing process and advanced packaging facilities under consideration as a possible fabrication partner, according to reporting confirmed by multiple outlets on July 2.

What's new

The talks are described as exploratory: no chip design work has started, and Anthropic has not yet settled on what the processor would do, how powerful it needs to be, or how it would fit into a server rack. The company may ultimately choose not to proceed. Asked about a potential Samsung partnership, Anthropic told TechCrunch it "had nothing further to add" beyond a general statement on its hardware strategy: Anthropic said "a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy."

To signal it's serious about building custom silicon, Anthropic has hired Clive Chan, an early member of OpenAI's own chip team, according to reporting. The move follows an April report that Anthropic was already considering building its own chips in response to ongoing supply constraints across the industry.

Samsung has a pre-existing financial relationship with Anthropic: Samsung, SK Hynix, and Micron all participated in Anthropic's $65 billion fundraising round in May. Samsung also already manufactures chips for Nvidia and is separately working with both Nvidia and Google on chip-making initiatives in South Korea.

Context

A custom chip effort would put Anthropic in the same position OpenAI reached roughly a week earlier, when it confirmed a custom chip partnership (codenamed "Jalapeño") with Broadcom. Google has run its own TPU program for years, and Amazon builds Trainium chips in-house — leaving Nvidia's general-purpose GPUs as the remaining major frontier lab not (yet) paired with a captive custom-silicon program of its own. Anthropic currently relies on a mix of Nvidia GPUs, Google TPUs, and Amazon Trainium chips, and is reportedly also in discussions with Microsoft and UK chip startup Fractile — a multi-vendor posture the company has repeatedly emphasized rather than a wholesale pivot to custom silicon.

Why it matters

Every frontier lab now training and serving frontier models at Anthropic's scale faces the same bottleneck: GPU supply is finite, expensive, and controlled by one dominant vendor. A custom chip — even a narrow one purpose-built for a specific inference or training workload — is a hedge against that dependency and a potential cost lever at scale. That OpenAI and Anthropic have both moved on custom silicon within the same week suggests the economics of frontier-scale compute have crossed a threshold where even labs without Google's or Amazon's hardware history are willing to take on the multi-year engineering risk of chip design. If Anthropic proceeds, Samsung's role as both an investor and a prospective foundry partner would tie one of Anthropic's financial backers more tightly into its infrastructure roadmap.