Anthropic in talks with Samsung to manufacture custom AI chips on 2nm process

Anthropic is in early-stage talks with Samsung Electronics to manufacture custom AI chips, The Information first reported, with Bloomberg subsequently confirming the discussions. The project remains at an exploratory phase: Anthropic has not finalized the chip's function, performance targets, or its role within the company's server infrastructure.

Samsung is being evaluated as a foundry partner specifically for its 2nm process node and advanced packaging capabilities, according to TrendForce. Anthropic is also in discussions with several chip design firms, but has not yet moved into detailed design, testing, or manufacturing stages.

Why Samsung?

The talks build on an existing relationship. During Anthropic's Series H funding round in May 2026, Samsung Electronics was named one of three "strategic infrastructure partners" alongside SK Hynix and Micron, cited for their roles supplying memory, storage, and logic chips. Samsung is the only one of the three memory giants that also operates an advanced logic chip foundry — making it the natural candidate for a custom silicon partnership.

Anthropic has also been quietly staffing up for a hardware push. The company hired Clive Chan, who previously helped build OpenAI's custom chip program, signalling that custom silicon has been on the roadmap for some time.

Trainium and GPUs remain core, for now

Anthropic confirmed to Bloomberg that Amazon's Trainium chips, Google TPUs, and NVIDIA GPUs will continue to anchor its computing strategy, and declined to comment on custom chip plans. The custom chip effort, if it proceeds, would likely target specific inference or training workloads where proprietary silicon could deliver better performance per watt — not replace its cloud partnerships outright.

Racing against OpenAI

The timing matters. Just days before the Samsung talks surfaced, OpenAI unveiled "Jalapeño" — a custom inference chip developed with Broadcom that the company says delivers superior performance per watt compared to competing processors. The AI chip race is no longer confined to hyperscalers: the leading AI model companies are now building silicon to match their models, not just rent it.

If Anthropic follows through, it would join Amazon, Google, Microsoft, Meta, and OpenAI in owning a piece of the semiconductor stack — a shift that has significant implications for Nvidia's long-term position as the default AI compute provider.