Anthropic eyes custom chips with Samsung

Watching rival OpenAI push deeper into custom silicon, Anthropic is taking its first serious steps toward building hardware of its own.

According to The Information, the Claude creator is in early-stage discussions with Samsung Electronics to manufacture a custom AI chip. The prospective deal could see Anthropic leverage Samsung's cutting-edge nanometer process and advanced packaging technologies.

The effort is still in its infancy. Anthropic has not yet kicked off detailed design work, nor has it finalized what the chip will actually do, how much compute power it will pack, or how it will integrate into data center server clusters.

Poaching talent and eyeing alternatives

Rumors of Anthropic looking into in-house silicon first surfaced in April via Reuters as a defense against chronic GPU shortages. The effort gained real teeth recently when Anthropic hired Clive Chan, an early member of OpenAI's custom silicon group.

When pressed for comment, Anthropic told TechCrunch that maintaining a diversified hardware stack spanning Google, Amazon, and Nvidia remains central to its compute strategy, declining to comment specifically on talks with Samsung.

The timing is hardly a coincidence. Just last week, OpenAI teamed up with Broadcom to reveal Jalapeño, its custom-built inference processor engineered for superior performance-per-watt efficiency. With Google scaling its internal TPUs and Amazon pushing custom cloud silicon, staying entirely dependent on merchant silicon is becoming a competitive liability.

For Samsung, the discussions represent a massive opportunity. The South Korean tech giant is aggressively battling TSMC for advanced AI foundry contracts. While Samsung already supplies Nvidia and collaborates with the chipmaker on software and a domestic AI factory, landing Anthropic alongside past discussions with Google would be a major stamp of credibility for its advanced packaging and foundry operations.

Why it matters

For everyday developers building on Claude, don't expect cheaper API tokens anytime soon. Designing and deploying production-grade silicon takes years, meaning Anthropic's current reliance on Amazon Trainium, Google TPUs, and Nvidia GPUs will remain unchanged for the foreseeable future.

Across the broader landscape, this confirms that frontier AI labs view hardware independence as an existential requirement. As training costs soar and inference workloads scale, relying exclusively on Nvidia's margin-heavy GPUs is no longer a sustainable long-term playbook.

We think knocking on Samsung's door is a clever strategic move. TSMC's advanced nodes and CoWoS packaging lines are essentially sold out to Nvidia, Apple, and AMD for months to come. By engaging Samsung, Anthropic likely gains far more leverage, lower fabrication costs, and dedicated capacity that TSMC simply cannot guarantee right now.

The lingering question mark is execution. Anthropic is starting from scratch without a defined microarchitecture or server blueprint, while OpenAI and hyperscalers are already iterations ahead. Paired with historical market skepticism surrounding Samsung's foundry yields, this ambitious silicon gamble looks promising on paper, but it faces a steep climb to production.