Anthropic builds custom AI chips for Claude as revenue triples

Anthropic wants its own chips. The company behind Claude is in early discussions with Samsung to design a custom AI processor, a move that would put it in the same lane as OpenAI, Meta, and Google, all of which have pursued their own silicon strategies in recent years.

Anthropic’s revenue climbed above $30 billion in 2026, up from roughly $9 billion at the end of 2025.

What Anthropic is actually building, and what it isn’t

To be clear about what’s happening here: no chip exists yet. There is no dedicated design team, no finalized architecture, and no production agreement with Samsung.

What Anthropic has done is open a conversation with Samsung about leveraging the company’s 2-nanometer manufacturing process for a potential custom chip. The specifics, including what the chip would actually do and which Claude workloads it would target, remain undecided.

Designing an advanced AI chip costs approximately $500 million, which is not a number most companies casually commit to.

Anthropic has spent years working with a range of hardware partners, including Google TPUs, Amazon’s custom chips, and Nvidia’s processors.

Why every major AI lab wants to own its own silicon

OpenAI has pursued custom chip development. Meta has built its own training hardware. Google has run on TPUs for years. Custom chips can be optimized specifically for your workloads, which means faster inference, lower power consumption, and better economics per query.

Samsung’s 2-nanometer process is among the most advanced in commercial production, putting it in direct competition with TSMC, which manufactures chips for Nvidia, Apple, and AMD.

What this means for investors watching the AI hardware space

Samsung stands to benefit from any confirmed partnership, particularly as it works to close the gap with TSMC in the advanced logic market. A high-profile AI lab choosing Samsung’s 2nm process would validate the foundry’s technology and attract other potential customers.

Anthropic is not abandoning Nvidia hardware, and the discussions with Samsung are nowhere near the stage where that would even be a question.

If the company can reduce its per-query compute costs through optimized hardware, margins improve without requiring any change to pricing or model quality. At a revenue run rate that has tripled in roughly a year, even small efficiency gains represent significant dollar amounts. The risk is that chip development is genuinely hard, expensive, and slow, and early-stage discussions have a long road before they produce anything that runs a single Claude inference.

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