Anthropic has confirmed it is hiring a team to design custom AI chips. The stated goal is co-developing hardware and models to run Claude faster and more efficiently. Business Insider broke the news first, and Anthropic confirmed it to TechCrunch.

Last month The Information reported that Anthropic was exploring Samsung as a manufacturing partner. The company already has compute agreements with AWS, Google, Nvidia, and AMD, but buying GPUs from others is no longer enough.

This is not an isolated move. In June, OpenAI unveiled Jalapeño, its inference chip built with Broadcom. Google has used TPUs for years, Meta develops MTIA accelerators. Competition between labs is shifting from weights to silicon.

For those building agents, the indirect effect is on cost per task. When a lab designs silicon alongside its model, optimizations target specific operations: faster inference, wider contexts at lower cost, more efficient tool-use. It’s the same logic behind why, as we reported on July 28, Anthropic identified chip controls as its primary governance lever. Now it wants that lever in-house.

The story is in its early days: there are job postings, not chips. But the direction is clear, and anyone relying on Claude’s APIs for production agents should watch how the underlying infrastructure evolves.

In detail

Until now, AI labs have played the compute game by buying as many Nvidia GPUs as possible. Anthropic has agreements with AWS, Google Cloud, Nvidia, and, as of July, a 5 billion dollar investment from AMD for 2 gigawatts of MI450 GPUs. The race for compute has become the deciding factor: whoever has more silicon trains larger models and serves more users.

But buying GPUs from others has a structural limit. Nvidia’s general-purpose chips must serve many customers with different workloads. A chip designed for a single model can optimize the operations that model performs most: the attention matrix, token decoding, routing in MoE models. Google has done this with TPUs for years, and it’s partly why Gemini has inference costs that competitors struggle to match.

The key word is co-design. Anthropic talks about designing hardware and models together, where decisions on how each works are made in parallel, not sequentially. If you know your model uses a certain attention structure, you build the silicon to run it well, instead of adapting the model to an existing chip.

OpenAI made the same choice in June with Jalapeño, the inference chip built with Broadcom. Meta has its MTIA. Google has TPUs. The trend is clear: frontier labs want to control the entire stack, from weights to silicon.

For those building agents, the concrete effect is on price and latency. If Anthropic can optimize silicon for the operations Claude performs most (reading long contexts, calling tools, reasoning in steps), cost per task drops. This is the direction the market has taken for weeks: when performance converges, price per task decides, and labs try to lower it from the hardware side too.

The constraints are real. There are job postings and a listing, not working chips. The development cycle for a custom chip takes years, and Anthropic is just starting. The Samsung partnership is at the scouting stage, as The Information reports, not production. And there’s an open question: if every lab controls its own silicon, model portability and open standards become more fragile. Anthropic created MCP, the protocol for connecting agents to external systems, but hardware verticalization adds a layer of lock-in that the protocol doesn’t solve.