Anthropic says it will build its own chip, and says it will keep buying everyone else's

Anthropic has confirmed that it will design the chips that run Claude itself. Yet the same announcement stated that hardware from Amazon's AWS, Google, Nvidia and AMD will remain essential.

Anthropic is hiring an AI chip design team (TechCrunch, 2026-08-05)

출처: techcrunch.com

A company that says it will build its own chips also says it will keep buying everyone else's. The two statements do not collide. What sits between them is the time it takes for a chip that is only now being designed to end up in a live server.

Why build silicon now

The cost of running a model like Claude splits into two parts. There is the cost of training the model, and there is the cost incurred every time the finished model answers a user. The second one is called inference, and it keeps growing as long as users keep arriving. Training ends once. Inference is money that goes out every day the service is alive. So shaving even a few percent off the electricity and time behind a single answer changes the shape of the whole income statement. A general-purpose accelerator is built to handle both training and inference reasonably well. A chip cut to fit one company's own model can drop the functions that model never uses, and save power and silicon area in the process. That is what Anthropic means when it says it will co-design hardware and models.

Years just to reach tape-out

A chip does not appear the moment the design is finished. Handing the circuit layout to a fab and printing the first prototype is called tape-out. Reaching that step usually takes a new team years, and validation and production ramp come after it. A chip from a team that is only now posting job ads is that far away from carrying a service. Until then the chips that run Claude still have to be bought. That is why Anthropic talks about using hardware from several vendors at once. The in-house chip is not a replacement for what it runs today. It is something that will sit beside it years from now.

What the job ad asks for

The listed salary runs from $320,000 to $485,000. The requirement is not publications but direct personal experience taking a chip through tape-out and out the door. Anthropic is looking for people who have shipped, not people who have studied.

The ad also tells you there is no semiconductor organisation inside the company yet. Design tools, fab agreements and verification staff all still have to be brought in from outside.

What Anthropic is buying with a chip

The point of custom silicon is unit cost, not a place on a performance table. OpenAI said in June that Jalapeno cut its costs in half. That figure is OpenAI's own claim and no outside verification has appeared. The Kobeissi Letter called in-house AI chips the next big trend. If that holds, then from the seller's side of the accelerator market, the largest customers are each assembling a design team. The line to watch from here is the price list. If Anthropic says its own chip is in production service and Claude's pricing or throughput moves with it, this hiring round changed the cost structure. If the rates and limits stay where they are long after the chip talk, this stage is closer to leverage in a supplier negotiation. When the renter starts building, whether price or volume moves first is what shows up here.