Anthropic, OpenAI and Amazon are all racing to build their own AI chips

The scramble to design custom silicon is really one bet: loosen Nvidia's grip on the cost of running AI.

This week made something clear. The biggest names in AI no longer just want to buy their chips — they want to build them.

On July 2, TechCrunch reported that Anthropic is in early talks with Samsung to develop a custom AI chip of its own. It doesn't stand alone. The move follows OpenAI's reported effort to build a custom inference processor with Broadcom, and lands the same week Amazon confirmed it is designing its own AZ3 and AZ3 Pro chips for Echo and Fire TV devices — silicon Amazon says improves wake-word detection by more than 50%, with plans to expand into Kindle and Ring hardware from 2027.

Three different companies, three different strategies, one direction: away from off-the-shelf hardware and toward chips designed in-house.

Why everyone suddenly wants their own silicon

Almost every serious AI system today runs on Nvidia's GPUs. That has made Nvidia one of the most valuable companies on earth — and made its chips scarce, expensive, and rationed. For a company whose entire business is training and serving AI models, that is a dependency on a single supplier for its single largest cost.

Designing your own chip is a way to take that back. Custom silicon can be tuned for the exact way a company's models run, which can mean lower cost per query, more predictable supply, and less exposure to Nvidia's pricing and waitlists.

The real shift

For the last few years the AI race was about one question: who has the best model. The next race is quieter and more industrial — who can run their models most cheaply, reliably, and at scale. When Anthropic, OpenAI and Amazon all move the same way in the same week, it stops looking like a coincidence and starts looking like the industry's next battleground.