Anthropic, the AI company behind the Claude chatbot, is in discussions to rent server chips designed by Microsoft, according to people familiar with the matter. The talks signal a growing push by major cloud providers to develop custom AI hardware and reduce their reliance on Nvidia's dominant chips.
Why Microsoft's custom chips matter
Microsoft has been building its own AI accelerators, known internally as Athena, to power the heavy computing loads behind large language models. The chips are designed specifically for the kind of work Anthropic does — training and running advanced AI systems. By renting these chips rather than buying Nvidia's off-the-shelf GPUs, Anthropic could get hardware tailored to its needs without the upfront cost of building its own silicon.
The deal isn't final, and terms are still being hammered out. But if it goes through, it would make Anthropic one of the first major external customers for Microsoft's in-house chip design. Microsoft has primarily used its own chips for internal workloads and for its Azure cloud customers running AI models. Renting the chips to a standalone company like Anthropic would be a new step.
The shift away from Nvidia
Nvidia's H100 and B200 GPUs have become the go‑to hardware for AI, but they're expensive and hard to get. Cloud providers have been scrambling for alternatives. Amazon builds its Trainium and Inferentia chips. Google has its Tensor Processing Units. Now Microsoft's Athena joins the list, and Anthropic's interest shows that even well‑funded AI labs are looking beyond Nvidia.
Custom chips can offer better performance per watt and per dollar for specific AI tasks. They also give the chip designer — and the companies that rent them — more control over the software stack. That's a big deal when your entire business depends on squeezing every bit of efficiency out of the hardware.
But switching isn't easy. AI models are often optimized for Nvidia's CUDA platform. Moving to a custom chip requires rewriting large chunks of code, and the performance gains have to be worth the hassle. Anthropic's talks suggest they think the payoff is there.
What's at stake for Anthropic
Anthropic has raised billions of dollars from investors including Google and Zoom. It needs access to massive computing power to train its next‑generation models. Right now, that usually means Nvidia chips, either bought directly or rented through cloud providers. A deal with Microsoft would give Anthropic another option — and possibly better pricing or performance.
The company has been vocal about the need for more competition in AI hardware. Co‑founder Dario Amodei has said the industry's dependence on a single chip supplier creates risk. Renting Microsoft's chips doesn't end that dependence overnight, but it's a step toward diversifying the supply chain.
For now, the talks are private. No deal has been signed. Both Anthropic and Microsoft declined to comment. The next few weeks could determine whether the partnership actually happens — and whether other AI labs start making similar calls.
