Microsoft unveils Maia 300 AI chip that aims to reduce dependence on Nvidia and AMD

Microsoft (MSFT) has launched its latest AI chip, the Azure Maia 300, in an effort to reduce its reliance on third-party processors such as Nvidia (NVDA) and AMD (AMD).

Unveiled Monday at the OCP 2026 Global Summit in San Jose, the Maia 300 will join Microsoft’s Maia 200 and Maia 100 chips in what the company said will be multi-gigawatt deployments starting next year.

The Windows maker said this would make its Maia chips the “leading AI inference system for Microsoft products and services.”

The company said the Maia 300 delivers 40% more tokens per megawatt than “leading GPU-based systems,” a clear coup for Nvidia, and that a rack of 96 chips delivers seven times the rack-scale token throughput of Microsoft’s Maia 200.

Tokens are a basic unit of measurement in AI systems that can represent parts of words, images and videos.

The more tokens per megawatt, the more efficient an AI rack is.

An image of Microsoft's Maia 300 processor. — Microsoft Maia 300 processor. (Image: Microsoft) · Microsoft

The Maia 200 chip currently powers its Microsoft production workloads, including Microsoft 365 Copilot, as well as OpenAI models (OPAI.PVT), Microsoft MAI models, and open source models. This provides 1.7 times more tokens per watt for AI capabilities in Microsoft’s Teams, Copilot and Cowork software compared to other GPU-based systems, the company said.

Microsoft and fellow hyperscalers Amazon (AMZN), Google (GOOG, GOOGL), and Meta (META), as well as AI labs OpenAI and Anthropic (ANTH.PVT), already offer or are working on their own custom chips.

This is a safeguard against over-reliance on processors from Nvidia and AMD. Producing their own semiconductors also allows companies to tailor them to their specific needs.

540.23 +5.16 (+0.96%)

At 10:35:21 a.m. EDT. Open market.

This is a potential problem for Nvidia, which regularly generates more than 50% of its quarterly revenue from hyperscalers.

Additionally, Amazon and Google are now leasing their custom chips to third-party customers, posing a direct threat to Nvidia’s business. Additionally, Amazon plans to sell chips for customers to install in their own data centers rather than managing them from Amazon’s own facilities.

This doesn’t necessarily mean hyperscalers will stop buying Nvidia chips altogether. Customers still want access to a certain number of chips for different workloads, which means the Amazons and Microsofts of the world will need to continue to stock up on the most in-demand processors.

And for now, these are Nvidia’s.