Microsoft will increase its use of its own AI chips instead of Nvidia GPUs, CNBC reported on Saturday.

Microsoft unveiled its third-generation AI accelerator, Maia 300, the report said. Microsoft plans to have AI request processing handled mainly by its own chips starting with Maia 300. Mass production of Maia 300 is scheduled for 2027.

Microsoft introduced its first AI chip, Maia 100, in 2023 and released Maia 200 in January. Chief Executive Satya Nadella said in July that Maia 200 delivers 30 percent higher performance per dollar than the latest hardware the company is operating.

Maia 200 is currently used to run OpenAI models on the Azure cloud.

Microsoft plans to install 96 Maia 300 chips in a single data-center rack. That is twice as many as Maia 200. AI token throughput per rack is up to 7 times higher, and tokens processed per megawatt are 40 percent higher than GPUs. Jim Kliwein (짐 클리웨인), a Microsoft technical fellow, said some Microsoft 365 Copilot features will become faster with Maia 300.

Scott Guthrie (스콧 구스리), a Microsoft vice president, said in a blog post the company will deploy the Maia family at the scale of several gigawatts. Alistair Speirs (알리스테어 스피어스), a Microsoft general manager, said, "The industry is changing very quickly, so we are leaving options open," and added, "We will deploy chips that match the right model at the right time."

Evercore ISI analysts forecast that Maia will account for about 10 percent of Microsoft's AI chip spending by the first quarter of 2027.

Unlike Google's TPU or Amazon's Inferentia and Trainium, Maia is not provided to external companies. Patrick Moorhead (패트릭 무어헤드), chief executive of Moor Insights & Strategy, said, "Microsoft has not yet had anyone outside the company test this chip."