Microsoft to rely mainly on its own Maia chips for AI inference

what happened

Microsoft plans to increase its use of in-house artificial-intelligence accelerators after years of depending on Nvidia graphics-processing units. CNBC reports that, beginning with the third-generation Maia chip unveiled on Monday, the company expects its own silicon to handle most incoming requests. Azure remains Microsoft’s main AI strength, running models for clients, while Microsoft 365 Copilot and GitHub Copilot provide AI services for office workers, developers and other users.

Microsoft launched its first AI chip, Maia 100, in 2023 and introduced Maia 200 in January this year. Production of Maia 300 is scheduled to begin in 2027. Microsoft expects to install 96 Maia 300 chips in each data-center rack, twice the number used for Maia 200. Compared with GPUs, the new version can process up to seven times more AI tokens per rack and 40% more tokens per megawatt. The company is targeting multi-gigawatt deployments across the Maia family, although Alistair Speirs did not give a timetable.

CEO Satya Nadella previously said Maia 200 delivers 30% better performance per dollar than the latest-generation hardware in Microsoft’s fleet and is handling OpenAI models. Microsoft will continue working with Nvidia and AMD on model training, while its accelerators are not yet available for hourly rental by other companies. Evercore ISI analysts estimate that Maia could account for about 10% of Microsoft’s AI-chip spending by the first quarter of 2027.

why it matters

The move gives Microsoft greater control over chip supply and computing costs as demand for AI capacity remains strong. Nvidia is still expected to remain significant, particularly for training models, making the announcement relevant to both the semiconductor industry and the cloud-computing market.

instruments