Microsoft to Deploy Maia 300 in Data Centers Starting Next Year, Cutting Nvidia Reliance

On October 12, 2026, Bloomberg was the first to report that Microsoft is preparing to install the latest version of its own chips in data centers, aiming to power its AI software mostly with homegrown hardware and further reduce its reliance on third-party AI chips. cnBeta then reported that Microsoft officially unveiled Maia 300, its third-generation in-house AI chip, with plans to deploy it to production environments starting in 2027 to support AI inference behind products and services such as Copilot. Microsoft said that compared with current leading GPU systems, Maia 300 is expected to process 40% more tokens per megawatt of power, and that in a full rack configuration of 96 accelerators, rack-level token throughput could reach up to 7x that of the previous-generation Maia 200. The chip uses TSMC's N2P process and a chiplet design. Maia 300 is the third product in the Maia line, following the Maia 200 announced in January, which is already used for AI workloads in several production environments including Microsoft Copilot and GitHub Copilot. Scott Guthrie, executive vice president of Microsoft's cloud business, said the company plans to scale Azure Maia deployments to multiple gigawatts starting in 2027 and make them the primary AI inference system for Microsoft products and services. Microsoft will still continue to buy Nvidia and AMD GPUs for training and inference tasks requiring different compute capabilities.

Later the same day, Wall Street Journal's Chinese edition cited Bloomberg in reporting that Microsoft announced Maia 300 will begin data center deployment next year, designed specifically for AI inference, with an expected 40% higher token throughput per megawatt than mainstream GPU systems, and plans for gigawatt-scale deployment to gradually reduce reliance on Nvidia GPUs. The report noted that unlike Google and Amazon, which have opened their in-house chip compute to cloud customers, Microsoft currently has no intention of renting out Maia compute directly to external customers. Rani Bork, who heads Microsoft's data center hardware business, also commented. CNBC reported that by relying more on its own chips, Microsoft stands to expand margins and bolster its AI stack.

Reports differ on the deployment timeline: cnBeta said deployment to production environments is planned to start in 2027, while the Wall Street Journal's Chinese edition sa