Microsoft announced Maia 300, a custom AI accelerator for Azure, on October 12, 2026, and expects deployment to begin in 2027. The chip is intended for AI inference—the computing that generates model responses—and Axios reports that Microsoft’s plan prioritizes its own products and services. (Microsoft’s announcement; Axios’s report)
Maia 300 is aimed at inference for Microsoft services
The planned role is to generate responses for Microsoft’s own products and services. Axios reports that this does not necessarily extend to third-party workloads running on Azure. Inference is distinct from training, the computing used to build or refine AI models; Axios says Nvidia remains strong in training.
Microsoft also described plans to expand its Maia systems toward multi-gigawatt deployments, with the expansion starting in 2027. Axios reported that reaching that scale would take longer than 2027.
What Microsoft’s 40% claim measures
Microsoft says Maia 300 is expected to produce 40% more AI tokens per megawatt than leading GPU-based systems. The figure compares tokens generated with a measure of power capacity; it is not a claim that Maia 300 is 40% faster in every workload. (Microsoft’s announcement)
The reported manufacturer
Yahoo Finance’s October 12 video description identified Taiwan Semiconductor Manufacturing Company (TSMC) as Maia 300’s manufacturer. A SemiWiki report published August 10 had described Microsoft as negotiating for TSMC capacity. (Yahoo Finance; SemiWiki )
