Microsoft on Oct. 12 introduced Maia 300, its third-generation in-house AI chip, and said the Maia family will scale to multi-gigawatt deployments from 2027 as the primary inference system for Microsoft products and services.
According to an official blog post by Scott Guthrie, executive vice president of Microsoft Cloud and AI, Maia 300 will begin volume deployment in 2027. That points to a gradual shift in the AI compute Microsoft uses to answer user requests, moving inference workloads from Nvidia GPUs toward its own chips.
Maia 300 is built with inference first
Microsoft said Maia 300 is expected to produce 40% more tokens per megawatt than leading GPU systems. In a fully configured rack with 96 accelerators, token throughput can reach as much as seven times that of the previous-generation Maia 200.
Each chip includes eight HBM4 memory stacks, with peak bandwidth of 18 TB/s and a maximum power draw of 950 watts. A full rack consumes 167 kW, with total HBM capacity reaching 27.6 TB.
Microsoft said it designed Maia 300 as an inference-first architecture aimed at long-context reasoning, latency-sensitive conversations, and AI agent tasks that require continuous multi-step execution.
Maia 200 is already deployed in several regions
The previous Maia 200 generation has already been deployed in the United States, Europe, and Australia. Microsoft said it is currently used in products including Microsoft Copilot, running OpenAI models, Microsoft’s own MAI models, and open-source models. GitHub Copilot and Microsoft Foundry are expected to follow before year-end.
Training workloads still depend on Nvidia and AMD
Microsoft is not replacing Nvidia across the board. Alistair Speirs, a Microsoft general manager, told CNBC that model training will still be done in partnership with Nvidia and AMD. Asked about the detailed timeline for multi-gigawatt deployment, he declined to provide specifics, saying, "This industry changes quickly, and we want to deploy the right chip for the right model at the right time."
CNBC, citing a note Evercore ISI analysts sent to clients last week, said Maia is estimated to account for about 10% of Microsoft’s AI chip spending by the first quarter of 2027.
Unlike Google’s TPU and Amazon’s Trainium, Maia is not rented out to external customers. Patrick Moorhead, chief executive of Moor Insights & Strategy, said, "No one outside Microsoft has tested these chips yet."
The report also noted that Microsoft had previously been reported to be working with TSMC on mass production of Maia 300.