Microsoft plans 2027 deployment of Maia 300 inference chips

Microsoft announced on October 12, 2026, that its third-generation Azure Maia 300 accelerator would begin production deployment in 2027. The company said the broader Maia family would start scaling toward multi-gigawatt deployments that year. Reuters reported that Microsoft could deploy more than one million Maia 300 chips in 2027. That figure is a potential deployment, not installed capacity or a public chip-count commitment from Microsoft.

Production use and deployment timetable

Microsoft says Maia 200 already serves production workloads, including Microsoft Copilot, and runs OpenAI, Microsoft MAI and open-source models. It plans to add GitHub Copilot and Microsoft Foundry workloads by the end of 2026. Those current and planned Maia 200 uses do not establish that Maia 300 is serving production traffic.

Scott Guthrie, Microsoft’s executive vice president for Cloud + AI, said the company intends Maia to become the primary inference system for its products and services. That goal describes a planned shift in how Microsoft serves AI requests. The company has not specified how quickly Maia would take over those workloads or how many Maia 300 chips it will install in 2027.

The multi-gigawatt target also applies to the Maia family, rather than an announced quantity of Maia 300 capacity already operating. Microsoft has not said when capacity at that scale will become operational.

Rack design and performance claims

Microsoft specifies 96 accelerators, 27.6 TB of high-bandwidth memory and a 167-kilowatt power envelope for each Maia 300 rack. Its architecture description specifies eight HBM4 memory stacks and 288 GB of memory per accelerator. Microsoft says the accelerator’s compute chiplets use TSMC’s N2P process. These are design specifications, not measurements from deployed Maia 300 systems.

The 167-kilowatt rack specification makes power and cooling capacity material requirements for a large deployment. It does not show how much datacenter capacity Microsoft has secured for Maia 300 or how many racks it can bring online next year.

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems and up to seven times Maia 200’s rack-scale token throughput. These are company comparisons, not independently verified production results. Reuters reported that Microsoft did not name the GPU systems in the efficiency comparison. Production performance and its effect on total operating cost remain unproven.

Implications for Microsoft’s chip purchases

Making Maia the primary system for Microsoft’s inference workloads would give the company greater control over the hardware and software used to serve them. It could also reduce reliance on purchased GPU systems. Neither outcome establishes that Microsoft has reduced its accelerator purchases or operating costs.

GeekWire reported that Nvidia and AMD remain part of Microsoft’s infrastructure plans. The amount of third-party accelerator capacity Microsoft will continue to buy remains uncertain. The practical scale of the Maia shift will depend on installations and production workloads that have yet to be reported.