Microsoft details Maia 300 AI chip rollout

Microsoft introduced its third-generation Maia 300 AI accelerator on October 12, 2026, outlining plans to expand its role in handling artificial intelligence workloads. The company said the Maia family will become the primary inference system for its products and services as deployments scale beginning in 2027. This marks a shift toward greater reliance on internally designed silicon after years of depending heavily on third-party graphics processing units.

The rollout is described as a multi-gigawatt expansion of Maia capacity, with plans discussed that include potential deployments exceeding one million Maia 300 chips in 2027. This planned scale is intended to support growing demand for AI model inference across Microsoft’s services.

Microsoft highlighted several performance and efficiency claims for the Maia 300 relative to existing GPU-based infrastructure. The company said Maia 300 delivers about 40% more tokens, or useful compute, per megawatt than leading GPU-based systems used for similar workloads. It also said Maia 300 offers up to seven times the rack-scale token throughput of the prior-generation Maia 200.

In terms of physical configuration, a Maia 300 rack is designed to hold 96 accelerators and draws roughly 167 kilowatts of power. These specifications are intended to improve throughput per rack while managing power consumption at the data center level. Some coverage has noted variation in published performance figures across different briefings, reflecting differences in how rack- and watt-level metrics have been framed, but Microsoft’s stated headline gains focus on token throughput and energy efficiency.

While expanding its in-house Maia program, Microsoft said Nvidia and AMD hardware will remain important in its infrastructure mix. The company plans to continue relying on Nvidia and AMD chips for model training workloads, even as Maia becomes the primary system for inference. This division reflects different requirements for training large models versus serving them in production.

Microsoft also indicated that Maia accelerators will be reserved for its own internal use. The chips are not being made available as rentable cloud instances to external customers. Instead, Maia capacity is intended to underpin Microsoft’s own AI-driven products and services directly.

Coverage of the announcement included estimates from some analysts about the financial impact of the Maia program. One assessment projected that Maia chips could account for roughly 10% of Microsoft’s AI chip spending by the first quarter of 2027. This suggests that even at large deployment scale, internally designed silicon would complement rather than fully replace external GPU purchases in the near term.

Taken together, the Maia 300 launch, the planned multi-gigawatt rollout starting in 2027, and the continued partnership with Nvidia and AMD outline a hybrid AI hardware strategy. Microsoft is moving to reduce its dependence on third-party GPUs for inference while maintaining a diverse supplier base for training workloads and overall capacity.