Microsoft plans Maia 300 rollout for 2027, claims efficiency gains

Microsoft announced Maia 300 on October 12 and plans production deployments in 2027. It expects the system to deliver 40% more tokens per megawatt than leading GPU-based systems. Tokens are the pieces of text an AI model processes or generates. Microsoft also expects up to seven times Maia 200's token throughput in a full rack.

A Maia 300 rack combines 96 accelerators within a 167-kilowatt power envelope. Microsoft says Maia 200 already runs production workloads, including Copilot. GitHub Copilot and Microsoft Foundry are due to follow by year-end. Maia 300's design separates computing, memory control, and input-output work into different chiplets.

Microsoft projects a 40% efficiency gain and seven times the rack throughput. Those figures measure different things. One measures power efficiency, while the other measures total rack output. Microsoft has not published independent results or promised lower customer prices. Maia 300 is due in 2027, while Maia 200 runs production workloads today.

AI systems keep using power every time they answer a request, even after training ends. A cloud company that builds its own chips can tailor the hardware to its services. More output per megawatt could help at sites where limited power restricts how much computing they can run. Microsoft still needs to show that the gain holds for the models and response times customers need.

Microsoft's 2027 production results need to show which models run on Maia 300 and how much energy each useful answer takes. Customer prices will show whether Microsoft passes any savings on to users. A rollout date answers neither question.