Maia 200 is serving; Maia 300 is not

On October 12, Scott Guthrie, Microsoft's executive vice president for cloud and AI, wrote that Azure Maia 200 is already serving Microsoft production workloads, including Microsoft Copilot, and supporting the MAI team as it develops models. GitHub Copilot and Microsoft Foundry are planned to follow by the end of the year. Across these workloads, Maia 200 runs OpenAI models, Microsoft's own MAI models, and leading open-source model families.

Maia 200 is deploying across the United States, Europe, and Australia. For built-in AI in Teams, Copilot, and Cowork, Microsoft says it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure. The figure is an upper value, for those built-in capabilities, not an average across every workload.

[1]
One rectangular charcoal chip, with a blue-gray edge and a small rust-colored mark at a corner.
One rectangular charcoal chip, with a blue-gray edge and a small rust-colored mark at a corner., AI-generated illustration, not a news photograph

Multi-gigawatt scale is a 2027 plan

Guthrie writes that Microsoft will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, and make Maia the primary inference system for Microsoft products and services. That is a plan, not capacity already in place.

Maia 300 production deployments begin in 2027. Microsoft expects Maia 300 to deliver 40 percent more tokens per megawatt than leading GPU-based systems, from efficiency across the whole system. In a full 96-accelerator configuration, it delivers up to 7 times the rack-scale token throughput of Maia 200. The 40 percent is an expectation. The 7 times is an upper value, and the comparison is Maia 200, not an outside GPU.

[1]

The Maia 300 specifications are on the page

The post describes Maia 300 as an inference system for long-context, memory-intensive, decode-heavy work. Eight HBM4 stacks provide 18.0 TB/s of peak memory bandwidth within a 950-watt TDP. Compute chiplets use TSMC's N2P process. Low-precision formats include enhanced MXFP4, to cut memory and bandwidth for computation and the KV cache. The post does not give a context length that results from that.

A Maia 300 rack puts 96 accelerators inside a 167-kilowatt power envelope, with 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth. Microsoft points to a separate technical paper for architectural detail. This blog does not reproduce that paper's measurements.

[1]