Microsoft has announced a scale-up of its in-house Maia accelerators starting in 2027. Maia 200 already serves production workloads including Microsoft Copilot, with GitHub Copilot and Foundry expected before year end.
Maia 300 targets long context, decoding and sustained agent tasks. Microsoft schedules production deployments for 2027 and describes 96-accelerator racks. It expects 40% more tokens per megawatt than its GPU comparison systems and reports up to seven times Maia 200’s rack-scale token throughput.
Maia 300 is not a W42 production release. Energy and throughput comparisons are Microsoft’s, with different scopes; they must not be added or generalized to all models. This is also distinct from MAI-Base-2 pretraining.
Cloud providers seek end-to-end inference optimization rather than relying solely on the GPU market.