Microsoft Maia 300: 40% More Tokens per Megawatt, Ships 2027

Microsoft says its third-generation Azure Maia 300 inference accelerator will deliver 40% more tokens per megawatt than leading GPU-based systems and up to 7x the rack-scale token throughput of Maia 200, according to an October 12, 2026 post by Scott Guthrie, EVP of Cloud + AI. Production deployments start in 2027, and Microsoft plans to scale the Maia family to multi-gigawatt deployments, making it the primary inference system for Microsoft products and services.

The numbers on the wire

• 40% more tokens per megawatt versus leading GPU-based systems (Microsoft’s expectation; the GPU baseline is not named).
• Up to 7x rack-scale token throughput versus Maia 200, in a full 96-accelerator rack.
• More than 2.5x performance per watt versus Maia 200, per the Azure architecture deep dive.
• Rack: 96 accelerators in a 167-kilowatt envelope, 27.6 TB of HBM and 1.7 PB/s of aggregate HBM bandwidth.
• Per accelerator: eight HBM4 stacks, 288 GB and 18 TB/s, within a 950-watt TDP.
• Scale-up: superpods of up to 12,288 accelerators (128 racks) on an Ethernet-based custom AI transport.

What changed from Maia 200

Maia 300 moves to a multi-chiplet design that separates compute, memory-controller and I/O dies. Compute chiplets are built on TSMC N2P. Microsoft says the split adds 36% more compute FLOPs, 2.5x the HBM bandwidth, 33% more memory capacity, about 2x the HBM bandwidth per watt and 14% more scale-up bandwidth. The rack doubles from 48 to 96 accelerators.

The design targets long-context, decode-heavy and agentic inference. Microsoft added an enhanced MXFP4 format and sub-3-bit storage formats to shrink memory and KV-cache traffic, and moved memory to a non-uniform layout across chiplets so data sits near the compute that uses it. In-rack links are electrical; optics carry the second network tier across racks.

Where Maia 200 stands

Maia 200 is now serving Microsoft Copilot and supporting the MAI model team, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. It runs OpenAI models, Microsoft’s MAI models and open-source model families, and is deploying across the United States, Europe and Australia. Microsoft says it delivers up to 1.7x more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork.

On the record

“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.” — Scott Guthrie, EVP, Cloud + AI, Microsoft

What to watch

Microsoft gave no gigawatt figure, deployment timeline beyond 2027, or named GPU baseline for the 40% claim. The receipts to track: the first Maia 300 production region, whether Foundry customers can rent Maia capacity directly, and how much of Microsoft’s inference fleet shifts off third-party GPUs. Microsoft says more details come at Ignite 2026 next month in San Francisco. Microsoft’s EMEA newsroom carries the same headline figures.

FAQ

What is Microsoft Maia 300?

Microsoft’s third-generation in-house AI accelerator, designed for inference on Azure and built as a full system with silicon, racks, networking and software co-designed.

When will Maia 300 ship?

Microsoft says production deployments begin in 2027.

How fast is Maia 300 versus Maia 200?

Up to 7x the rack-scale token throughput in a 96-accelerator rack and more than 2.5x the performance per watt, according to Microsoft.