Microsoft makes its own Maia chips its main AI inference system from 2027, and unveils Maia 300

Microsoft will make its own Azure Maia chips the main system that runs AI for its products and services, starting in 2027, and it announced the next generation, Maia 300, today. Scott Guthrie, Microsoft's executive vice president for Cloud + AI, published the plan on the Official Microsoft Blog on October 12, and Satya Nadella called it "our biggest bet yet on our own silicon".

WHAT CHANGES

- From 2027, Microsoft will scale Maia to multi-gigawatt deployments and make it the primary inference system for its own products and services. - Maia 200 already serves production work, including Microsoft Copilot, and supports Microsoft's MAI model team. GitHub Copilot and Microsoft Foundry are due to follow by the end of this year. - On those workloads Maia 200 runs OpenAI models, Microsoft's own MAI models and open-source model families, per the post.

MAIA 300, AS MICROSOFT DESCRIBES IT

- Eight HBM4 memory stacks, 18.0 TB/s of peak memory bandwidth, 950 W per chip. - Compute chiplets built on TSMC's N2P process. - Support for low-precision formats, including an enhanced MXFP4, which Microsoft says shrinks memory and KV-cache needs for long contexts. - A rack holds 96 accelerators in 167 kW, with 27.6 TB of HBM and 1.7 PB/s of total memory bandwidth. - Production deployments start in 2027.

MICROSOFT'S OWN FIGURES

These are Microsoft's claims, not independent measurements:

- Maia 300 is expected to deliver 40% more tokens per megawatt than "leading GPU-based systems". - A full 96-accelerator Maia 300 rack delivers up to 7x the rack-scale token throughput of Maia 200. - Maia 200 delivers up to 1.7x more tokens per watt than leading GPU infrastructure for the AI features in Teams, Copilot and Cowork.

The post does not name Nvidia or say which GPU systems it compared against.

WHAT WE DO NOT KNOW

- Whether Azure customers will be able to rent Maia 300 directly, or whether it will lower what they pay per token. - Which GPU systems and models the 40% figure was measured on. No third-party benchmark exists yet. - How much of the OpenAI model traffic Microsoft serves will move to Maia, beyond the workloads the post names.

Microsoft says it will share more at Ignite 2026 in San Francisco next month.