Microsoft Unveiled Maia 300 on October 12, Saying It Will Make Maia Its Main Inference System From 2027
On October 12 Microsoft announced Maia 300, an inference chip built on TSMC's N2P process, and said it will scale the Maia family to multi-gigawatt deployments from 2027. Microsoft projects 40% more tokens per megawatt than leading GPU systems; that is the company's own estimate.
Microsoft says its in-house Maia chips will become "the primary inference system for Microsoft products and services." It announced the third-generation Maia 300 in a blog post on Monday, October 12, and said it will scale the Maia family to multi-gigawatt deployments starting in 2027.
What Microsoft claims
Microsoft says it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200 in a 96-accelerator configuration. Both figures are Microsoft's projections; Maia 300 begins production deployments in 2027, the company says.
The hardware
According to Microsoft, Maia 300 uses eight HBM4 stacks for 18.0 TB/s of peak memory bandwidth within a 950-watt TDP, with compute chiplets built on TSMC's N2P process. A rack holds 96 accelerators within 167 kilowatts and 27.6 TB of HBM, the company says.
SiliconANGLE reports that Maia 300's 288 GB of HBM matches Nvidia's Rubin accelerator but that it has about 20% less memory bandwidth than Rubin; that comparison is SiliconANGLE's own. It also reports that Microsoft plans superpods of 12,288 accelerators linked over a custom Ethernet-based network.
Where Maia 200 stands
Microsoft says Maia 200 already serves production workloads including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by the end of the year, and runs OpenAI models, Microsoft's MAI models and open-source models.
Limits
- Performance and efficiency numbers are Microsoft's own and have not been independently measured.
- The Rubin comparison is SiliconANGLE's analysis, not Microsoft's.