Microsoft unveils Maia 300 inference chip, with production deployments from 2027
Microsoft unveiled Maia 300, the third generation of its Azure Maia inference chips, pairing eight HBM4 stacks with TSMC N2P compute chiplets in a 950-watt package. It expects 40% more tokens per megawatt than leading GPU-based systems, Microsoft's own figure for a chip not yet deployed. Production deployments start in 2027; Maia 200 already serves Microsoft Copilot.
What it means for youif you plan AI capacity on Azure, treat the 40% as Microsoft's expectation until Maia 300 is deployed, and watch for more detail at Ignite next month.
With Nora, AI hardware analyst. Hosted by Vivian.
Sources
- Microsoft: Microsoft to expand Azure Maia system to multi-gigawatt deployments blogs.microsoft.com, Oct 12, 2026
Read the transcript
VivianMicrosoft unveiled Maia 300, its third-generation chip for running AI models, and says production deployments start in 2027.
VivianNora, our AI hardware analyst. What's inside it?
NoraMaia 300 is built for inference, the work of running models once they're trained. Microsoft says moving data is increasingly the bottleneck in large inference systems. So each chip pairs 8 HBM4 memory stacks, with 18 TB/s of peak memory bandwidth, inside a 950 W package. The compute chiplets are built on TSMC's N2P process.
VivianWhy does memory matter so much for running a model?
NoraMicrosoft designed it for long-context, memory-heavy work, like long conversations and multi-step agent tasks. It also supports low-precision formats, including an enhanced MXFP4, which cut the memory and bandwidth each step needs. Microsoft says that allows larger context windows and higher throughput in the same power envelope.
VivianMicrosoft's headline is 40% more tokens per megawatt than leading GPU-based systems. How should we read that?
NoraAs an expectation. It's Microsoft's own figure, for a chip that isn't deployed yet, and the post doesn't name the GPU systems it compares against. At rack scale, Microsoft puts 96 accelerators in 167 kW, and says that delivers up to 7x the token throughput of a Maia 200 rack.
VivianAnd Maia 200 is already running?
NoraYes. Microsoft says Maia 200 serves production work, including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. Microsoft plans to share more at its Ignite conference next month, and Maia 300 production deployments start in 2027. Microsoft wants Maia to be the main inference system for its own products and services.
VivianSo, for you: if you plan AI capacity on Azure, treat the 40% as Microsoft's expectation until Maia 300 is deployed, and watch for more detail at Ignite next month.
VivianDailyChat is AI-generated. Vivian and Nora are AI characters, not real experts. Facts come from the sources we link. Not professional advice. See you tomorrow.