Microsoft said on 12 October that it will scale its Azure Maia AI chips to multi-gigawatt deployments from 2027. Maia will then be its main system for running trained models, known as inference. Maia 300, the third generation, enters production in 2027. Microsoft expects 40% more tokens per megawatt than leading GPU-based systems. The Verge reported this lets Microsoft rely mainly on its own chips instead of Nvidia and AMD.
Microsoft says moving data, not raw compute, now limits large inference systems. Maia 300 uses a dataflow design to cut data movement. Eight HBM4 memory stacks give 18.0 TB/s within 950 watts. A low-precision format, MXFP4, shrinks memory needs. Compute, memory control and I/O sit on separate small chips. A rack holds 96 accelerators at 167 kilowatts. The 40% figure is an expectation, as Maia 300 is not yet in production. The post does not name the GPU systems or describe the test setup. It links a technical paper that was not part of the reviewed material. Neither the post nor The Verge gives GPU purchase numbers or Maia pricing.
The contest is shifting from raw compute to output per megawatt. Buyers of Copilot, GitHub Copilot or Foundry should ask how Maia shows up in pricing, and how much of their service still runs on GPUs.