Microsoft says its next Azure Maia chip should deliver 40% more tokens per megawatt than leading GPU systems, and it wants that system carrying inference for its own products from 2027. Scott Guthrie, executive vice president for the cloud group, says Microsoft will scale the Maia family to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for Microsoft products and services (Microsoft's blog). Production of Maia 300 starts the same year.
In a full rack of 96 accelerators, Microsoft expects up to seven times the token throughput of Maia 200. You still can't rent it by the hour, so there is no customer price to set beside a GPU. Model training, including work tied to OpenAI and Anthropic, still runs with Nvidia and AMD (CNBC).
Copilot now, the ramp later
Maia 200 already runs Microsoft Copilot. GitHub Copilot and Microsoft Foundry follow on Maia 200 by the end of the year. The 2027 ramp date is soft. Maia 300 production is planned for that year, and the multi-gigawatt build isn't locked, so Microsoft can match the silicon to the models it is serving.
On that chip Microsoft also runs OpenAI models, its own MAI models, and leading open-source families, and the same hardware supports the team building the MAI models. Deployments are underway in the United States, Europe, and Australia. On Teams, Copilot, and Cowork, Microsoft says Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure.
This is the third Maia accelerator. Microsoft introduced the first in 2023. Maia 200 is the generation in production now, and Nvidia's GPUs still fill the data centers Microsoft already runs. For general-purpose computing, Microsoft uses its own CPUs plus Intel and AMD. Maia is a different job. It's the inference bet.
What the company learned running Maia 200 is what it says shaped Maia 300. The new chip is aimed at long-context reasoning, latency-sensitive conversation, and multi-step agent tasks, including clinical workflows in Dragon Copilot. Microsoft says the bottleneck is moving data, not raw compute, so Maia 300 keeps more of that work on the chip. Silicon takes years, so the hardware and software teams start together.
If you pay for Azure AI or Copilot seats, treat the 2027 Maia plan as Microsoft's cost and latency story. Don't rebuild your stack this quarter. Watch Ignite for which products actually move onto it.
Photo: Jiaqian AirplaneFan / Wikimedia Commons, CC BY 3.0.
