Photo: Derrick Coetzee (Public domain), via Wikimedia Commons (cropped) (source)
Microsoft plans to scale its in-house Azure Maia AI chips to deployments measured in gigawatts starting in 2027, with a new third-generation part, Maia 300, becoming the main system it uses to run AI models for its own products and services, Scott Guthrie, the company's executive vice president of Cloud + AI, wrote in a Microsoft blog post on Monday, Oct. 12.
The chip was unveiled at the OCP Global Summit in San Jose, Yahoo Finance reported. Maia 300 is built for inference, the computing used to generate answers from trained models, rather than for training them.
What Microsoft is claiming
Guthrie wrote that the company expects the chip "to deliver 40% more tokens per megawatt compared to leading GPU-based systems." Tokens are the units of text and other data that AI models read and produce. Configured as a full rack of 96 accelerators, he wrote, it reaches "up to 7x the rack-scale token throughput of Maia 200," the previous generation.
According to the company, each Maia 300 uses eight HBM4 memory stacks for 18.0 terabytes per second of peak bandwidth within a 950-watt power rating. Its compute chiplets are made on TSMC's N2P process. A full rack draws 167 kilowatts and holds 27.6 terabytes of high-bandwidth memory.
Maia 200, which The Verge noted was announced only in January, is already serving production work including Microsoft Copilot and is deployed in the United States, Europe and Australia, Microsoft said. The company said it delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork. DatacenterDynamics reported that Maia 200 is built on TSMC's 3-nanometer technology.
Microsoft AI chief executive Mustafa Suleyman told The Verge: "At that scale we need a platform that we own, and that we co-design with our models."
Nvidia's customers build their own
Axios framed the launch as evidence that every major Nvidia customer is now also a competitor. Amazon, Google and Meta, along with AI labs OpenAI and Anthropic, already offer or are developing custom chips, according to Yahoo Finance, which said Nvidia regularly earns upwards of half its quarterly revenue from the largest cloud companies.
StorageReview reported that Nvidia and AMD hardware remains in Microsoft's training fleet and that Maia is not offered as an instance Azure customers can rent. The outlet also noted that the 40% claim does not name the GPU, model or serving setup it was measured against, so it cannot be checked outside Microsoft.
What's next
Maia 300 production deployments are scheduled to begin in 2027, Guthrie wrote. He said Microsoft will share more about its infrastructure strategy at Microsoft Ignite 2026 next month in San Francisco, and that GitHub Copilot and Microsoft Foundry are slated to move onto Maia 200 by the end of this year.
Sources
- Microsoft: Microsoft to expand Azure Maia system to multi-gigawatt deployments (2026-10-12)
- The Verge: Microsoft’s new AI chip makes it less reliant on Nvidia’s GPUs (2026-10-12)
- Yahoo Finance: Microsoft reveals Maia 300 AI chip as it works to reduce dependence on Nvidia, AMD (2026-10-12)
- DatacenterDynamics: Microsoft unveils Azure Maia 300, says chip will be its "primary inference system" (2026-10-12)
- StorageReview: Microsoft Maia 300 Puts 288GB of HBM4 at 18TB/s on an N2P Chiplet and 96 Accelerators in a 167kW Rack (2026-10-12)
- Axios: Every major Nvidia customer is now a competitor as Microsoft reveals new Maia chip (2026-10-13)
Motion Media News corrects errors promptly. To report one, email hello@motionnews.studio.
