Scott Guthrie, who leads Microsoft's cloud and AI business, said in a post on the company's official blog on October 12 that the Azure Maia family of in-house chips will expand to multi-gigawatt deployments starting in 2027, at which point Maia will become the primary inference system for Microsoft products and services. The same post unveiled the third-generation chip, Maia 300, which is slated for production deployment in 2027. CEO Satya Nadella reshared the announcement on social media the same day.
"making Maia the primary inference system for Microsoft products and services"
Maia 300 specifications
Microsoft's per-chip figures: eight HBM4 stacks, peak memory bandwidth of 18.0 TB/s, a thermal design power of 950 W, and compute chiplets built on TSMC's N2P process. The chip supports low-precision formats such as an enhanced version of MXFP4, which are meant to shrink memory and bandwidth footprints and support longer context windows.
A full rack holds 96 accelerators and draws 167 kW, with 27.6 TB of total HBM capacity and 1.7 PB/s of aggregate bandwidth. The interconnect between accelerators is tuned for tensor parallelism and runs a custom Ethernet-based transport protocol. On the architecture side, Microsoft treats on-chip SRAM as a software-managed scratchpad to cut down on repeated reads and writes to HBM. One line in the post sums up the thinking:
"Moving data efficiently, rather than raw compute power, is increasingly the bottleneck in large inference systems."
All of the performance figures come from Microsoft. Compared with leading GPU systems, it claims 40% more tokens per megawatt, and with all 96 accelerators in place, rack-level token throughput of up to 7 times that of Maia 200. Reuters noted in its coverage that Microsoft did not say which system it was comparing against.
What Maia 200 runs today
The previous generation, Maia 200, is already deployed in the US, Europe and Australia. It carries production traffic for Microsoft Copilot and is also used by Microsoft's own MAI model team. Microsoft says it delivers up to 1.7 times the tokens per watt of leading GPU facilities for the built-in AI features in Teams, Copilot and Cowork. GitHub Copilot and Foundry, Azure's model-serving platform, are scheduled to come on board by the end of the year. The models running on it include OpenAI's, Microsoft's in-house MAI models and major open-source models.
The three generations side by side
The Maia program has not had a smooth run. The first-generation Maia 100 debuted at Ignite in November 2023 and then saw limited internal use for years. Maia 200 launched in January this year, and when Nadella mentioned it in June, only two data centers were using it, with plans to expand to more locations, including overseas.
In August, The Information reported that Microsoft was negotiating Maia 300 capacity with TSMC, targeting delivery of more than 300,000 chips in 2027 and, over the longer term, more than one million. Microsoft did not confirm the numbers at the time; Andrew Wall, general manager of Azure Maia, said only that outside figures did not reflect the scale of the program and that the company expects demand served by Maia to be measured in gigawatts. Putting 'multi-gigawatt' in the official blog turns that remark into a written commitment, but the post still gives neither a chip count nor capital spending figures.
From 100 to 300, this is the first time the official wording has reached the level of 'primary inference system.' The qualifier is inference: the blog says nothing at all about which chips will be used for training. Nor does it say whether Azure cloud customers outside Copilot will be able to rent Maia compute.
Microsoft says it will share more on Maia strategy and technical details at Ignite 2026 in San Francisco in November. A technical white paper is already posted on the Azure website, but whether the numbers hold up in third-party testing will have to wait until the chips are actually in service.
Sources: Microsoft official blog, Azure Maia 300 technical white paper, Reuters, CocoLoop, Techzine, The Information; the Microsoft blog was used to check chip and rack specifications, the basis of the per-megawatt and per-watt token comparisons, and the deployment regions.