Microsoft Maia 300 explained: 2027 deployments and a multi-gigawatt plan

Maia 300 is Microsoft's next in-house AI accelerator, and it is built for inference, which means running trained models rather than training them. On Oct. 12, 2026, Microsoft said Maia 300 production deployments will begin in 2027. It also said it will scale the Azure Maia family to multi-gigawatt deployments starting the same year, which would make Maia "the primary inference system for Microsoft products and services." Today's chip, Maia 200, already serves Microsoft production workloads, including Microsoft Copilot. Every figure below comes from Microsoft's own posts. None has been independently benchmarked.

What Microsoft announced on Oct. 12, 2026

Microsoft made the announcement in two posts: a Microsoft Blog post by Scott Guthrie, Executive Vice President, Cloud + AI, and a technical Azure Blog deep dive by Rani Borkar, President of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, Corporate Vice President, Systems Planning and Architecture.

• Maia 200 now: serving Microsoft production workloads, including Microsoft Copilot, and deploying across the United States, Europe and Australia.
• Next on Maia 200: GitHub Copilot and Microsoft Foundry are "to follow by end of year." The post is dated October 2026 but does not state the year in that sentence.
• Models: Maia 200 runs OpenAI models, Microsoft's own MAI models and "leading open-source model families."
• Maia 300: production deployments to begin in 2027.
• Scale: multi-gigawatt Maia deployments starting in 2027.
• Next update: Microsoft says it will share more at Microsoft Ignite 2026 in San Francisco, listed for Nov. 17–20, 2026.

Maia 300 hardware, by the numbers

Per accelerator, according to Microsoft:

• Process: the compute chiplets are made on TSMC's N2P node. The design splits work across chiplets, and the deep dive says this raises compute FLOPs by 36%.
• Memory: eight HBM4 stacks with 288 GB of capacity and 18 TB/s of peak bandwidth.
• Power: a 950-watt TDP.
• Number formats: "enhanced MXFP4 and sub-3-bit storage formats" for low-precision inference.
• Memory layout: software-managed on-die SRAM, and memory that is now non-uniform (NUMA) across chiplets instead of uniform. Microsoft gives no SRAM capacity.

Per rack and beyond:

• Rack: 96 accelerators on 24 four-accelerator trays, within a 167-kilowatt rack. That is twice as many accelerators per rack as Maia 200.
• Rack memory: 27.6 TB of HBM and 1.7 PB/s of total HBM bandwidth.
• Networking: a custom AI transport over Ethernet. In-rack links are electrical. Second-tier switches connect up to 128 racks into a superpod of up to 12,288 accelerators, and any accelerator in another pod is at most three switch hops away.
• Cooling: liquid cooling removes more than 80% of the rack's heat. The rack can connect to facility water or use a heat-exchanger unit in air-cooled facilities.
• Power delivery: dual power feeds, high-capacity power shelves and a 50 V busbar.

How it compares with Maia 200

Microsoft's Jan. 26, 2026 Maia 200 post gives the baseline: TSMC 3-nanometer, over 140 billion transistors, over 10 petaFLOPS at FP4 and over 5 petaFLOPS at FP8, 216 GB of HBM3e at 7 TB/s, 272 MB of on-chip SRAM, a 750 W SoC TDP, and clusters of up to 6,144 accelerators.

The two posts are consistent with each other. 18 TB/s against 7 TB/s matches the deep dive's "2.5 times" memory bandwidth, and 288 GB against 216 GB matches its "33% more" capacity. The memory gain is much larger than the 36% compute gain, which suggests Microsoft is focused on the memory-bound part of serving large models.

How to read the efficiency claims

Microsoft gives four headline ratios. Each one measures something different, so they cannot be multiplied together:

1. Up to 7x rack-scale token throughput versus Maia 200. This applies only to a full 96-accelerator rack, and that rack has twice as many chips as a Maia 200 rack.
2. More than 2.5x performance per watt versus Maia 200. This is a generation-over-generation comparison.
3. 40% more tokens per megawatt than "leading GPU-based systems." Microsoft does not name the GPU systems it compared against.
4. Up to 1.7x more tokens per watt for Maia 200 versus leading GPU infrastructure. This covers built-in AI features in Microsoft Teams, Copilot and Cowork.

Chaining claim 2 and claim 4 does not give you Maia 300's lead over GPUs. Microsoft's own figure for that is claim 3, the 40%, and its baseline is unspecified. Neither post publishes full benchmark methods.

Why the 2027 plan matters

If Maia becomes the primary inference system for Microsoft products and services, most of Copilot-style traffic would move onto Microsoft's own silicon. Microsoft states the scale in gigawatts of power, and it states the efficiency claims in tokens per watt and per megawatt. That puts the focus on how much output the company gets from a fixed power budget. Maia 200 already runs OpenAI, MAI and open-source models, so the hardware is meant to serve models built outside Microsoft as well as its own.

What to watch

• Ignite 2026 (Nov. 17–20): Microsoft has promised more detail there.
• End of year: whether GitHub Copilot and Microsoft Foundry move onto Maia 200.
• 2027: whether Maia 300 production deployments and the multi-gigawatt scale-up start on schedule.
• Benchmarks: independent results, or the workloads and configurations behind each vendor ratio.

Limitations: this article relies on Microsoft's own announcements, and none of the performance figures has been independently tested.