Microsoft Plans Multi-Gigawatt Maia Rollout, Led by N2P-Based Maia 300

Microsoft said on Monday that it will scale its in-house Maia chips to multi-gigawatt deployments beginning in 2027, making the company's own silicon the primary inference system for its products and services.

"We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services," Scott Guthrie, executive vice president of Cloud + AI, wrote in a post on the company's official blog. The vehicle for that expansion is Maia 300, the third generation of the chip, which Microsoft said will begin production deployments in 2027.

The phrase matters because of what it measures. Microsoft did not give a chip count or a dollar figure. It described the program in power, the unit that increasingly sets the pace of AI build-outs, and it framed Maia 300's main selling point the same way: the company said it expects the chip to deliver 40% more tokens per megawatt than "leading GPU-based systems." That figure is Microsoft's own projection for hardware not yet in production, and the company did not name the GPU systems it used as the baseline.

Maia 300 also changes how the chip is built. Maia 200, introduced in January, is made on TSMC's 3-nanometer process. Maia 300 splits the work across chiplets, separating compute, memory-control and input/output functions. Only the compute chiplets use TSMC's N2P process. Microsoft said this places compute where the advanced process delivers the greatest gain in performance per watt, while the memory-controller and I/O functions sit on dies optimized for their roles.

TSMC describes N2P as an extension of its 2-nanometer family with "further performance and power benefits on top of N2," and lists its volume production as scheduled for the second half of 2026. That timing fits Microsoft's 2027 schedule. Microsoft did not say how many wafers it has booked or where the chips will be packaged.

The figures Microsoft did publish are mostly about memory, which is where inference workloads increasingly run into limits. Each Maia 300 accelerator carries eight stacks of HBM4 memory with 288 gigabytes of capacity and 18 terabytes a second of peak bandwidth, within a 950-watt thermal design power. Maia 200 had 216 gigabytes of HBM3e at 7 terabytes a second within 750 watts. Microsoft's Azure engineering blog puts the generational change at 2.5 times the memory bandwidth, 33% more capacity and roughly twice the bandwidth per watt.

Compute grows more slowly. Microsoft said the chiplet design raises the system's compute FLOPs by 36% and scale-up bandwidth by 14%. It did not give a headline petaflop figure for Maia 300, unlike in January, when it said each Maia 200 delivered more than 10 petaflops at 4-bit precision. The uneven gains are deliberate. Microsoft describes Maia 300 as "inference-first," built for long-context reasoning, conversational responses that need low latency, and agentic tasks that run for many steps. Those jobs tend to be held back by moving data in and out of memory, not by raw arithmetic. Lower-precision formats, including what Microsoft calls enhanced MXFP4 and storage formats below 3 bits, are meant to cut the bytes moved for each token.

The rack is where the numbers compound. A Maia 300 rack holds 96 accelerators in 24 trays of four, double the density of Maia 200, inside a 167-kilowatt power envelope. Microsoft said the rack provides 27.6 terabytes of HBM and 1.7 petabytes a second of aggregate memory bandwidth. Each tray carries its own host processor, so there are no separate host trays. The rack also has six switch chips in three liquid-cooled trays, dual power feeds and a 50-volt busbar. An Ethernet-based two-tier network connects as many as 12,288 accelerators into what Microsoft calls a superpod. Against Maia 200, the company claims up to seven times the rack-scale token throughput and more than 2.5 times the performance per watt.

Those multiples are Microsoft's figures, and the company has not published the workloads or settings behind them. The seven-times throughput claim compares a full 96-accelerator rack with an earlier rack holding half as many chips, so much of the gain comes from packing in more silicon and power, not from each chip doing more. The per-watt claim is the better guide to whether more megawatts will produce proportionally more output.

Maia 200 is already doing paid work. Microsoft said it is serving production workloads including Microsoft Copilot and supporting the company's MAI superintelligence team, with GitHub Copilot and Microsoft Foundry due to follow by the end of the year. According to the company, the chip is being deployed across the United States, Europe and Australia, and delivers up to 1.7 times more tokens per watt than "leading GPU infrastructure" for built-in AI features in Teams, Copilot and Cowork.

Microsoft gave no sites, regions or dates for Maia 300 beyond the start of production deployments in 2027. It also did not say how much of the multi-gigawatt plan will run on Maia 300 rather than Maia 200, or how quickly the capacity will come online. Those gaps make the announcement a statement of direction rather than a build schedule.

Maia 200 has already featured in Microsoft's investor messaging. On the July 29 earnings call, chief executive Satya Nadella said: "Maia 200 continues to scale. It delivers 30% better performance per dollar than the latest-generation hardware in our fleet and is now supporting both OpenAI and MAI models." He added that Microsoft is "seeing 40% better performance per watt when running MAI models on Maia 200." Maia is not named in the company's 10-K, which refers only to "custom-built silicon."

The chips sit inside the fastest-growing part of Microsoft's spending. Capital expenditures were $41bn in the fiscal fourth quarter to June 30, and chief financial officer Amy Hood said "roughly two thirds of our capex was for short-lived assets, primarily CPUs and GPUs." For the quarter to September she guided to spending of "over $50 billion." The 10-K lists $329.1bn of leases not yet commenced, primarily for data centers, starting between fiscal 2027 and 2033. Microsoft has not said how much of its spending goes to Maia.

Investors get the next read on Oct. 28, when Microsoft reports fiscal first-quarter results after the market closes. Actual spending against the "over $50 billion" guide, and any update on Maia's role, will show how much of the 2027 plan is already in the budget.

For the supply chain, the commitment matters in two directions. TSMC gains a large hyperscale customer for N2P compute chiplets, along with the advanced packaging needed to join them to memory and I/O dies. Memory makers gain a buyer for eight HBM4 stacks per accelerator. For GPU suppliers, the message is narrower than it first sounds. Microsoft's own wording limits Maia's primary role to inference for Microsoft's own products and services. It says nothing about training frontier models, and nothing about the GPU capacity Azure rents to outside customers.

Microsoft's case rests on co-design: tuning the chip, rack, network and software together for workloads it already runs at scale. Its first-generation chip, Maia 100, was unveiled in 2023. Maia 200 has moved into production. Maia 300 is the first generation that Microsoft has tied to a power commitment large enough to change how it plans data centers. Whether the 2027 date holds will depend on N2P yields, HBM4 supply and the company's ability to get grid power at gigawatt scale, and Microsoft did not address any of those risks on Monday.