Microsoft Bets Big on Maia 300, Targeting Multi-Gigawatt Inference From 2027

Microsoft today announced Maia 300, its third-generation in-house chip, built primarily for inference, and Satya Nadella says the company will scale the Maia family to multi-gigawatt deployments starting next calendar year. The goal, in his words: "making it the primary inference system for our own product and services."

Gigawatts, ladies and gentlemen. Not megawatts, not racks, gigawatts, plural, with a "multi" in front. As we previously reported, the chip is designed to be the "primary AI inference system" for Microsoft's products and services, and it makes the company less reliant on Nvidia's GPUs.

The headline numbers

Nadella says Maia 300 delivers 40% more tokens per watt compared to leading GPU-based systems. Scott Guthrie, executive vice president of Microsoft's cloud group, put it this way: "We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems." Guthrie expresses the efficiency claim per megawatt rather than per watt; both executives cite a 40 percent improvement.

Guthrie adds that in a full 96-accelerator configuration, Maia 300 "delivers up to 7x the rack-scale token throughput of Maia 200." That is a generational leap, by Microsoft's own account.

The hardware specs, per TechRadar's reporting:

• Eight stacks of HBM4 with a claimed peak bandwidth of 18TB/s within a 950W thermal design power.
• Compute chiplets built on TSMC's N2P process, in a multi-chiplet design that separates compute, memory-controller and input/output functions.
• Support for low-precision formats, including an enhanced MXFP4, which Microsoft says will reduce memory and bandwidth demands.
• A full rack of 96 accelerators in a 167kW power envelope, with 27.6TB of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth.

Microsoft also says specialized software can use on-chip SRAM as a scratchpad, easing pressure on HBM during demanding workloads. For networking, Maia 300 will use high-bandwidth connections between accelerators for tensor parallelism, plus a custom AI transport based on Ethernet infrastructure.

The 2027 timeline

Nadella's "next calendar year" means 2027, matching Guthrie's timetable for scaling the Azure Maia family "starting in 2027." TechRadar reports that Maia 300 itself will begin deploying in 2027.

What carries the load before then is Maia 200. Nadella says Microsoft is already deploying it across the US, Europe and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot. TechRadar reports that Microsoft says Maia 200 is already running production workloads including Microsoft 365 Copilot and models developed by its Microsoft AI team. GitHub Copilot and Microsoft Foundry deployments are expected to follow by the end of this year.

What it means

This is a power move, and I mean that in every sense. Microsoft is saying its own Maia family will become the primary inference engine for its products and services, reducing its reliance on Nvidia GPUs. The announcement does not say Nvidia is out. It is about inference, the work of actually serving answers to users, and it is about Microsoft's own products and services first.

The efficiency claims are the commercial heart of it. Tokens per watt is the currency Microsoft is spending in this pitch, and a 40 percent edge over "leading GPU-based systems" is the kind of number that moves capital plans. Microsoft's comparison, as presented, does not name which GPU systems it means.

What to watch

The announcement leaves several big questions open:

• Whether Maia 300 will be offered to external Azure customers or stay inside Microsoft's own products.
• What share of Microsoft's overall inference capacity Maia will carry compared with Nvidia and AMD hardware.
• Where the multi-gigawatt Maia data centers will be built.
• How Maia 300 stacks up in benchmarks against specific Nvidia GPUs such as Blackwell or Hopper. None are in the announcement.

The plan is to start in 2027 and scale from there. Between now and then, watch the Maia 200 rollout to GitHub Copilot and Foundry, the next expansion of a strategy already running in production. The gigawatts are plural, the ambition is enormous, and the receipts come next year.