Microsoft has now published the specifications for Maia 300, its next in-house AI chip, and the headline number is a promise rather than a result. Scott Guthrie, the company’s executive vice president for cloud and AI, wrote on 12 October that Microsoft “expect[s]” Maia 300 to get 40% more tokens out of each megawatt than “leading GPU-based systems”. Production deployments begin in 2027, so nothing has been measured.
Our earlier coverage treated Maia 300 as an announced plan with no figures attached. The detail has arrived: a chip, a rack and a network, all described in numbers. The calendar has not moved.
Two generations are in play, and only one of them exists. Maia 200 is in production today. Microsoft says it serves Copilot, is deploying across the United States, Europe and Australia, and supports the MAI team, which builds Microsoft’s own models. GitHub Copilot and Microsoft Foundry, the company’s platform for building AI apps, are due to follow by the end of the year. The chips handle OpenAI models, open-source families and Microsoft’s own MAI models.
Maia 200 also carries the only efficiency figure in the package that describes hardware already running. Microsoft reports up to 1.7 times more tokens per watt for the built-in AI in Teams, Copilot and Cowork, against leading GPU infrastructure. That is still Microsoft’s number, on Microsoft’s workloads, and it is an “up to.”
Both comparisons lean on the same phrase, and Microsoft never defines it. The announcement and the architecture deep dive by Rani Borkar and Saurabh Dighe say “leading GPU-based systems” and stop there. No vendor, no chip generation, no test setup. A ratio against an unnamed baseline cannot be checked, and the baseline will have moved by the time Maia 300 ships. A 40% edge over today’s GPUs could shrink or vanish against what GPU makers sell in 2027.
Of the specifications, the memory figures say the most about what Microsoft is aiming at. Each Maia 300 accelerator pairs eight HBM4 memory stacks holding 288 GB with 18 TB/s of bandwidth, inside a 950-watt power budget. Against Maia 200 that is 2.5 times the bandwidth and 33% more capacity. Microsoft describes the chip as built for long-context, memory-hungry workloads where the model spends its time generating text, and moving data, not computing, is the bottleneck.
The 7x figure needs a careful read. Microsoft says a full rack of Maia 300 delivers up to seven times the token throughput of a Maia 200 rack. But the new rack holds 96 accelerators, double the density of the old one, so much of that jump is more chips rather than better ones. The like-for-like number is the one Microsoft gives for performance per watt: more than 2.5 times. Both compare the company’s own two products, which is a fair claim but a different one from beating a competitor.
The scale is the other signal. Racks draw 167 kilowatts, and Microsoft plans to link up to 128 of them into a 12,288-accelerator group. Guthrie says Maia deployments will reach multi-gigawatt scale from 2027, making it the main chip Microsoft uses to serve its own products. Satya Nadella called it “our biggest bet yet on our own silicon” in a post on X.
That is a statement about dependence. Microsoft is saying that most of the AI it serves to its own customers should eventually run on chips it designed, not chips it buys. Whether that saves money depends on the very comparison it has left unnamed.
Microsoft has said it will share more at Ignite in San Francisco in November. The first useful thing it could add is a named GPU, a named workload and a test someone else can rerun.
Based on Microsoft’s announcement by Scott Guthrie and its architecture deep dive by Rani Borkar and Saurabh Dighe, both published 12 October 2026.
