Microsoft said on 12 October 2026 that from 2027 the chips it designs itself will be the main system running the AI behind its own products, including its assistant Copilot. The chip that has to carry that plan is Maia 300, the third generation of the Maia family, and Microsoft's claim for it is efficiency: 40% more output from the same electricity than the leading systems built on graphics chips, the kind most AI runs on today. That figure is a company expectation, aimed at competitors Microsoft does not name, and the company has not said how many of the chips it intends to install.
Two jobs, and Maia does one of them
AI does two different jobs. One is building a model, when software learns patterns from very large amounts of text and other data. The other is running a finished model while people are using it, which the industry calls inference. Bloomberg's account of the announcement gives the examples of summarizing a call and answering a research question. Maia chips do that second job only; they do not build models.
Microsoft has been doing that work at scale with the previous generation. Maia 200 is running in production in data centers in the United States, Europe and Australia, serving Copilot, OpenAI's models and Microsoft's own MAI models, and Microsoft says it delivers up to 1.7 times more output per watt than leading GPU infrastructure for the AI features built into Teams, Copilot and Cowork. GitHub Copilot and Microsoft Foundry are due to run on Maia 200 by the end of 2026. Maia 300 was announced less than a year after Maia 200, and Andrew Wall, Microsoft's general manager for Maia, has said the company is shortening the time between finishing a chip design and running it in a data center.
What 40% more from the same electricity means
A token is a piece of text about as long as a short word, and a model writes its answers one token at a time. Tokens per megawatt therefore measures how much answering a system produces from a fixed amount of electricity, where a megawatt is a million watts. Microsoft expects Maia 300 to produce 40% more tokens per megawatt than what the company describes as leading GPU-based systems, a number the cloud chief Scott Guthrie put in its announcement.
Electricity is the measure because of what limits a data center. Wall described the trade-off in an interview with the Seoul Economic Daily: for years the highest-density racks with the most computing power were what mattered, but they require high power and do not necessarily fit existing data centers, which take years to build. The design goal for Maia, he said, was not the highest-power rack but the configuration that fits where it is installed and gets the most out of each watt.
The hardware follows that idea. A Maia 300 rack holds 96 chips in a unit that draws 167 kilowatts, with 27.6 terabytes of high-bandwidth memory and 1.7 petabytes per second of memory bandwidth across the rack, according to the technical account Microsoft published. High-bandwidth memory is stacked close to the processor, where it holds the model and the conversation in progress; without enough of it, a chip spends much of its time waiting for data instead of producing answers. Each Maia 300 chip carries eight of the newest HBM4 memory stacks inside a 950-watt power limit, and its computing parts are made by the chipmaker TSMC on a manufacturing process called N2P. Microsoft says a full rack delivers up to seven times the token throughput of a Maia 200 rack, and that up to 12,288 of the chips can be linked across 128 racks. All of those are company figures, and Microsoft has not described the mix of work behind the 40% comparison, the systems it compared against, or any test done outside the company.
The chips Microsoft buys stay in the fleet
The plan is not an exit from Nvidia. Wall said Microsoft expects to keep a heterogeneous fleet: "We expect to continue maintaining a heterogeneous fleet for Microsoft workloads. That means we will deploy Nvidia and AMD GPUs alongside scaling up Maia." He gave that description to the Seoul Economic Daily. In the same interview he separated two kinds of chip: general-purpose graphics chips, such as those from Nvidia and AMD, which suit any task, and specialized accelerators, such as those from Groq or Cerebras, which are built for a narrow range. Maia is meant to sit between the two, optimized for the work Microsoft cares about most, and the work he names first is the assistants people talk to.
On rack design he made a direct comparison, saying Microsoft's rack holds 96 chips at 167 kilowatts while Nvidia's rack design uses higher-power racks and holds 72 accelerators. That comparison is Microsoft's own, and the systems behind the 40% figure are still unnamed.
Who gets the first chips, and what Microsoft will not say
Wall said the systems first deployed in 2027 "will focus primarily on Copilot, GitHub Copilot, Microsoft AI and Azure Foundry". Foundry is the service outside companies use to build with Microsoft's AI tools, so it is a route by which customers who are not Microsoft would use Maia hardware. He said capacity is also slated for Seoul but declined to give timing. Asked about scale, he said Microsoft is "not disclosing exact chip counts" while describing deployments on the order of several gigawatts. A figure of more than a million chips next year appeared in secondary coverage of the launch; Microsoft has not confirmed it. The "biggest challenge across the industry", he said, is supply chains for processors and memory.
The argument is cost
The reason to build your own chip, in Wall's telling, is cost rather than capability. He said the performance of particular models is gradually reaching its limits. "In the end it is a question of intelligence per cost," he said, "so customers will begin to look at what they can do to optimize costs." Satya Nadella, Microsoft's chief executive, said in a post on X that the work is "our biggest bet yet on our own silicon". Mustafa Suleyman, who runs Microsoft AI, told The Verge that the company needs a platform it owns and co-designs with its own models. The spending behind that is large: Microsoft itself projects capital spending of about $175 billion for the calendar year, after changing how it accounts for data center leases.
The test begins in 2027
Maia 300 production deployments begin in 2027, and Wall said the company plans to expand through 2028. If the efficiency holds once the chips are carrying real traffic, each answer Microsoft's services produce costs less, and the company buys proportionally less of the hardware it currently relies on for inference. The 40% stays a company expectation until the chips carry real traffic, and the workload behind it, the comparison and the chip count are still Microsoft's to disclose.