Microsoft says its new Maia 300 chip will be the main engine behind its own AI starting in 2027

Microsoft unveiled Maia 300 on Monday, Oct. 12, 2026, the third generation of its custom AI accelerator, and said it will scale the Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia "the primary inference system for Microsoft products and services." Reuters reports the company plans to deploy potentially more than a million of the chips next year.

Training gets the headlines, but inference, the part where the model actually answers you, is the bill that never stops coming. Every Copilot reply, every agent that grinds through a task for an hour, is running on somebody's chips, and until now that somebody was mostly Nvidia. With Maia 300 Microsoft is saying out loud that from 2027 its own silicon becomes the default for its own products, and if Reuters' "more than a million chips" holds, that is a lot of work that no longer needs Nvidia or AMD hardware. The fine print is that this is all Microsoft's own math: "40% more tokens per megawatt" against an unnamed GPU, on a chip that ships next year into a market that moves every six months. But the direction is unmistakable. Google has TPUs, Amazon has Trainium, OpenAI is building its own, and now Microsoft wants the meter running on hardware it owns. Power, not chips, is the real ceiling now, which is why every number in this launch is per watt.

Microsoft announced Maia 300, the third generation of its Azure Maia custom AI chip program, on Monday, Oct. 12, 2026, in a post by Scott Guthrie, executive vice president of Cloud + AI. A custom AI chip is one a company designs for its own computers instead of buying a general-purpose graphics card. Microsoft says it will scale the Azure Maia family to “multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.” Inference is the work of running a finished model so it can answer you. Production deployments of Maia 300 begin in 2027. Those lines are Microsoft’s.

Reuters reports Microsoft plans to deploy potentially more than a million Maia 300 chips next year to back its internal AI services. That count is Reuters’. It is not a figure printed on Microsoft’s own launch post.

Microsoft’s performance claims: about 40% more tokens per megawatt than “leading GPU-based systems,” and, in a full 96-accelerator configuration, up to 7 times the rack-scale token throughput of Maia 200. A token is a small chunk of text the model reads or writes. A megawatt is a million watts of power. Throughput, here, is how many of those tokens a full rack can produce. Reuters notes Microsoft did not name the GPU systems it compared against. Those comparisons are Microsoft’s.

Specs Microsoft published: an inference-first design for long-context, memory-heavy and agentic workloads. Long-context means the model can keep a long conversation or document in mind. Agentic work is when the model takes a chain of steps on someone’s behalf. The chip uses a multi-chiplet layout that separates compute, memory-controller and I/O functions. A chiplet is a smaller piece of silicon that is later joined to others. Compute chiplets sit on TSMC’s N2P process, a 2-nanometer-class manufacturing step. Reuters reports the other chiplets use cheaper 3-nanometer technology. Eight HBM4 memory stacks give 18.0 TB/s of peak memory bandwidth inside a 950-watt TDP. HBM4 is a fast stacked memory sitting next to the chip. TDP is the heat and power the package is designed to handle. The chip supports low-precision formats including enhanced MXFP4, which store numbers in fewer bits so more of the model fits in memory. Networking uses an Ethernet-based custom AI transport to tie many chips together. Those lines are Microsoft’s, except the 3-nanometer line, which is Reuters’.

Each Maia 300 rack holds 96 accelerators in a 167-kilowatt power envelope, with 27.6 TB of HBM capacity and 1.7 PB/s of aggregate memory bandwidth. A rack is a tall cabinet of these chips. Reuters compares that with about 142 kilowatts for Nvidia’s latest GB300 racks with 72 GPUs. Those rack figures are Microsoft’s. The Nvidia comparison is Reuters’.

The predecessor, Maia 200, announced in January, is now serving Microsoft production workloads including Microsoft Copilot and supporting the MAI team’s next-generation model work, with GitHub Copilot and Microsoft Foundry to follow by the end of the year. Microsoft says Maia 200 runs OpenAI models, its own MAI models and leading open-source models, is deploying across the US, 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. Those lines are Microsoft’s.

Microsoft says the two generations will support GitHub’s agentic coding, model serving in Microsoft Foundry, inference and data generation for the MAI team, and latency-sensitive work such as Dragon Copilot for clinicians. Latency-sensitive means the wait has to stay short. Those uses are Microsoft’s.

Rani Borkar, president of Azure Hardware Systems and Infrastructure, told Reuters: “It’s basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system.” Reuters notes Microsoft expects to spend more than $175 billion in capital expenditures this calendar year. Capital expenditures are the money spent on buildings, chips and other long-lived kit. Microsoft says it will share more at Microsoft Ignite next month. The quotation and the spending figure are Reuters’. The Ignite line is Microsoft’s.

The picture is Microsoft’s Azure Maia 300 prototype AI chip, its package marked Microsoft Azure Maia 300, photographed on a dark blue stage. It is the company’s launch photograph. The frame does not print a calendar date.

In plain terms, Microsoft said on Monday that Maia 300 is its third custom AI chip and that, from 2027, Maia becomes the main system that runs its own products. Reuters says the company may put more than a million of the chips into service next year. Microsoft claims about 40% more tokens per megawatt than unnamed leading GPU systems, and up to 7 times the rack-scale throughput of Maia 200. Each rack holds 96 chips in 167 kilowatts, which Reuters sets against about 142 kilowatts for Nvidia’s latest 72-GPU GB300 racks. The older Maia 200 is already in production, including on Copilot.