Microsoft Unveils Maia 300 and Plans Multi-Gigawatt Maia Deployments Starting in 2027

Microsoft on October 12, 2026 unveiled Maia 300, the third generation of its in-house AI accelerator, and said it will scale the Maia family to multi-gigawatt deployments starting in 2027. Maia is to become "the primary inference system for Microsoft products and services," Scott Guthrie, executive vice president of Cloud + AI, wrote in the announcement.

Inference is the work of running a trained model to answer requests, as opposed to training it. Nvidia GPUs fill Microsoft's existing data centers, CNBC notes, and the plan puts Microsoft alongside Google and Amazon, which have their own AI chips, and OpenAI, which built one with Broadcom (The Next Web). Microsoft introduced its first AI chip, Maia 100, in 2023 and announced Maia 200 on January 26, 2026, according to CNBC and Microsoft.

What did Microsoft announce?

Microsoft said on October 12, 2026 that Maia 300 will begin production deployments in 2027 and that the Maia family will reach multi-gigawatt scale from that year. Microsoft says it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to seven times the token throughput of a Maia 200 rack.

Microsoft did not name the GPU systems behind the 40% figure, as The Next Web points out. The chip is built for long-context reasoning, latency-sensitive chat and multi-step agent tasks, Guthrie wrote.

The predecessor is already in service. Maia 200 now serves Microsoft production workloads including Microsoft Copilot, runs OpenAI models, Microsoft's own MAI models and open source model families, and is deploying across the United States, Europe and Australia, according to Microsoft. GitHub Copilot and Microsoft Foundry are to follow by the end of 2026. Microsoft claims Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot and Cowork.

The multi-gigawatt goal comes without a schedule. Alistair Speirs, a Microsoft general manager, declined to give a timeline, telling CNBC that the company wants to "reserve a little bit of optionality there, because this is an industry that's changing very fast..."

How does Maia 300 compare with Maia 200?

On Microsoft's own published specifications, Maia 300 raises memory bandwidth per chip from 7 TB/s to 18 TB/s and doubles the number of chips per rack to 96, while chip power rises from 750 W to 950 W. Its compute chiplets move from TSMC's 3 nanometer process to TSMC N2P.

The Maia 200 rack count of 48 follows from Microsoft's statement in its architecture article that Maia 300 doubles accelerator density. The same article says the chiplet design adds 36% compute, 2.5 times the memory bandwidth and 14% more scale-up bandwidth, and that a Maia 300 rack delivers more than 2.5 times the performance per watt of Maia 200. HBM, or high bandwidth memory, is the fast memory stacked next to the chip that holds model weights and the running context of each conversation.

The comparison set has changed since January. Microsoft's Maia 200 announcement measured the chip against Amazon's third generation Trainium and Google's seventh generation TPU; the Maia 300 announcement compares it with Maia 200 and with GPU-based systems it does not name.

Why is Microsoft moving its inference work onto Maia?

Microsoft presents Maia as a way to get more AI output from the same power and money. "You can think of power translating pretty directly into operational cost for something like a data center as well," Speirs told CNBC.

The company has made the cost case before. CEO Satya Nadella told analysts in July 2026 that Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet," and finance chief Amy Hood said in September that Microsoft's own chips have helped with "margin improvement over the course of a decade," according to CNBC.

Each Maia 300 rack holds 96 chips in a 167 kilowatt power envelope, according to Microsoft, while Nvidia's GB300 racks hold 72 GPUs and draw about 142 kilowatts (Reuters via The Next Web). Dividing one gigawatt by that rack power works out to about 6,000 Maia 300 racks, or roughly 575,000 chips, before cooling and other data center overhead. In August, The Information reported that Microsoft was in talks with TSMC to make more than 300,000 Maia 300 chips for 2027 (The Information via The Next Web), which works out to a little over half a gigawatt of rack power on the same basis.

Nvidia stays in the picture. Microsoft works with Nvidia and AMD to train models, Speirs told CNBC, and Nvidia CEO Jensen Huang spoke at a Microsoft event the week before, where the company presented the Nvidia-powered Surface Laptop Ultra, CNBC notes. Microsoft AI CEO Mustafa Suleyman ties the chip to Microsoft's own models: "At that scale we need a platform that we own, and that we co-design with our models," he told The Verge.

What we don't know yet

• Which GPU systems Microsoft used for its 40% tokens per megawatt claim, and how the test was run.
• How many gigawatts Microsoft plans, and on what schedule beyond the 2027 start.
• Whether Microsoft will pass any cost savings on to Copilot, GitHub Copilot or Foundry customers.
• Whether Maia 300 will ever be offered to outside companies as a rentable Azure service.