Microsoft to deploy Maia 300 AI chips from 2027

What Microsoft announced on October 12

Microsoft announced Maia 300, the third generation of its Azure Maia AI accelerators, in a blog post by Scott Guthrie, executive vice president of Cloud + AI. The post sets 2027 as the start of production deployments.

Maia to become Microsoft's main inference system

Microsoft plans to make Maia the primary inference system for its products and services, scaling the chip family to multi-gigawatt deployments starting in 2027. Inference is the work an AI model does after training, when it reads a prompt and writes a response.

Multi-gigawatt refers to the electricity the deployment draws. The post does not give a total gigawatt figure or name the datacenter regions.

Built for long context, chat and agents

Microsoft designed Maia 300 for three newer kinds of AI work. The post lists three:

• Reasoning over long context, where a model reads a large amount of text at once
• Conversations where response time matters
• Agents that carry out many steps in a row

According to Microsoft, moving data efficiently, more than raw compute, is becoming the bottleneck in large inference systems. Maia 300 roughly doubles HBM bandwidth per watt over Maia 200 to address that.

What the 40% and 7x figures are measured against

The two headline numbers use different baselines, so they cannot be read as one claim. A token is the unit of text an AI model processes; tokens per megawatt measures how much work a system gets out of its power budget.

Maia 300 performance claims and their baselines (Microsoft and Azure blogs)

Figure	Compared with	Metric
40% more	Leading GPU-based systems	Tokens per megawatt
Up to 7x	Maia 200	Token throughput of a 96-chip rack
More than 2.5x	Maia 200	Rack performance per watt

Expected 40% more tokens per megawatt than leading GPUs

Microsoft expects Maia 300 to produce 40% more tokens per megawatt than leading GPU-based systems. The post presents this as an expectation tied to efficiency gains across the whole system. Microsoft has not published a measured result behind it.

Up to 7x Maia 200's rack throughput at 96 chips

A full rack of 96 Maia 300 chips delivers up to seven times the token throughput of a Maia 200 rack. That figure compares whole racks, and since a Maia 300 rack holds twice as many chips, it is not a per-chip comparison. The Azure engineering post adds a per-watt figure: the Maia 300 rack does more than 2.5 times the work of Maia 200 for the same power.

Microsoft has not named the GPU baseline

Neither Microsoft post identifies the GPU system behind the 40% figure or the conditions of the comparison. Nobody outside Microsoft can check the number yet.

OpenAI, by contrast, reported 1.5–1.9x the work per watt of Nvidia GB200 and GB300 for its own inference chip. Those results and their test conditions are covered in OpenAI Jalapeno Chip: Inference Benchmarks 2026.

Maia 300 chip and rack specifications

Microsoft describes Maia 300 as a full system, with the chip, rack, cooling and network designed together. Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, who leads Azure hardware planning and architecture, published an architecture write-up the same day.

Maia 300 key specifications (Microsoft and Azure blogs)

Item	Maia 300
Memory	8 HBM4 stacks, 288 GB
Memory bandwidth	18.0 TB/s
Power (TDP)	950 W
Compute chiplet process	TSMC N2P
Rack	96 chips, 167 kW
Rack HBM	27.6 TB, 1.7 PB/s

288 GB of HBM4 at 18 TB/s in a 950 W chip

Each Maia 300 carries eight stacks of HBM4, for 288 GB of capacity and 18.0 TB/s of peak bandwidth. HBM is stacked memory placed right beside the processor, so it can move data quickly. Compared with Maia 200, bandwidth is 2.5 times higher and capacity is up 33%.

The chip splits compute, memory control and input/output into separate chiplets, smaller dies packaged together as one chip. The compute chiplets are built on TSMC's N2P process, and the whole chip is rated at a 950 W TDP, the maximum power it is designed to draw.

96 chips in a 167 kW liquid-cooled rack

A Maia 300 rack fits 96 chips on 24 trays of four, inside a 167 kW power budget. That is double the chip density of a Maia 200 rack.

Liquid cooling takes away more than 80% of the rack's heat, while air handles memory, storage and other parts. Microsoft says the rack can connect to facility water or, in air-cooled buildings, use a heat-exchanger unit.

Up to 12,288 chips in one superpod

A two-tier, Ethernet-based network links up to 128 racks, or 12,288 chips, into what Microsoft calls a superpod. Links inside a rack are electrical, and links between racks are optical. Any chip can reach any other chip in the superpod directly.

Where Maia 200 already runs

Maia 200 is already handling production work inside Microsoft, and the company says what it learned there shaped Maia 300. Microsoft runs Maia as a multigenerational program, and Maia 200 is the generation in service now.

Where Maia 200 runs today (Microsoft blog)

Item	Detail
In production	Microsoft Copilot and other services
Due by year-end	GitHub Copilot, Microsoft Foundry
Models	OpenAI, Microsoft MAI, leading open-source families
Regions	United States, Europe, Australia

Copilot now, GitHub Copilot and Foundry by year-end

Maia 200 serves Microsoft Copilot and supports the MAI team, Microsoft's in-house model group, as it builds new models. GitHub Copilot and Microsoft Foundry, the company's platform for building AI apps, are set to move onto it by the end of 2026.

The chip is not tied to a single model vendor. It runs OpenAI models, Microsoft's own MAI models and leading open-source model families. Maia 200 is rolling out across the United States, Europe and Australia.

Up to 1.7x more tokens per watt than leading GPUs

For the AI features built into Teams, Copilot and Cowork, Microsoft says Maia 200 produces up to 1.7 times more tokens per watt than leading GPU infrastructure. As with the Maia 300 claim, the GPU is not named.

What to watch at Ignite in November

Microsoft has committed to a 2027 start, but several details are still open. The company says it will share more about its AI strategy at Microsoft Ignite 2026, held November 17–20 in San Francisco and online.

• Which GPU system the 40% figure is measured against, and under what conditions
• How many gigawatts the Maia rollout will reach, and in which regions
• Whether Azure customers will be able to rent Maia capacity, which the announcement does not address

Until Microsoft fills those gaps, read the 40% figure as a projection against an unnamed GPU, and the 7x figure as a rack-to-rack comparison with Maia 200 rather than a per-chip gain. The 1.7x figure for Maia 200 rests on the same unnamed GPU baseline.