Microsoft Unveils Maia 300, Plans Multi-Gigawatt AI Rollout From 2027

Microsoft plans to expand its Azure Maia chip family to deployments measured in multiple gigawatts beginning in 2027, putting its internally designed hardware at the center of the infrastructure serving its AI products. The company announced the expansion alongside Maia 300, its next-generation AI accelerator, on October 12.

The plan moves Maia beyond a supporting role. Scott Guthrie, Microsoft’s executive vice president for Cloud and AI, said the family would become the primary inference system for Microsoft products and services. Inference is the computing performed when an already-trained model processes a request and produces an answer, rather than the earlier work of training the model.

Microsoft’s announcement places the forthcoming rollout alongside an existing deployment. Maia 200 is already handling production workloads, including Microsoft Copilot, and supporting the company’s MAI model development team. GitHub Copilot and Microsoft Foundry are scheduled to follow by the end of 2026.

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. Tokens are the units of information that language models process and generate. Measuring their output against power consumption gives operators one way to assess how much AI work they can support within available electrical capacity.

A separate comparison concerns Microsoft’s own hardware. In its Maia 300 overview, the company said the new systems are expected to provide up to seven times Maia 200’s token throughput at rack scale.

The business question is how those improvements translate into the cost of delivering services. Greater output within a fixed power allowance could help Microsoft accommodate more requests without a proportional increase in electrical capacity. It would not, by itself, establish lower overall costs: equipment, facilities, cooling, software and utilization also affect the economics of an operating system.

Maia 300’s engineering changes address more than raw computing speed. Microsoft’s technical description emphasizes keeping information near the processing resources that need it, reducing movement across the system and coordinating work among accelerators. These are central design considerations for models handling lengthy inputs and tasks involving multiple steps.

The accelerator uses a chiplet design that separates computing, memory-controller and input-output functions. Its computing chiplets are manufactured using TSMC’s N2P process. Microsoft specifies eight HBM4 memory stacks per accelerator, providing 288 gigabytes of capacity and peak memory bandwidth of 18 terabytes per second.

At rack level, the design combines 96 accelerators within a 167-kilowatt power envelope. Microsoft also describes a network connecting as many as 128 racks into a “superpod” of 12,288 accelerators.

Cooling is another deployment consideration. According to Microsoft, liquid cooling removes more than 80% of the rack’s thermal load, while air cooling supports remaining components. The design can connect to facility water or use a heat exchanger in an air-cooled environment. Those options concern how the equipment fits into data centers, alongside its computational specifications.

Microsoft identifies coding, model deployment and conversational services among the applications intended to benefit from Maia 300. Its product announcement points to GitHub, Microsoft Foundry, Copilot, Teams and Dragon Copilot, connecting the hardware program to services with different response-time and processing requirements.

Those workloads create a practical test for the new system. A configuration that delivers high aggregate output must also handle the mix of requests that reaches a production service. Longer inputs, concurrent sessions and multistep tasks can place different demands on memory and networking, which Microsoft has made explicit priorities in the design.

The transition also gives Microsoft a larger responsibility for matching its software to its infrastructure. Developing hardware around its own services offers an opportunity to improve that fit, but projected specifications do not yet show the financial outcome of the 2027 deployment.