Microsoft has announced plans to deploy its custom-designed AI chip, Azure Maia, at a multi-gigawatt scale starting in 2027. The company intends to make it a key inference platform across its products, according to executive Scott Guthrie. Maia 200 is already powering Microsoft Copilot in the United States, Europe, and Australia, and the company plans to extend it to GitHub Copilot and Foundry later this year.
The choice to bring the infrastructure powering AI in-house
Discussions about AI tend to focus on which models are the smartest. But as the number of users grows, so does the computing power needed to generate each response. Who runs this process, known as inference, on what hardware, and with how much electricity? This announcement takes on that foundational question.
Microsoft says it will deploy Maia at a multi-gigawatt scale starting in 2027 and make it a key inference system for its products. CEO Satya Nadella reportedly described it as “the biggest bet we’ve ever made on our own silicon.” The move suggests Microsoft is looking beyond relying solely on chips procured from outside suppliers and integrating chip design and operations into its cloud strategy.
One notable point is the range of models Maia will support. The company says it will run OpenAI models, Microsoft’s own MAI series, and open-source models. Rather than being a specialized chip tailored to a particular model, it appears to be intended as a general-purpose inference platform for the diverse AI applications running on Azure.
Maia 200 is already powering Copilot
This is more than a plan for the future. Maia 200 is already running Microsoft Copilot in the United States, Europe, and Australia. Support for GitHub Copilot and Foundry is also expected to expand later this year.
Microsoft says that, in Teams, Maia can process up to 1.7 times as many tokens per watt as leading GPUs. A token is roughly the smallest unit AI uses to process text. Higher throughput per unit of electricity means more users can be served with the same infrastructure, directly affecting operating costs. However, this figure comes from Microsoft, and details about the comparison and test conditions are not available in the published information.
Maia 300 targets long inputs and AI agents
The successor, Maia 300, is scheduled to ship in 2027. Its intended uses include processing long inputs, fast interactions, and agentic tasks. An agent is an AI that autonomously carries out multiple steps in response to a user’s instructions. Because agents may call a model repeatedly for a single request, they tend to require more computing power than conventional chat interactions.
Microsoft has also published the chip’s specifications. Each chip has eight stacks of HBM4 memory, for a total capacity of 288 GB, bandwidth of 18 TB per second, and a rated power consumption of 950 watts. HBM is a type of high-speed memory used for AI; its capacity and speed become increasingly important when handling large models or long contexts.
Microsoft claims Maia 300 delivers 40% more token throughput per megawatt than earlier GPU systems. It also says a rack made up of 96 chips can generate up to seven times as many tokens as a Maia 200 rack. Details are outlined in a technical report, with more information expected at Ignite 2026 in November.
Why throughput per watt is becoming a competitive battleground
What stands out in this announcement is that performance is being measured not just by “speed,” but by how much can be processed per unit of electricity. The fact that the scale of the plan is described in gigawatts—a unit of power—is itself telling. As data centers expand, securing enough electricity is becoming a greater constraint than the number of chips.
A difference of tens of percentage points in power efficiency can change the volume of AI services that can run within the same power budget. Designing chips in-house makes it possible to fine-tune them for a company’s own services and potentially achieve greater efficiency gains. This appears to be one reason Microsoft is doubling down on its in-house approach.
However, the public information does not make clear whether “multiple gigawatts” refers to the chips’ power consumption or the capacity of the overall facilities. The actual scale of deployment and the conditions under which the performance figures were measured are likely to become clearer through Ignite 2026 and the technical report. The value of AI will increasingly depend not just on how smart its models are, but also on the electricity and semiconductor design capabilities that support them. Maia’s trajectory reflects that shift.
Comments