Microsoft has announced plans to deploy its self-designed AI chip, Azure Maia, at multi-gigawatt scale starting in 2027. The company plans to make it a key inference platform across its product lineup, an initiative announced by executive Scott Guthrie. Maia 200 is already powering Microsoft Copilot in the U.S., Europe and Australia, and the company plans to extend it to GitHub Copilot and Foundry later this year.
Choosing to Control the Infrastructure That Runs AI
When people talk about AI, attention tends to focus on which models are the smartest. But as the number of users grows, so does the computing required to generate each response. Who runs this process, known as inference, on what hardware and with how much power? This announcement delves into that underlying infrastructure.
Microsoft says it will deploy Maia at multi-gigawatt scale starting in 2027 and make it a core inference system for its products. CEO Satya Nadella reportedly described it as “the biggest bet we’ve ever made on our own silicon.” The company appears intent on embedding chip design and operation into its cloud strategy, rather than relying solely on chips sourced from outside vendors.
One point worth noting is the range of models Maia will run. Microsoft says it will support OpenAI models, its own MAI series and open-source models. Rather than a specialized chip tailored to a single model, Maia appears intended to serve as a general-purpose inference platform for the diverse AI workloads running on Azure.
Where Things Stand: Maia 200 Is Already Powering Copilot
This is more than just a plan for the future. Maia 200 is already running Microsoft Copilot in the U.S., Europe and Australia. Support for GitHub Copilot and Foundry is also expected to expand later this year.
Microsoft says Maia can process up to 1.7 times as many tokens per watt as leading GPUs in Teams. A token is roughly the smallest unit AI uses to process text. Higher throughput per unit of power means more users can be served with the same infrastructure, directly affecting operating costs. However, this figure comes from Microsoft, and the publicly available information does not specify the comparison targets or measurement conditions.
Maia 300 Targets Long Inputs and AI Agents
The successor, Maia 300, is scheduled to ship in 2027. Its intended workloads include processing long inputs, fast interactions and agentic tasks. An agent is an AI that autonomously performs 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 than conventional conversational AI.
Microsoft has also disclosed the chip's specifications. Each chip has eight stacks of HBM4 memory, with a total capacity of 288GB, bandwidth of 18TB per second and a rated power consumption of 950 watts. HBM is high-speed memory used in AI systems; its capacity and speed become increasingly important when handling large models or long contexts.
Microsoft claims Maia 300 will deliver 40% more tokens per megawatt than the preceding GPU systems. It also says a rack comprising 96 chips can generate up to seven times as many tokens as a Maia 200 rack. The details are outlined in a technical report, with more information expected at Ignite 2026 in November.
Why Processing per Unit of Power Is Becoming a Competitive Battleground
What stands out in this announcement is that performance is measured not just by “speed,” but by how much can be processed per unit of power. The fact that Microsoft describes the scale of its plans in gigawatts, a unit of power, is itself symbolic. As data centers expand, securing electricity is becoming a constraint even more significant than the number of chips.
A difference of tens of percentage points in power efficiency can change how much AI service can run within the same power allocation. Designing its own chips allows Microsoft to fine-tune them for its services and potentially achieve greater efficiency gains. This seems to be one reason the company is doubling down on developing its own hardware.
At the same time, the public information does not make clear whether “multiple gigawatts” refers to the chips' power consumption or the capacity of the facilities as a whole. The actual deployment scale and the conditions under which each performance figure was measured are likely to be clarified at Ignite 2026 or in the technical report. Maia's rollout reflects a broader shift: AI's value will depend not only on the intelligence of its models, but also on the power and semiconductor design capabilities that support them.
Comments