Microsoft to expand Azure Maia chips to multi-gigawatt deployments
The Official Microsoft Blog ● Covered by 4 sources
Microsoft says Azure Maia chips will scale to multi-gigawatt use starting in 2027. Maia 200 is already in production; Maia 300 is built for the AI workload crunch.
Based on reporting by The Official Microsoft Blog — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Microsoft is pushing its Maia chip program from a single product line into the backbone of its AI infrastructure. Maia 200 is already running production workloads, including Microsoft Copilot, and the company says it will expand the family to multi-gigawatt deployments starting in 2027. Maia 300 is the next step, with production deployments set to begin in 2027 as well.
The company is not treating these chips as side projects or lab toys. Maia 200 is supporting the MAI team as it develops next-generation models, and GitHub Copilot and Microsoft Foundry are expected to follow by the end of the year. Across those workloads, Microsoft says Maia 200 is running OpenAI models, its own MAI models and leading open-source model families.
The bigger point is that Microsoft is designing hardware around the kind of AI it actually runs. Maia 300 is aimed at long-context reasoning, latency-sensitive conversational systems and multi-step agent tasks. Microsoft says the new chip should deliver 40% more tokens per megawatt than leading GPU-based systems, and in a full 96-accelerator setup it offers up to 7x the rack-scale token throughput of Maia 200.
There is a lot of system design wrapped into that claim. Maia 300 uses a software-defined dataflow architecture, eight HBM4 stacks, a 950-watt TDP, a multi-chiplet design and compute chiplets built on TSMC’s N2P process node. Microsoft also says the rack version packs 96 accelerators into a 167-kilowatt envelope, with 27.6 TB of HBM capacity and 1.7 PB/s of aggregate HBM bandwidth.
Maia 200 is already being deployed across the United States, Europe and Australia, and Microsoft says it has delivered up to 1.7x more tokens per watt for built-in AI capabilities in Teams, Copilot and Cowork versus leading GPU infrastructure. That early deployment seems to be the point: build, measure, then use the results to shape the next chip. Microsoft is betting that the boring parts of AI infrastructure — bandwidth, memory, power, rack density — are where the real advantage will come from.
My take — AI-written commentary, not fact-checked reporting
Microsoft is doing the sensible thing here: less Silicon Valley poetry, more watts, bandwidth and rack density. The industry keeps talking about intelligence like it falls out of the sky; meanwhile, the bill still shows up in the datacenter. Open-ish model support on a tightly controlled Microsoft stack is exactly the sort of pragmatism big enterprises buy.
Read more about this at: The Official Microsoft Blog