Microsoft outlines its Maia 300 AI chip

Microsoft describes Maia 300 as the third generation of its Azure Maia program and says Maia will become the primary inference system for its products and services.[1] The company calls the chip’s architecture inference-first, optimized for long-context, memory-intensive and decode-heavy workloads.[1] The announced use cases include long-context reasoning, latency-sensitive conversational experiences and sustained, multi-step agentic tasks.[1] That positioning distinguishes the stated focus from a general claim that Maia 300 will replace every accelerator: Microsoft’s announcement describes a particular workload strategy and a planned role in its own products and services.[1]

Microsoft says a Maia 300 rack will contain 96 accelerators and operate within a 167-kilowatt power envelope.[1] The rack is specified to provide 27.6 terabytes of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth.[1] Each accelerator has eight HBM4 stacks, with 288 gigabytes of capacity and 18 terabytes per second of bandwidth, according to Microsoft.[2] Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and says a full 96-accelerator configuration can reach up to seven times Maia 200’s rack-scale token throughput.[1] Microsoft presents these as expectations and specifications.[1][2]

Microsoft plans to begin production deployments of Maia 300 in 2027 and says it will start scaling the Azure Maia family to multi-gigawatt deployments that year.[1] The company says a Maia 300 superpod can scale to 12,288 accelerators.[2] Microsoft also says the chip uses a custom AI transport based on Ethernet, while its compute chiplets use TSMC’s N2P process node.[2][1] The announcement sets out a future deployment plan; it does not say that Maia 300 is already serving production workloads. Microsoft says Maia 200, by contrast, is already serving production workloads, including Microsoft Copilot.[1]

Microsoft’s plan gives its own products and services a primary inference system, but the evidence does not describe Maia 300 as a wholesale replacement for Nvidia hardware.[1] CNBC reported that Nvidia GPUs remain installed in Microsoft’s existing data centers and that Microsoft continues to partner with Nvidia and AMD to train models.[3] CNBC also reported that Microsoft’s AI accelerators are not available for other companies to rent by the hour.[3] Evercore ISI analysts estimated that Maia would account for about 10% of Microsoft’s AI-chip spending by the first quarter of 2027.[3] That estimate offers one outside view of the program’s expected scale, rather than a confirmed spending figure.