Monday brought word from Microsoft of Azure Maia 300, a third-generation in-house AI accelerator, and with it a promise: the Maia line will reach multi-gigawatt deployments beginning in 2027. Scott Guthrie, the executive who runs Cloud + AI, described the chip not as one component but as an entire intelligent system.
Years of stocking data centers with Nvidia GPUs are giving way to a different plan. Incoming inference requests should, in Microsoft’s intent, be handled largely by its own silicon. Across Microsoft’s products and services, Maia 300 takes over as the primary system.
A particular workload shape guided the design. Long context, memory-heavy work and decode-bound agent tasks are what Microsoft says its inference-first architecture targets, with compute, memory bandwidth, networking and power balanced at cloud scale.
Specific figures accompanied the announcement. Within a 950-watt envelope, eight HBM4 stacks supply 18.0 TB/s of peak bandwidth. One rack holds 96 accelerators drawing 167 kilowatts; 27.6 TB of HBM and 1.7 PB/s of aggregate bandwidth span the scale-up domain.
Versus leading GPU systems, Microsoft claims 40 percent more tokens per megawatt, plus as much as seven times the rack-scale throughput of Maia 200. TSMC’s N2P node builds the compute chiplets. Production begins in 2027, and November’s Ignite should bring more detail.
Copilot is already served by Maia 200 in the United States, Europe and Australia. By the end of the year, GitHub Copilot and Foundry follow. A firm deployment timeline was declined by Alistair Speirs, a general manager, who said the company wants optionality as models change.
