Microsoft startet 2027 Produktion des Maia 300 und plant Azure-Deployment im Gigawatt-Massstab

Microsoft has set 2027 as the start of production deployments for Maia 300, its next-generation AI accelerator, and says the broader Azure Maia program will scale to multi-gigawatt deployments from that year. The announcement signals a deeper push to design AI infrastructure around the workloads Microsoft runs across its cloud and software products. 1

Microsoft describes Maia 300 as a system for long-context reasoning, latency-sensitive conversations, and sustained agent tasks. Its design targets inference—the stage where a trained model responds to user requests—rather than positioning the chip only as a training accelerator. That focus makes memory bandwidth, data movement, and power use central parts of the system design. 1

The company says Maia 300 combines eight HBM4 memory stacks, peak bandwidth of 18 terabytes per second, and a 950-watt thermal design power. A full rack configuration would contain 96 accelerators, use 167 kilowatts, and provide 27.6 terabytes of HBM capacity with 1.7 petabytes per second of aggregate HBM bandwidth. Microsoft also says the compute chiplets use TSMC's N2P process. These are manufacturer specifications and plans, not independent measurements of a shipping product. 1

Microsoft's headline performance comparison is also its own claim: it expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems, and up to seven times the rack-scale token throughput of Maia 200 in a 96-accelerator configuration. UPI reported at launch that no third-party companies had yet tested Maia 300, so those comparisons should be read as projections awaiting outside validation. 1

Maia 200 is already serving Microsoft production workloads, including Copilot, and supports the company's MAI model team. Microsoft says GitHub Copilot and Microsoft Foundry are expected to follow before the end of 2026. Its stated strategy is to co-design chips, software, and services around real workloads, while maintaining an open, heterogeneous platform rather than relying on a single model or processor. 1

That does not mean Microsoft is abandoning merchant GPUs. The company is expanding its own accelerator family alongside a wider hardware ecosystem. If Maia 300 reaches production at the planned scale, Microsoft's gains will depend not only on chip specifications but on how reliably its software stack can keep accelerators, memory, and networking busy in live inference workloads.

The key test begins in 2027: whether the promised efficiency holds outside Microsoft's own comparisons, and whether the new racks translate into lower-cost, dependable capacity for Azure customers.