Microsoft plans to deploy its Maia AI accelerators at multi-gigawatt scale beginning in 2027 and make the family a primary inference system for its products and services. Maia 200 already serves Microsoft Copilot and the company’s MAI team, while GitHub Copilot and Microsoft Foundry are due to follow by year-end.
The next Maia 300 is designed for long-context, memory-heavy inference. Each accelerator has 288 GB of HBM4 memory with 18 TB per second of bandwidth within a 950-watt power envelope. A rack contains 96 accelerators, and optical networking can join 128 racks into a 12,288-accelerator superpod. Microsoft uses one Ethernet-based fabric for communication within and across racks, including moving key-value cache state from prompt-processing workers to token-generation workers.
Microsoft expects Maia 300 to produce 40% more tokens per megawatt than leading GPU systems, but it did not identify those systems, define a standard workload or explain what it counts as a useful token. Practitioners cited by InfoQ questioned whether the metric accounts for context length and time.
Customer access is also unresolved. Every named workload is a Microsoft service, and the company has not said whether Azure customers will be able to provision Maia capacity directly. More technical information is promised at Microsoft Ignite.