Microsoft Announces Maia 300 AI Inference Chip for 2027 Deployment
Microsoftが独自AI推論システムMaia 300を発表。96基搭載ラックやHBM4、2027年の導入計画を解説。
Microsoft announced Azure Maia 300 on October 12, 2026. Production deployments are planned for 2027, with a multi-gigawatt expansion of the Maia family.
The accelerator targets long-context inference, low-latency conversation and multi-step agent workloads. Its design optimizes compute, memory and networking together.
A full rack contains 96 accelerators in a 167-kW power envelope. Microsoft claims up to seven times the rack-scale token throughput of Maia 200; these are company figures.
Eight HBM4 stacks deliver up to 18 TB/s of memory bandwidth per accelerator. A chiplet design separates compute, memory control and I/O functions.
Microsoft expects 40 percent more tokens per megawatt than leading GPU-based systems, but results depend on the comparison configuration and workloads.
The announcement points to greater in-house AI infrastructure for services such as Copilot and Microsoft Foundry. Actual deployment and efficiency should be assessed in 2027.