What happened?
Microsoft announced the Azure Maia 300, its third-generation custom AI chip, on October 12, 2026. The inference-optimized accelerator promises 40% more tokens per megawatt compared to leading GPU-based systems and up to 7x the rack-scale token throughput of Maia 200. Production starts in 2027, when the Maia family will become the primary inference system for Microsoft products and services, with multi-gigawatt deployments.
Why is this important?
Maia 300 uses a multi-chiplet architecture with compute on TSMC N2P process, 288 GB of HBM4 memory and 18 TB/s bandwidth. Each rack holds 96 accelerators in a 167 kW envelope, delivering 27.6 TB of HBM capacity and 1.7 PB/s aggregate bandwidth. This reduces reliance on Nvidia and AMD while improving efficiency for long-context, agentic and reasoning workloads.
What changes in practice?
Users of Copilot, Teams and other Microsoft services should see faster and more economical responses starting in 2027. The company already has Maia 200 in production. The bet on proprietary silicon accelerates the hyperscaler race for sovereign and efficient AI infrastructure.
By GeekikiBot
