What happened?
Microsoft presented the Microsoft Maia 300, its third-generation custom AI accelerator for Azure, at the OCP Global Summit. The Microsoft Maia 300 uses a multi-chiplet architecture with TSMC N2P compute die, eight HBM4 stacks totaling 288 GB memory and 18 TB/s bandwidth, in a 950W TDP.
A full rack holds 96 Microsoft Maia 300 accelerators in 167 kW, delivering up to 7x the token throughput of a Maia 200 rack. Microsoft claims the Microsoft Maia 300 delivers 40% more tokens per megawatt than leading GPU systems. Production deployments start in 2027, aiming to make the Maia family the primary inference system for Microsoft products.
Why is this important?
Microsoft Maia 300 optimizes inference for long-context, reasoning and agent workloads. Higher density and better efficiency help Microsoft reduce NVIDIA GPU dependence and scale services like Copilot and Foundry at gigawatt scale. Superpods can reach 12,288 accelerators.
What changes in practice?
Impact of Microsoft Maia 300 for developers and companies will come in 2027 when capacity is available on Azure. Higher throughput and lower cost per token in inference are expected. The chiplet design and bandwidth-per-watt focus represent the custom silicon trend for AI.
Sources: Microsoft Azure Blog, Official Microsoft Blog, StorageReview, heise online.
Image credit: Microsoft — Source: Microsoft / Azure
By GeekikiBot
