Microsoft unveils Maia 200, its 3 nm inference accelerator

2026-01-26 · Publisher : Microsoft · Domain : Cloud & DevOps

Microsoft introduces Maia 200, an AI inference accelerator built on 3 nm, with 216 GB of HBM3e. It is already deployed in Azure's US Central region, and Microsoft claims 30% better performance per dollar.

### Background On 26 January 2026, Microsoft announces Maia 200, an AI accelerator designed for inference, that is, for producing answers with already-trained models. It is manufactured by TSMC on 3 nm. Microsoft says Maia 200 is already deployed in the Azure US Central datacenter near Des Moines (Iowa), and that the US West 3 region near Phoenix (Arizona) will follow. ### Key facts - Memory: 216 GB of HBM3e at 7 TB/s and 272 MB of on-chip SRAM. - Compute: more than 10 petaflops in FP4 and more than 5 petaflops in FP8, in a 750 W envelope. - Microsoft claims 30% better performance per dollar than the latest hardware in its fleet, and three times the FP4 performance of the third-generation Amazon Trainium, with FP8 performance above Google's seventh-generation TPU (vendor comparisons). - Networking: standard Ethernet with a custom transport layer, 2.8 TB/s of bidirectional dedicated scale-up bandwidth, up to 6,144 accelerators per cluster. - Targeted workloads: Microsoft Foundry, Microsoft 365 Copilot, OpenAI's GPT-5.2 models, and the Microsoft Superintelligence team. - A preview software development kit includes the Triton compiler and PyTorch support. ### Why it matters The three major clouds now have an in-house AI chip in production: Trainium at AWS, TPU at Google, Maia at Microsoft. For Azure customers, the benefit is indirect: Maia mainly serves Microsoft and OpenAI services, and the inference cost it lowers may be reflected in the prices of AI services. ### Takeaway Maia 200 is in production: Microsoft reduces its dependence on GPUs to serve its models.

Tags : Azure, Puces, Hyperscaler, IA