Particle.news
Download on the App Store

Microsoft Launches Maia 200 AI Accelerator for Azure With Inference Focus

The in-house accelerator reflects Microsoft's bid to reduce reliance on Nvidia by tailoring silicon for large-model inference.

Overview

  • Microsoft says Maia 200 delivers 10 PFlops of FP4 throughput with 216 GB of HBM3E, a 1.4 TB/s interconnect, and sub‑900‑watt power draw.
  • Azure will offer Maia 200 instances first in the US Central region, followed by US West 3 in Phoenix.
  • The company claims better performance per dollar than Google TPU v7 and AWS Trainium 3 by roughly 30 percent, with pricing not yet disclosed and independent benchmarks pending.
  • Deployments can scale to as many as 6,144 interconnected nodes, and Microsoft’s Superintelligence Team is already using the chip for synthetic data generation and reinforcement learning.
  • Maia 200 is built on TSMC’s N3P process, and a successor codenamed Clea (Maia 300) is planned for U.S. fabrication with Intel to strengthen supply security.