Microsoft Unveils Maia 200: The 3nm Beast Designed to Break NVIDIA’s AI Grip

Microsoft unveiled the Maia 200, a 3nm AI inference chip with 140B transistors, delivering >10 PFLOPS (FP4), >5 PFLOPS (FP8), sub-750W TDP, and 216GB HBM3e at 7 TB/s, claiming superior cost-performance over Google TPU and AWS Trainium; initially deployed in Iowa data centers to power Copilot and future GPT-scale models, it targets large-scale inference workloads, uses Ethernet interconnects instead of InfiniBand, and aims to reduce reliance on NVIDIA GPUs for inference while leaving high-end training to NVIDIA.