Microsoft on January 26, 2026 unveiled the second generation of its in-house AI chip, along with software tools aimed at challenging one of Nvidia’s key advantages with developers.

The new Maia 200, an AI inference accelerator chip fabricated on TSMC’s 3nm process, designed to improve the efficiency of large-scale AI workloads.

The chip features native FP8/FP4 tensor cores, a redesigned memory system with 216GB HBM3e memory at 7TB/s, and data movement engines to support massive models.

It delivers over 10 petaFLOPS in 4-bit precision and over 5 petaFLOPS in 8-bit, with a power envelope of 750W.

Maia 200 is deployed in Microsoft’s US Central datacenter near Des Moines, Iowa, with plans for expansion.

It integrates with Azure and is compatible with tools like PyTorch and Triton, allowing developers to optimize models for the hardware.

🔗 Source: Microsoft

Recent Microsoft developments