TAIPEI (Taiwan News) — Microsoft on Monday launched its second-generation in-house AI chip, Maia 200, manufactured using TSMC’s advanced 3 nm process.

The chip went live this week at Microsoft’s Iowa data center. It will be deployed at a second facility in Arizona, following the company’s first-generation Maia chip introduced in 2023, per Reuters.

Like Nvidia’s upcoming Vera Rubin chip, Maia 200 features high-bandwidth memory. It also includes FP8/FP4 tensor cores, 216GB HBM3e memory at 7 TB/s, 272 MB of on-chip SRAM, and optimized data movement engines.

Microsoft said Maia 200 delivers three times the FP4 performance of Amazon’s third-generation Trainium and surpasses Google’s seventh-generation TPU in FP8 workloads. It offers 30% better performance per dollar compared with previous-generation hardware.

The chip will support Microsoft’s AI infrastructure, powering GPT-5.2 models from OpenAI and Microsoft 365 Copilot. Its design accelerates domain-specific data generation and reinforcement learning for downstream AI training.

Microsoft is also previewing a Maia SDK with PyTorch integration, a Triton compiler, optimized kernel libraries, and a low-level programming language for fine-grained control. Developers can port models across accelerators and optimize performance on the chip.