By: Nana Appiah Acquaye
Microsoft has announced the launch of its newest AI accelerator, Maia 200, now online in Azure. Designed for industry-leading inference efficiency, the accelerator delivers 30% better performance per dollar compared to existing systems and is optimized for large-scale AI workloads.
Built on TSMC’s 3nm process, Maia 200 features native FP8/FP4 tensor cores, 216GB HBM3e memory with 7TB/s bandwidth, 272MB of on-chip SRAM, and advanced data movement engines. Microsoft says the system outperforms competing hardware, offering three times the FP4 performance of Amazon Trainium’s third generation and FP8 performance above Google’s seventh-generation TPU.
Microsoft Chairman and CEO Satya Nadella said, “Maia 200 represents a major step forward in our AI infrastructure, enabling customers to run advanced models faster and more cost-effectively. By combining innovation in silicon design with Azure’s global platform, we are giving businesses and developers the tools to unlock new AI possibilities.”
Maia 200 will serve multiple models, including OpenAI’s GPT-5.2, and will support Microsoft Foundry and Microsoft 365 Copilot. Its capabilities also extend to synthetic data generation and reinforcement learning, accelerating the creation of high-quality, domain-specific datasets for model training.
The accelerator is currently deployed in Microsoft’s US Central datacenter near Des Moines, Iowa, with expansion planned to the US West 3 region near Phoenix and additional locations. Microsoft is also previewing the Maia SDK, offering developers PyTorch integration, a Triton compiler, optimized kernel libraries, and low-level programming access for fine-tuned control and model portability across heterogeneous hardware.
