Microsoft Officially Enters the AI Chip Race With the Launch of Maia 200

Microsoft has kicked off 2026 with the official launch of the Maia 200 AI chip. That’s the new in-house silicon built for large-scale AI deployments. In the announcement post, the company says the chip is designed to push performance per dollar, especially for newer models that make heavy use of low-precision compute.

Maia 200 is built on TSMC’s 3nm process, and delivers top-notch performance

Microsoft’s Maia 200 AI chip is built on TSMC’s 3nm process and packs over 140 billion transistors. It is designed specifically for low-precision compute. Microsoft claims the chip can offer more than 10 petaFLOPS of FP4 performance. That’s not all; it also delivers over 5 petaFLOPS at FP8, all within a 750W power envelope.

To address memory demands, Maia 200 comes with 216GB of HBM3e memory, which offers up to 7 TB/s of bandwidth. That comes with 272MB of on-chip SRAM and custom data movement engines. The company further notes that this memory setup is designed to keep performance consistent even as models grow larger and more complex.

Ethernet-based scaling at cloud scale

As far as system is concerned, Maia 200 uses a two-tier scale-up network built on standard Ethernet rather than proprietary fabrics. Each chip exposes 2.8 TB/s of bidirectional bandwidth, which supports predictable performance across clusters of up to 6,144 chips. Nit to forget, Microsoft has used four Maia chips inside a single tray, which are connected directly using non-switched links. This keeps communication local, while helping reduce power use and overall costs.

The AI chip is built for for Azure from day one

Maia 200 is already live in Microsoft’s US Central datacenter region, with more regions planned. The AI chip has been integrated directly with Azure’s control plane,which allows for security, telemetry, diagnostics, and management at both the chip and rack levels.

Last but not least, Microsoft is also previewing the Maia SDK, which includes PyTorch support, a Triton compiler, optimized kernels, and low-level programming tools for developers who want extra control. Just in case you are looking for some extra information, feel free to check here.