Microsoft brings Maia 200 online in Azure, positioning its custom AI chip as an alternative to Nvidia and rival cloud silicon

Maia 200: Microsoft doubles down on first-party AI compute

Microsoft has introduced its second-generation AI chip, the Maia 200, and says the accelerator is now online in Azure. The move is part of the company’s broader effort to reduce reliance on scarce, high-cost third-party AI hardware while improving the economics of serving AI workloads at hyperscale. Microsoft framed Maia 200 as a potential alternative to industry-leading processors from Nvidia and also as a competitive answer to custom chips built by cloud rivals.

According to the report, Maia 200 is built using Taiwan Semiconductor Manufacturing Co.’s 3-nanometer process. Microsoft also described its system approach: multiple chips connected inside each server, relying on Ethernet rather than the InfiniBand standard commonly used in many high-performance AI clusters. The architecture choice highlights how hyperscalers are experimenting with networking and system design to squeeze more throughput and better utilization out of each data-center rack.

Performance claims and a push toward wider availability

Microsoft executives said Maia 200 is designed for inference efficiency and claimed improved performance per dollar versus existing systems in the company’s fleet. Microsoft also suggested broader access over time, signaling that the chip will not remain exclusively an internal tool. Developers, academics and AI labs were pointed toward preview access through a software development kit, a typical path for hyperscalers that want an ecosystem to form around new silicon.

Where the chip will be used first

Early Maia 200 deployments are expected to support internal work as well as Microsoft’s AI services. The report said some of the first units would go to Microsoft’s Superintelligence team led by Mustafa Suleyman, and that the chips will also be used to power enterprise Copilot experiences and to run models that Microsoft offers to cloud customers. The broader context is straightforward: demand for AI compute remains intense, and any additional supply—especially supply a cloud provider controls end-to-end—can become a strategic advantage.

In the near term, Maia 200 will be watched for two things: whether it truly lowers inference costs at scale, and whether Microsoft can ramp deployments quickly enough to matter in a market where compute availability can determine which AI products ship first.

• Microsoft says Maia 200 is online in Azure and targets inference efficiency.
• Built on a 3-nanometer process and connected within servers using Ethernet.
• Microsoft signals wider customer availability and preview tooling over time.
• First deployments support internal teams and Microsoft’s AI services.