Microsoft has unveiled the Maia 200, a proprietary AI accelerator specifically designed for “inference”—the execution of AI models. With this new processor, Microsoft promises to drastically reduce the cost of generating AI tokens, while delivering performance that surpasses that of competitors like Amazon and Google.
Made by Scott Guthrie (Executive Vice President Cloud + AI), the announcement marks a significant step in Microsoft’s strategy to reduce its reliance on third-party chip manufacturers like Nvidia. According to Microsoft, the Maia 200 is the most powerful first-party chip from any major cloud provider to date.
Powerhouse at 3 nanometers
The Maia 200’s specifications are impressive. The chip is fabricated on TSMC’s cutting-edge 3nm process and contains over 140 billion transistors. The focus is entirely on efficiency for large-scale workloads:
The chip delivers over 10 petaFLOPS of 4-bit precision (FP4). This is three times faster than the third-generation Amazon Trainium, says Microsoft.
To quickly power massive models like GPT-5.2, the chip is equipped with 216GB HBM3e memory with a bandwidth of 7 TB/s.
The hardware offers a 30% improvement in performance-per-dollar compared to current generation hardware in Azure datacenters.
Fuel for GPT-5.2 and Copilot
The new chip isn’t just a theoretical success; it’s already working in practice. Microsoft confirms that the Maia 200 will be used for the latest OpenAI models, including GPT-5.2. The chip will also power Microsoft 365 Copilot and support the internal Superintelligence teams in generating synthetic data for training future models.
What makes the Maia 200 unique is its speed of implementation. Thanks to an innovative “pre-silicon” development method, AI models could be fully operational within days of the first physical chips leaving the factory.
Global rollout and liquid cooling
The first units are already operational in the Azure region in Iowa (US), with Phoenix, Arizona, next on the list. Due to the high power consumption of 750W per chip, Microsoft uses an advanced closed-loop liquid cooling system to manage the enormous heat in the data centers.
While Microsoft is primarily deploying the chips for its own use and Azure customers, it is also making the Maia SDK available today to developers, AI startups and academics. This allows them to optimize their models for this new architecture.
