Microsoft Plans Mass Rollout of Custom Maia AI Chips Starting in 2027

Microsoft plans to deploy its in-house Maia artificial intelligence processors on a massive scale starting in 2027, making the custom silicon the primary hardware for running AI workloads across its product lineup, as reported by TechXplore (https://techxplore.com/news/2026-10-microsoft-homegrown-ai-chip-nvidia.html). The strategy aims to cut Microsoft's infrastructure reliance on Nvidia as operational demand for AI compute escalates.

The planned deployment will draw several gigawatts of electricity across Microsoft's global data center footprint, an energy load equivalent to the output of several commercial nuclear reactors. While Nvidia maintains dominant market share in chips used to train AI models, cloud operators and chipmakers are increasingly competing for inference workloads—the ongoing execution of trained models to generate responses and run autonomous software agents. Competitors including Amazon, Alphabet's Google, and AMD are also building out custom hardware for inference.

"Today marks a new phase for Microsoft's AI infrastructure," Scott Guthrie, Microsoft executive vice president, wrote in a blog post announcing the rollout.

Microsoft disclosed details for two generations of the hardware architecture. The current Maia 200 processor is operational in data centers across the United States, Europe, and Australia. It handles workloads for Microsoft 365 Copilot and Teams, running OpenAI models, Microsoft's in-house MAI models, and leading open-source models. According to Microsoft, the Maia 200 delivers up to 1.7 times more output per watt than leading GPU-based systems on Copilot and Teams features.

Its successor, the Maia 300, is scheduled to launch in 2027. Microsoft claims the Maia 300 will generate 40% more AI output from the same power draw compared to leading systems, and more than double the energy efficiency of the Maia 200. Taiwan Semiconductor Manufacturing Co. (TSMC) will fabricate the computing components for the new chip.

To support density and power efficiency, Microsoft designed server racks to hold 96 Maia chips per cabinet—doubling previous density—while processing up to seven times the workload. In an accompanying engineering post, Microsoft noted that architecture teams focused on reducing power wasted when transferring data internally. The systems are primarily liquid-cooled but remain compatible with legacy air-cooled data center facilities.