Microsoft Unveils Maia 300 AI Chip for 2027 Deployment

Microsoft Corp. has unveiled Maia 300, its newest internally developed artificial intelligence accelerator designed for inference workloads. The company plans to make these custom chips the primary system for running AI in its products starting in 2027. This launch marks a significant shift toward using self-designed silicon for core operations to reduce dependence on industry partners like Nvidia.

Microsoft estimates the chip can process 40% more tokens per megawatt than leading GPU-based systems. One report suggests this comparison likely refers to products from Nvidia Corp., while another notes the company did not name its specific benchmarks in the announcement. Regardless of the reference, the goal is improved efficiency for every watt of power used during AI queries.

The hardware uses a design split into smaller pieces called chiplets, which are made using Taiwan Semiconductor Manufacturing Co.’s latest two-nanometer process. Each chip carries eight stacks of HBM4 memory with 288 gigabytes of combined capacity to support large models. This architecture allows computing parts to keep information in internal SRAM cache to avoid sending data across the chip unnecessarily.

To manage power and data flow, Microsoft includes a device control chiplet alongside two auxiliary chiplets for movement tasks. The system also compresses data into lightweight formats such as MXFP4 to reduce memory footprint and data movement. Executives stated these narrower representations allow the system to support larger models and longer contexts within the same power envelope.

Microsoft plans to deploy the chips in trays that hold four accelerators apiece, placed inside liquid-cooled racks containing 96 chips total. A full rack delivers up to seven times the output of the previous-generation Maia 200 chip but draws about 167 kilowatts of power. By comparison, Nvidia’s latest GB300 racks hold fewer GPUs and use less power per system according to Reuters reporting.

To connect the chips, Microsoft uses a custom Ethernet-based network stack that adjusts bandwidth allocation for specific workloads. This design allows jobs to use capacity across rack boundaries without needing extra hardware, removing the need for separate host server components. The company intends to group these racks into superpods with thousands of accelerators linked together for increased performance.

Microsoft’s move is part of a larger trend among tech giants to design their own AI chips and reduce reliance on Nvidia. Google and Amazon have their own AI chips, and OpenAI has built its own chip with Broadcom. However, Microsoft and Nvidia remain big partners, with Nadella noting Microsoft will be among the first cloud providers to deploy next-generation systems based on Nvidia’s Vera Rubin.

The current generation Maia 200 is already in use for services like Copilot and runs models from OpenAI as well as Microsoft’s own. Analysts estimate that Maia will account for about 10% of Microsoft’s spending on AI chips by early 2027. The company says it will share more details next month at its Ignite conference in San Francisco.