Microsoft Plans to Run Its AI Products Primarily With Its Own AI Chips

Microsoft intends to primarily utilize its custom Maia AI chips for powering its AI products starting in 2027, according to the company's recent announcement. The Maia 300, the third generation of Microsoft's custom AI accelerator, is expected to be deployed at multi-gigawatt scale next year, potentially amounting to over one million Maia 300 chips by that year.

Built specifically for AI inference, the Maia 300 chip is optimized for tasks such as long conversations, AI reasoning, AI agents, and multi-step tasks. Microsoft claims that Maia 300 can generate 40% more tokens per megawatt of power compared to leading GPU-based systems, although the specific competing GPU models were not disclosed.

Each Maia 300 rack features 96 chips and consumes around 167 kilowatts of power, making it a denser system in terms of accelerator count, although Microsoft's primary focus is on improving output per watt. The Maia 300 utilizes TSMC's N2P process for manufacturing and incorporates eight stacks of HBM4 memory per rack. Microsoft has already deployed the previous-generation Maia 200 for workloads related to Copilot, OpenAI models, and Microsoft's MAI models, and plans to continue using Maia 200 for GitHub Copilot and Microsoft Foundry by the end of the year.

The Maia program aligns with a broader industry trend among large cloud providers and AI companies to develop their own chips, following similar strategies adopted by Google and Amazon. Microsoft is working to secure sufficient manufacturing capacity for the rollout, with earlier reports suggesting discussions with TSMC to manufacture over 300,000 Maia 300 chips for 2027.

The company anticipates higher volumes in the future as demand for AI computing continues to grow, and more details about Maia 300 will be revealed at its upcoming Ignite conference in San Francisco.