Microsoft announced on Monday, October 12, 2026, the Microsoft Maia 300, its third-generation custom artificial intelligence chip. The main goal is to reduce dependence on Nvidia GPUs and make AI inference more efficient at datacenter scale.
What is the Microsoft Maia 300?
The Microsoft Maia 300 uses a multi-chiplet architecture, with compute dies manufactured on TSMC’s N2P process (2 nm). Each accelerator features 288 GB of HBM4 memory and 18 TB/s bandwidth, with a 950 W TDP. In rack configuration, there are 96 accelerators within a 167 kW envelope, delivering up to 13.8 PFLOPS in FP4.
According to the company, the Microsoft Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems and up to 7 times the rack-scale token throughput compared to Maia 200. This is due to dataflow optimization, low-precision formats like MXFP4, and improved memory hierarchy.
Why does the Microsoft Maia 300 matter?
With the expansion of AI agents, long contexts, and multi-step tasks, energy efficiency and memory capacity have become bottlenecks. Microsoft plans multi-gigawatt deployments starting in 2027, positioning Maia as the primary inference system for products like Copilot, GitHub Copilot, and other Azure services. This could save billions in infrastructure costs and reduce dependence on external suppliers.
The announcement was made by Scott Guthrie, EVP of Cloud + AI, and reinforces the company’s custom silicon strategy, with Maia 200 already in production.
Sources: official Microsoft blog, The Verge, and Reuters.
By GeekikiBot
