Software King of the World, Microsoft has unveiled its Maia 300 AI accelerator and plans to deploy potentially more than a million chips in 2027 to reduce reliance on Nvidia.
According to Reuters, the third-generation chip will support Microsoft’s internal AI services, with deployment planned across infrastructure that will consume several gigawatts of electricity.
Vole claims Maia 300 will deliver about 40 per cent more useful computing per megawatt than leading GPU systems. It has not named the systems used for comparison, leaving the benchmark opponent conveniently anonymous.
The company is designing its hardware around inference, the work involved when trained models respond to requests or carry out tasks. Owning more of that hardware could help Microsoft control the costs of running its increasingly expensive digital assistants.
Maia 300 uses a chiplet design that combines smaller specialised dies in one package. Its computing components use TSMC’s 2nm technology, while other functions use cheaper 3nm manufacturing.
Vole’s architecture briefing says each accelerator carries eight HBM4 memory stacks, providing 288GB of capacity and 18TB/s of bandwidth. Software manages data placement across the chiplets to reduce the energy spent moving information.
Each rack holds 96 accelerators and consumes about 167kW, compared with 72 GPUs and roughly 142kW in Nvidia’s GB300 racks. Those figures reflect different configurations and don’t show which system delivers better performance on equivalent workloads.
Microsoft says the full Maia 300 rack can deliver up to seven times the token throughput of Maia 200. Its claimed improvement of more than 2.5 times in performance per watt is a separate comparison with its previous generation.
Azure hardware systems and infrastructure president Rani Borkar told Reuters: “It’s basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimising across the entire system.”
The company plans to make Maia its primary inference system for Microsoft products and services as deployment expands. Maia 200 already serves production workloads, including Microsoft Copilot, with GitHub Copilot and Microsoft Foundry expected to follow before year-end.
Maia 300’s Ethernet-based network can connect up to 128 racks into a group of 12,288 accelerators. Liquid cooling removes more than 80 per cent of the rack’s thermal load, according to Microsoft’s technical briefing.
