Microsoft unveils Maia 300, its in-house inference chip set to replace GPUs in its services

Maia 300, Microsoft’s third-generation AI chip, promises 40 per cent more tokens per megawatt than GPU-based systems and will be deployed on a gigawatt scale from 2027.

On October 12, Microsoft introduced Maia 300, the third generation of its in-house designed AI accelerator. Engraved in N2P at TSMC, the chip adopts a chiplet architecture that separates compute, memory controller, and I/O, in order to devote more silicon surface area to pure computing. It is explicitly aimed at recent workloads: reasoning over long contexts, latency-sensitive conversations, and multi-step agentic tasks.

Less dependent on Nvidia and AMD

According to Scott Guthrie, Microsoft's head of cloud, Maia 300 is expected to produce about 40% more tokens per megawatt than the best GPU-based systems, and a full array of 96 accelerators would achieve up to seven times the throughput of Maia 200, which was announced only in January. The group plans multi-gigawatt deployments starting in 2027, with the goal of making Maia the primary inference system for its products, from Copilot to Foundry.

Mustafa Suleyman, head of Microsoft AI, sees it as a centerpiece for his own models: co-designing the hardware and models should help the company catch up with Google DeepMind, OpenAI and Anthropic.