Microsoft Bets Maia 300 Can Loosen Nvidia’s Grip
Microsoft Bets Maia 300 Can Loosen Nvidia’s Grip
Microsoft’s push into custom AI silicon has moved quickly. It introduced Maia 100 in 2023, followed with Maia 200 in January, and says the second-generation chip is already serving production workloads across Copilot, GitHub Copilot, Foundry and its in-house models.
Now comes Maia 300, the third-generation accelerator Microsoft unveiled as it seeks to shift the economics — and the control — of running AI services. The company says the new system is designed for long-context reasoning, fast conversational responses and multi-step agentic work, the demanding tasks increasingly central to its product plans.
The central promise is efficiency. Microsoft says a rack of 96 Maia 300 accelerators can deliver up to seven times Maia 200’s token throughput, while producing 40% more tokens per megawatt than leading GPU-based systems. “We need new architectures that will give us essentially better performance, better use of the power that we have,” Microsoft general manager Alistair Speirs said, tying power use directly to data-centre operating costs.
The rollout is scheduled for 2027, when Microsoft intends to scale the Maia family to multi-gigawatt deployments and make it the primary inference system for its own products and services. Satya Nadella called it the company’s “biggest bet yet on our own silicon.”
That is a meaningful challenge to Microsoft’s long reliance on Nvidia GPUs, which still fill its data centres and support major AI customers. Yet this is not a clean break: Speirs said Microsoft will continue partnering with Nvidia and AMD to train models, while Maia chips are reserved for Microsoft’s own operations rather than rented to outside cloud customers.
In other words, Maia 300 is less a declaration of independence than a calculated rebalancing — a bid to own more of AI’s costly inference layer without abandoning the hardware ecosystem that helped build it.
Write a comment