It was only in January that Microsoft introduced Maia 200. Before the year is out, the company has already revealed the next generation and attached a deadline to it: from 2027, the Maia 300 is slated to take on most of Microsoft’s AI work, leaving a smaller share for Nvidia’s GPUs.
The Maia 300 is Microsoft’s third-generation AI chip developed in-house, and the company says it is intended to serve as the “primary AI inference system” behind its own products and services. Because production rollout doesn’t start until 2027, what exists today is a roadmap backed by figures rather than hardware customers can actually rent.
The number Microsoft wants you to remember
Microsoft is leading with efficiency, and it put a concrete figure on that claim.
“We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems,” said Scott Guthrie, executive vice president of Microsoft’s cloud group. “And in a full 96-accelerator configuration, it delivers up to 7x the rack-scale token throughput of Maia 200.”
Those statements deserve a careful read. The 40 percent gain is what Microsoft “expects,” and the comparison is against unnamed “leading GPU-based systems.” The 7x figure is an “up to” ceiling that only applies to a complete 96-accelerator rack. No independent testing has backed up either number, and the chip isn’t due in production until 2027.
Nvidia and AMD get demoted
What matters more is how Microsoft intends to use those gains. The company currently sources AI chips from Nvidia and AMD, but it said the efficiency improvements should soon allow it to lean primarily on its own silicon instead.
“We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary AI inference system for Microsoft products and services,” Guthrie said.
The key word there is “inference.” Microsoft is describing the running of AI models that power its products and services, not the training of those models. Still, deployments measured in multiple gigawatts would represent a sizable portion of Microsoft’s AI compute that would no longer come from Nvidia and AMD.
Maia 200 is already doing real work
Part of what lets Microsoft make this case is that the current generation is already deployed. Maia 200 handles production AI workloads throughout the company, including Microsoft Copilot, GitHub Copilot, Microsoft Foundry and Microsoft’s own AI models.
“Maia 200 is the best performing AI infrastructure in our fleet, measured in real world metrics for AI services,” Guthrie said. “Maia 300 is a complete AI system designed for the new types of AI workloads emerging today: Long-context reasoning, latency-sensitive conversational experiences, and sustained, multi-step agentic tasks.”
It is exactly these workloads that have been pushing inference costs higher. Reasoning models and AI agents burn through far more tokens than a simple chatbot response, and Microsoft said Maia 300 is better equipped than the previous chip to deal with them.
Chiplets on TSMC’s N2P
Behind the efficiency improvements is a new chiplet architecture. Maia 300 is manufactured on TSMC’s N2P process node, and rather than placing everything on a single die, it separates compute, the memory controller and I/O into distinct pieces.
“This increases the silicon area dedicated to compute performance and allows each function to use the technology best suited to its role,” Microsoft’s Azure hardware and infrastructure team said.
The chiplet approach itself isn’t novel, as AMD has relied on it for years. Applying it to a custom inference chip, however, signals that Microsoft is prepared to invest engineering resources in hardware it would otherwise purchase from others.
Microsoft’s model ambitions depend on it
The chip is also tied to Microsoft’s push to develop its own frontier models and close the gap with Google DeepMind, OpenAI and Anthropic. Microsoft AI CEO Mustafa Suleyman framed control of the silicon as a component of that push.
“We have to work across the full stack, all the way down to the silicon,” Suleyman said. “At that scale we need a platform that we own, and that we co-design with our models. That’s why running our Microsoft AI models on top of the Maia chip is a critical focus for our team.”
Google has pursued this same strategy with its TPUs for years, and Microsoft is plainly after a similar degree of control. The 2027 rollout is the promise that carries the most weight. Should Maia 300 ship on schedule and its 40 percent efficiency claim stand up against whatever Nvidia offers at that time, Microsoft will have lowered its reliance on Nvidia. Should it be delayed, Microsoft will go on buying GPUs.