Microsoft moves to reduce dependence on Nvidia with new Maia 300 AI chips, shifts more AI workloads in-house

Microsoft wants to run more of its artificial intelligence business on chips it builds itself. The software giant on Monday unveiled its third-generation Maia 300 AI accelerator and announced plans to shift much of the computing behind its AI products away from Nvidia GPUs and onto its own silicon starting in 2027.

The move marks a significant shift for Microsoft, one of Nvidia’s biggest customers, as the company looks to gain greater control over the hardware running its Azure cloud and AI services. It comes as surging demand for AI computing has made specialized chips increasingly expensive and electricity a major constraint for data center operators.

Announcing the new chip, Scott Guthrie, Microsoft’s executive vice president of Cloud + AI, outlined the company’s ambitions in a blog post published Monday.

“As we move into the era of Super Intelligence, Microsoft is entering a new phase of infrastructure. We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.”

The announcement marks a significant change in Microsoft’s AI infrastructure strategy. Rather than relying primarily on outside chip suppliers, the company wants its own processors to handle the growing volume of requests flowing through its AI products.

The transition focuses on inference, the process of running trained AI models to generate answers, write code, and complete other tasks. Nvidia will remain an important supplier for training large AI models.

Microsoft said Maia 300 will become the primary system for AI inference across its products and services. Inference is the process of running trained AI models to generate responses, write code, answer questions, and perform other tasks.

The company is betting that designing chips around its own workloads can help it deliver faster AI services at lower operating costs.

According to Microsoft’s technical specifications, Maia 300 can process 40% more AI tokens per megawatt than leading GPU-based systems. A fully configured rack containing 96 Maia 300 accelerators can process up to seven times as many tokens as a rack using the previous-generation Maia 200.

Those figures are based on Microsoft’s own performance claims. The company has yet to demonstrate how its newest chips perform under independent testing.

Microsoft’s AI chip strategy puts Nvidia’s dominance to the test

Microsoft’s relationship with Nvidia has been central to its AI expansion. Nvidia GPUs run across Microsoft’s data centers, supporting Azure customers and AI models developed by companies including OpenAI and Anthropic.

That relationship is not ending. Microsoft still expects to rely on Nvidia and AMD processors for training large AI models. Its new strategy focuses primarily on inference, where the costs of processing millions of everyday AI requests can add up quickly.

The distinction matters. Training an AI model requires enormous computing resources, but running that model for millions of customers creates an ongoing expense. Companies that can reduce the cost of each AI response can improve margins as usage grows.

Microsoft has spent years preparing for this transition.

It introduced its first Maia chip in 2023, followed by Maia 200 in January 2026. The second-generation processor is already handling production workloads, including Microsoft Copilot and OpenAI models.

Microsoft CEO Satya Nadella told analysts in July that Maia 200 delivered “30% better performance per dollar than the latest generation hardware in our fleet.”

Maia 300 takes that effort further.

Microsoft plans to begin production deployments in 2027 and eventually scale the Maia family to multi-gigawatt installations across its data centers. The latest design doubles the number of accelerators per rack compared with Maia 200, increasing the amount of AI computing Microsoft can fit into its existing infrastructure.

The chips could support faster interactions across Microsoft 365 Copilot, GitHub Copilot, Teams, and other AI services.

Energy efficiency is another major factor. As AI data centers consume more electricity, getting more computing work from the same energy supply has become a competitive advantage.

“We need new architectures that will give us essentially better performance, better use of the power that we have,” said Alistair Speirs, a Microsoft general manager. “You can think of power translating pretty directly into operational cost for something like a data center as well.”

Microsoft is joining a broader industry effort to loosen Nvidia’s grip on AI infrastructure. Google has invested heavily in its tensor processing units, Amazon develops Trainium chips, and other companies are exploring alternatives to conventional GPUs.

Microsoft’s approach still has limits. Unlike Google’s and Amazon’s custom accelerators, Maia chips are not currently available for outside customers to rent independently through Azure, CNBC reported.

“Nobody has tested them outside of Microsoft yet,” said Patrick Moorhead, CEO of Moor Insights and Strategy.

For Nvidia, Microsoft’s announcement does not mean an immediate loss of its largest customers. The chipmaker remains deeply embedded in AI model training and cloud infrastructure.

But Microsoft’s strategy points to a bigger industry shift. The companies spending billions on AI infrastructure increasingly want to own more of the technology behind it.

For Microsoft, Maia 300 is less about replacing Nvidia overnight and more about reducing the cost of running AI at a scale that could eventually reach billions of daily interactions.