Microsoft to shift AI workloads to homegrown Maia chips, easing reliance on Nvidia

Microsoft said its third-generation Maia AI accelerator, unveiled on Monday, will be its primary silicon for incoming AI requests, according to CNBC. Production of the Maia 300 is slated to begin in 2027.

The move follows years of reliance on Nvidia's GPUs, which have been the standard for AI workloads, fill Microsoft's existing data centers, and are used to train models from OpenAI and Anthropic. Microsoft's AI strength has been its Azure cloud, along with its Microsoft 365 Copilot and GitHub Copilot products, though its in-house models have yet to make a major impression in a market dominated by OpenAI, Anthropic and Google.

Microsoft introduced its first AI chip, the Maia 100, in 2023, and touted the Maia 200 in January. CEO Satya Nadella told analysts in July that the Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet" and said it was handling OpenAI models. Finance chief Amy Hood said at a Goldman Sachs conference last month that use of its own chips has supported "margin improvement over the course of a decade."

For general-purpose work, Microsoft already uses its own central processing units alongside Intel and AMD silicon. The broader AI market has seen model leaders embrace chips from a range of suppliers, including Amazon, AMD, Cerebras and Google, as demand for advanced processors keeps growing and customers seek choice.

Analysis

Why This Matters

Microsoft is the latest large cloud provider to move meaningful AI workloads onto its own custom silicon, potentially shifting the balance of a market long dominated by Nvidia.
If a major customer like Microsoft reduces its Nvidia purchases, it could pressure GPU pricing and reshape the economics of AI infrastructure.
Cloud customers may benefit from more choice about where models run and from competitive pressure on price and performance.

Background

Microsoft's Maia program began with the Maia 100 in 2023, followed by the Maia 200 earlier this year and now the Maia 300, with production due to start in 2027. The company joins other hyperscale cloud providers that have developed custom AI chips, including Google, which has used its own TPUs for years, and Amazon, which designs its own accelerators. The wider industry has also embraced alternatives from AMD and Cerebras as demand for advanced processors has surged.

Key Perspectives

Microsoft: Sees its own silicon as a margin and performance play. Nadella has said the Maia 200 delivers better performance per dollar than the latest generation hardware in its fleet, and Hood credits in-house chips with supporting margin improvement over a decade. Its Azure cloud and Copilot products give it a large internal base of AI demand to supply.

Nvidia: Remains the default standard for AI workloads and model training, filling Microsoft's existing data centers and powering training runs for OpenAI and Anthropic. It is the incumbent supplier now watching one of its biggest customers become a competitor.

Chip rivals and customers: Amazon, AMD, Cerebras and Google already supply alternatives in AI silicon, and their adoption reflects a market desire for choice beyond Nvidia. A shift by Microsoft could accelerate that diversification.

Critics and skeptics: Microsoft's in-house models have not yet broken through in a market dominated by OpenAI, Anthropic and Google, and large-scale chip production carries execution risk. Delivering on the Maia 300's 2027 timeline will be the test.

What to Watch

Whether Maia 300 production starts on schedule in 2027 and how much of Microsoft's own AI traffic it absorbs by then.
Whether Microsoft eventually offers Maia chips to its cloud customers, extending their reach beyond internal use.
How Nvidia responds as its largest customers build credible alternatives to its GPUs.