Microsoft has introduced the Maia 300, its third-generation in-house AI chip, and it has a very specific job in mind. The company says the chip is designed to serve as the “primary AI inference system” behind Microsoft’s own products and services.
What Microsoft is promising
The headline metric is energy efficiency. Scott Guthrie, executive vice president of Microsoft’s cloud group, put a number on it.
“We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems.”
For the uninitiated, tokens are the small chunks of text an AI model reads and writes. More tokens per megawatt means more AI output for the same electricity bill, which matters when power has become one of the tightest constraints on data center growth.
Note the framing, though. Guthrie said Microsoft expects that result. It is a target, not a benchmark anyone has independently verified yet.
Advertisement
The chip will start rolling out to production in 2027. So customers and investors are evaluating a promise with a lead time measured in quarters, not weeks.
From tens of thousands to hundreds of thousands
The real story may be scale. Research findings indicate Microsoft is negotiating with TSMC to manufacture over 300,000 Maia 300 chips for delivery in 2027.
Longer term, the company is reportedly eyeing production exceeding a million units. That ambition depends on component availability and advanced packaging capacity, two areas where the entire industry has been fighting over a limited supply.
Compare that to the previous generation. The Maia 200, released in January 2026, saw limited production of only tens of thousands of units.
The Maia 200 itself was built on TSMC’s 3nm process, carried 216GB of HBM3e memory alongside substantial on-chip SRAM, and achieved 30% better performance per dollar than Microsoft’s previous hardware.
The custom silicon arms race
Microsoft is not alone in this. Google and Amazon have their own custom AI chips, and the big cloud providers have all concluded that relying entirely on one supplier for their most important hardware is an uncomfortable position.
Research findings indicate Microsoft has been in discussions about broader adoption of Maia chips beyond its own operations, with potential interest from customers such as Anthropic.
Those are discussions, not signed deals. Still, if Maia chips eventually serve outside customers, Microsoft would shift from merely trimming its own costs to competing more directly in the AI hardware market.
What this means for Microsoft, Nvidia, and the market
For investors watching the AI hardware space, several markers are worth tracking. First, whether Microsoft finalizes the TSMC volume deal at the reported scale. Second, whether the efficiency target survives real-world deployment in 2027. Third, and perhaps most interesting, is whether any outside customer like Anthropic actually commits to Maia hardware. That would turn a cost-cutting tool into a product, and change how the competitive map of AI chips looks.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.