Skip to main content

Tech

Microsoft Corp MSFT View profile

Microsoft bets on Maia 300 to break Nvidia's grip on its AI costs

The third-generation custom accelerator is the most concrete step yet in Microsoft's long campaign to stop buying expensive Nvidia chips it can build itself.

The image features a close-up of a prototype chip labeled 'Maia 300' housed within a sleek metallic frame. It is displayed against a blurred blue background, suggesting a technology-related environment, possibly related to cloud computing.

Microsoft has unveiled Maia 300, its most powerful custom AI chip, as the company steps up efforts to reduce a dependence on Nvidia that its chief executive Satya Nadella has privately described as making Microsoft "a very thin layer on top of Nvidia."

The announcement comes alongside reports that Microsoft is in talks with TSMC to secure more than 300,000 Maia 300 chips for delivery in 2027.

The longer-term ambition is to have more than one million units, a scale that would begin to make custom silicon a genuine structural part of Azure rather than a costly side project.

The economics

Microsoft told investors that running OpenAI and its own models on Maia 200 costs 30% to 40% less than running them on comparable Nvidia hardware.

Maia 300 is designed to push that gap wider, delivering up to seven times the rack-scale token throughput of Maia 200 and more than 2.5 times the performance per watt.

This would see 96 accelerators packed into a single 167-kilowatt rack alongside 18 terabytes per second of HBM4 memory bandwidth.

Pat Moorhead of Moor Insights and Strategy said: "At that scale, custom silicon isn't just about performance. It can become an important margin driver by improving the economics of delivering Microsoft's own AI services."

Closing the gap on rivals

Microsoft lags Google and Amazon, both of which have deployed custom silicon at meaningful scale for several years, and is also trying to persuade large AI customers, including Anthropic, to shift workloads onto Maia rather than the Amazon and Google chips they currently use.

Andrew Wall, Microsoft's general manager for Azure Maia, confirmed the scale of ambition: "We expect our Azure Maia deployments to support AI workload demand measured in gigawatts."

Austin Lyons of chip analysis firm Chipstrat said the architecture is built around the specific demands of agentic AI.

He noted it "attacks that directly with on-die SRAM, 18 TB/s of HBM4 bandwidth, and an Ethernet scale-up fabric that moves KV cache across racks."

Nvidia shares were down 1.6% in early Wall Street trading, suggesting investors have taken the hint, while Microsoft opened 1% higher.

Today’s Edition