Key points
- How Maia 300 compares with GPUs
- What Microsoft is ordering for 2027
- Where Nvidia and AMD fit now
Microsoft (MSFT) will scale its Azure Maia chips to "multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services," the company said Monday. Inference is the work of running a trained AI model to answer requests.
Microsoft also introduced Maia 300, its third-generation chip, which it expects to deliver 40% more tokens per megawatt than leading GPU-based systems. It didn't identify the comparison system. CEO Satya Nadella called the announcement "our biggest bet yet on our own silicon." A top Microsoft executive said the company will keep buying chips from Nvidia (NVDA) and AMD (AMD).
Maia 300 is built on TSMC's N2P process
Maia 300 uses eight stacks of HBM4 memory with peak bandwidth of 18 terabytes per second. Its compute chiplets are manufactured by Taiwan Semiconductor Manufacturing (TSM) using the N2P process.
Microsoft says a full 96-accelerator Maia 300 configuration delivers up to 7x the rack-scale token throughput of Maia 200, its current chip. Maia 300 will begin production deployments in 2027.
Maia 200 is deploying across the US, Europe, and Australia. Microsoft says it delivers up to 1.7x more tokens per watt than leading GPU infrastructure for built-in AI features in Teams, Copilot, and Cowork. Maia 200 runs OpenAI models, Microsoft's own MAI models, and open-source models. GitHub Copilot and Microsoft Foundry are due to move onto it by the end of the year.
Microsoft will keep buying Nvidia and AMD chips
Scott Guthrie, Microsoft's executive vice president of Cloud and AI, wrote Monday's announcement. In an interview with Tae Kim's Key Context newsletter published the same day, Guthrie said "we're ordering a lot of Maia 300 chips," and that the supply chain "is going to be seeing this in earnest." He didn't give a date for reaching multi-gigawatt scale. "It'll be a ramp," he said.
Guthrie said Maia 300 will also serve outside developers through Foundry and Microsoft's API service, not just Microsoft's own apps. Maia 400 and Maia 500 will follow, he said.
"We're big partners of NVIDIA and AMD, and we'll continue to be big partners of them," Guthrie said.
Google, Amazon, and Meta also design their own AI chips. Guthrie said Maia's instruction set supports "a broader set of scenarios than say a TPU," Google's in-house chip, and isn't tied to any single model.
As of 10:36 a.m. ET, Microsoft shares were up about 1% at $540.23. Nvidia was down 1.7% at $225.27, and AMD was down 3.3% at $588.10.
Guthrie also said "there's a real price war going on where model costs are coming down." For AI providers, the question is whether lower serving costs and greater paid usage can offset falling token prices.
"The more we can drive down the cost of inference, the more we can pass on savings to our customers," Guthrie said. Microsoft plans to share more at its Ignite conference in San Francisco next month.
