Microsoft's Maia 300 targets 300,000 TSMC chips in 2027 scale-up

Microsoft plans to launch its Maia 300 AI accelerator this fall and is negotiating with TSMC for delivery of more than 300,000 units in 2027 — a tenfold jump from the tens of thousands of Maia 200 chips currently in production, according to two people familiar with the matter.

"Microsoft continues to invest in custom chips as part of our long-term AI infrastructure strategy," Andrew Wall, general manager of Azure Maia at Microsoft, said. "We expect Azure Maia deployments to support AI workloads measured in gigawatts."

Maia 200 chips run OpenAI and Microsoft models at 30-40% lower operating costs than Nvidia's flagship accelerators, Microsoft disclosed to investors last month. Maia 300, designed specifically for Microsoft's own models, performs even better, according to people familiar with the matter. The company ultimately wants more than 1 million Maia 300 units but faces supply constraints and ongoing capacity negotiations with TSMC.

The scale-up would narrow the gap with Google, which Morgan Stanley estimates will produce more than 3 million TPUs this year and 5 million next year. Microsoft's Cobalt CPU line has already won adoption from OpenAI and Adobe across more than 25 data centers, providing a proof point for its custom silicon strategy.

Maia 200's Slow Start vs. the Ambition

Maia 200's rollout has lagged expectations. The chip was delayed last year after early testing missed internal targets, and as of late last month it remained deployed in only two data centers in the United States, according to a person familiar with the matter. CEO Satya Nadella said in June that two data centers were operational and that the company planned to expand deployment, including overseas.

The gap with competitors is stark. Google's TPU line has won adoption from multiple large customers, and Google has begun selling TPUs externally. Amazon's Trainium has similarly gained traction. Maia 200's only user is Microsoft itself.

The strategic urgency traces back to a 2022 email from Nadella, later made public through legal proceedings, in which he wrote: "We are just a thin layer on top of Nvidia, and all the intellectual property is in OpenAI's hands." The email referenced a business unit projected to lose $4 billion the following year, driven by the cost of running OpenAI models on Azure.

TSMC Capacity and the Nvidia Dependency

The Maia 300 production plan hinges on TSMC, which is simultaneously ramping 3nm output to 180,000 wafers per month by early Q4 2026 — two to three months ahead of schedule — and accelerating construction of its 1.4nm fab in Taichung. TSMC's capacity is booked through 2026 and into 2027 by Nvidia, AMD, Broadcom, and hyperscaler demand, making Microsoft's 300,000-unit request a competitive allocation battle.

Microsoft's fallback plan is to shift more internal AI workloads to Maia while continuing to rent Nvidia chips to Azure customers at premium prices. The company's MAI-Cyber-1-Flash model, a small in-house specialist, now carries about 90% of Microsoft's vulnerability-scanning workload, cutting its compute bill by roughly half — evidence that in-house silicon and models can displace rented frontier capacity.

Nadella said Microsoft's own AI chips are driving up to 40% efficiency gains. If Maia 300 delivers on that promise at scale, the economics of Azure's AI infrastructure shift materially — potentially saving billions in GPU procurement costs annually.

Microsoft's Maia scale-up, if realized, would reduce the company's dependence on Nvidia's pricing power and improve Azure margins. Nvidia, meanwhile, faces a structural question: as hyperscalers build in-house alternatives, its data center revenue concentration becomes more contested. TSMC stands to benefit regardless, as both Microsoft's Maia and Nvidia's Blackwell and Rubin platforms run on its fabs.

This article is for informational purposes only and does not constitute investment advice.