Microsoft’s Maia 200 and the hyperscalers’ custom-chip push raise the stakes for Nvidia

The big buyers of Nvidia’s AI hardware are also working to buy fewer of the same parts. In parallel with the race to build out power-hungry AI data centers, cloud and AI platform companies including Microsoft, Amazon, and Alphabet are expanding custom silicon programs designed to run more of their own training and inference workloads.

For Microsoft, the urgency is showing up in both spending plans and chip deployment. In its fiscal 2026 third-quarter updates, Microsoft said Azure and other cloud services revenue grew 40% year over year for the quarter ended March 31, 2026. Even with that acceleration, the company told investors it expects to remain capacity-constrained at least through calendar 2026 as it balances incoming supply and works to bring more compute online.

Microsoft also put a hard number on the scale of the buildout. In an earnings conference call, executives said Microsoft expects to invest roughly $190 billion in capital expenditures for calendar year 2026. The same call framed those investments as part of a broader effort to increase the pace of GPU, CPU, and storage capacity additions even as the company expects further constraints.

Chip strategy, in Microsoft’s case, is centered on Maia 200, an in-house AI accelerator designed for use inside Azure. Microsoft said Maia 200 deployment has already begun in select U.S. data centers and that it will soon power key Microsoft and OpenAI systems, including Microsoft 365 Copilot. Microsoft also described Maia 200 as part of a long-term plan to optimize its own AI infrastructure end-to-end, from chips to the services running on them.

Still, Maia is not positioned as an immediate wholesale replacement for third-party GPUs. The overview around the industry shift points to Microsoft using Maia as a way to claw back some portion of spending over time, while acknowledging that much of Azure’s AI work continues to rely on Nvidia GPUs. That distinction matters, because it suggests a transition where custom chips and Nvidia hardware can coexist for a period rather than abruptly displacing one another.

Amazon and Alphabet are pursuing similar aims with different silicon. Amazon has highlighted its custom chip business, including Trainium for AI acceleration, saying its chips business saw nearly 40% quarter-over-quarter growth in Q1 and that its annual revenue run rate is now over $20 billion. Alphabet, meanwhile, has long run its own TPU systems, but is taking steps to make TPU access available beyond its own walls, including through a new compute-as-a-service setup tied to Blackstone’s data center platform.

The effect on Nvidia is likely to be indirect and uneven. Hyperscalers still buy Nvidia in large volume even as they develop alternatives, and Nvidia also faces demand growth from customers that do not build their own chips. What investors will watch next is whether custom silicon ramps fast enough to change purchasing mix at the margin, and whether Nvidia can defend pricing and profitability as the bargaining power of the biggest AI infrastructure buyers increases.

Microsoft, Amazon, and Alphabet have not provided a full breakdown of how quickly internal chips will translate into dollar-level changes in their procurement from Nvidia, nor have they disclosed unit volumes or the share of total AI workloads that will move off third-party GPUs in 2026 and beyond. For now, the public picture is that companies are scaling custom chips while continuing to buy Nvidia, making near-term outcomes for Nvidia more sensitive to execution speed, supply availability, and customer adoption of the in-house platforms.