Nvidia (NVDA) sold its most valuable chips to the world’s biggest tech businesses in the AI boom.
What is disconcerting is that some of those same consumers now are paying big to establish alternatives.
Microsoft (MSFT) might be poised to demonstrate just how serious that threat has gotten.
The company plans to unveil its next-generation Maia 300 AI accelerator this fall, potentially as soon as September, according to The Information, as Microsoft pushes to reduce its reliance on Nvidia’s expensive processors.
Reuters reported that Microsoft has discussed manufacturing capacity with Taiwan Semiconductor Manufacturing (TSM) for more than 300,000 Maia 300 chips for delivery in 2027. Longer term, Microsoft reportedly wants capacity for more than 1 million units.
Microsoft disputed the claimed production numbers, saying they don’t reflect the breadth of the company’s effort but underlined bespoke silicon is a key part of its long-term AI infrastructure strategy.
Microsoft says its current Maia 200 accelerator delivers more than 30% better performance per dollar than the latest-generation hardware previously deployed in its fleet.
That distinction is important.
Microsoft doesn’t need Maia to replace Nvidia everywhere.
It just needs its own chips to be good enough, cheap enough, and plentiful enough to transfer a meaningful share of Azure’s AI workloads away from outside providers.
That may be the larger threat for Nvidia investors.
Its biggest clients don’t need to quit buying Nvidia chips.
They merely need to start needing fewer of them.
Microsoft wants to change the economics of its AI spending
Microsoft’s bet on bespoke chips starts with a simple problem.
Artificial intelligence is costly.
With increased demand for Copilot, Azure AI services, or third-party models, Microsoft needs additional computer power to run such products.
Much of that capability has involved purchasing accelerators from firms such as Nvidia.
That deal has helped make Nvidia one of the greatest winners of the AI boom.
Microsoft also had a strong incentive to build more of the silicon itself.
Maia was first introduced by the firm in 2023, followed by the second-generation Maia 200 accelerator in January 2026. Maia 200 is designed for inference, applying trained AI models to respond to cues and complete tasks, not training models.
The Maia 200 is built on TSMC’s 3-nanometer technology, has over 140 billion transistors, and delivers over 10 petaFLOPS of FP4 performance, says Microsoft. More important to shareholders, Microsoft believes it offers about 30% greater performance per dollar than the latest generation of hardware it has previously deployed in its fleet.
That’s the number investors need to watch.
At Microsoft’s scale, even small savings in the cost of processing AI requests can lead to significant savings.
Related: AMD just made bold move to challenge Nvidia
Microsoft data centers in Iowa and Arizona are already running Maia 200. During its fiscal third-quarter results call, Microsoft claimed millions of servers throughout its fleet now employ its own networking, security, and virtualization silicon, in addition to its first-party CPUs and accelerators.
Maia 300 would take that strategy considerably further.
If Microsoft is going to deploy hundreds of thousands of its own AI accelerators, custom silicon is less of a side project.
It’s becoming part of the economics of Azure.
Anthropic could give Maia 300 the validation Microsoft needs
The next stage is more difficult.
Designing an AI chip for Microsoft’s internal workloads is one thing.
Convincing major cloud customers to use it is another.
.Microsoft is encouraging customers like Anthropic to use Maia 300. That might be a big point of validation.
If Anthropic or another frontier AI startup picks Maia for substantial workloads, Microsoft would have evidence that its custom accelerator is competitive, not only within its own infrastructure but also as a product in Azure’s broader cloud ecosystem.
And that’s where the question of Nvidia gets more fascinating.
More Microsoft:
- Morgan Stanley resets Microsoft stock forecast ahead of earnings
- Bank of America doubles down on Microsoft stock ahead of earnings
- Citi revamps Microsoft stock price target for the rest of 2026
Microsoft is already trailing two competitors in the race.
Alphabet has spent years building Tensor Processing Units; Amazon has developed its Trainium family of AI accelerators. Amazon’s Trainium is gaining traction, and Google has started booking income from direct sales of its custom AI chips.
Microsoft has had delays on its other bespoke chip projects, too. A previous next-generation Maia project in 2025 had slipped at least six months because of design changes and staffing issues.
That is a history worth remembering.
Maia 300 will not guarantee success just because Microsoft plans to make it in volume.
But the reported production aspirations reflect just how much more serious the project has grown.
Nvidia’s customers are becoming a new class of competitors
The obvious temptation is to call the Maia 300 a “Nvidia killer.”
That would be overkill.
Nvidia has so much more than the GPU in competitive edge.
Its CUDA software ecosystem, networking solutions, developer partnerships, and huge installed base make it hard for rivals to displace it.
Microsoft keeps using hardware from Nvidia and AMD, as well as its own first-party silicon. In its business results call for the fiscal third quarter, the company described its AI infrastructure strategy as a blend of bespoke silicon and accelerators from third-party providers.
The latter is the incremental threat, not an existential one.
Microsoft doesn’t need to kick Nvidia out of Azure.
Maia has to accomplish enough inference, synthetic data creation, or other tasks to enhance the economics of Azure and lower the number of outside accelerators needed per dollar of AI revenue.
Google can do the same with TPUs. Amazon can do it with Trainium. If all three succeed, Nvidia faces an unusual situation. Its largest customers remain among its largest sources of demand. They are simultaneously designing technology intended to reduce that dependence.
What Nvidia investors should watch
- September:Maia 300 could be unveiled as soon as next month.
- 300,000+:Reported manufacturing capacity Microsoft has discussed with TSMC for 2027.
- 1 million+:Microsoft’s reported longer-term capacity ambition.
- 30%+: Microsoft’s claimed Maia 200 performance-per-dollar improvement over the latest generation previously in its fleet.
- Anthropic: A major Azure customer Microsoft reportedly hopes to persuade to use Maia 300.
- TSMC: The manufacturer Microsoft is reportedly negotiating with for significantly greater capacity.
Microsoft has a lot to prove after all.
The corporation disputes the quoted output figures. Microsoft custom-chip development hasn’t always gone as planned in the past Component supply and TSMC capacity could limit deployment
But the strategic orientation is not hard to see.
The world’s biggest cloud providers are gradually looking to control more of the silicon beneath artificial intelligence.
For Microsoft, success might imply cheaper inference costs, higher margins for Azure, and more control over restricted compute capacity.
The matter is more delicate for Nvidia.
It does not have to lose Microsoft as a customer for Maia 300 to matter.
All Microsoft has to do is shift enough workloads onto its own hardware to modify the number of Nvidia chips needed for every dollar of AI growth.
That’s why the next AI-chip conflict may look so different from the previous.
Perhaps the most threatening competitors for Nvidia are not other chipmakers.
They might be the clients who made Nvidia a powerhouse.
Related: Nvidia dominates AI chips, but BofA sees AMD closing in
