Three of the largest buyers of Nvidia (NVDA 5.93%) chips are also three of the companies trying hardest to need fewer of them. Amazon (AMZN 2.93%), Alphabet (GOOG 0.80%)(GOOGL 0.82%), and Microsoft (MSFT 2.55%) each design their own AI processors now, and each is pushing that silicon deeper into the data centers it's building at a breakneck pace.
It's easy to read that as a problem for the company whose graphics processing units (GPUs) have been the default pick for AI work. On Friday, the market seemed to agree that the company faces some risks, with Nvidia shares falling about 6% in a broad semiconductor sell-off. But the companies building their own chips are also buying record amounts of Nvidia's, and that contradiction sits at the center of the story.
Image source: Getty Images.
Amazon
Amazon has the most developed in-house chip story.
The e-commerce and cloud computing company's custom silicon business -- the Graviton processor, the Trainium AI chip, and the Nitro networking chip -- topped a $20 billion annual revenue run rate in the first quarter of 2026.
"If our chips business was a stand-alone business and sold chips produced this year to AWS and other third parties as other leading chip companies do, our annual revenue run rate would be $50 billion," said Amazon CEO Andy Jassy in the company's first-quarter earnings call.
Jassy added that the operation is now one of the top three data center chip businesses in the world.
None of this, however, has slowed Amazon's appetite for Nvidia. The company plans to spend about $200 billion on capital expenditures in 2026 -- much of it on infrastructure that still leans heavily on Nvidia GPUs to serve customers filling Amazon Web Services.
Alphabet
Alphabet has designed its tensor processing units, or TPUs, for more than a decade, and 2026 may be the year that work moves outside its own walls.
Google has since announced eighth-generation TPU systems, as part of a strategy that no longer treats the chips as internal-only. In May, Blackstone announced a joint venture with Google to offer TPUs as a rentable cloud service, with an initial $5 billion commitment and plans to bring 500 megawatts of capacity online in 2027. That followed an agreement to give AI lab Anthropic access to as many as 1 million TPUs, and an earlier reported leasing deal with Meta Platforms.
Now willing to offer access to its custom chips outside its own cloud, this puts it in more direct competition with Nvidia than it used to be. Even so, reports surfaced this week that Google signed a multiyear SpaceX cloud deal involving access to about 110,000 Nvidia GPUs; Alphabet keeps buying even as it builds.
Microsoft
Microsoft may be the furthest behind of the three on custom silicon.
Its effort centers on the Maia accelerator, and the second-generation Maia 200 only recently went live in some data centers to serve some of the work behind Microsoft 365 Copilot and partner OpenAI's models.
Still, the vast majority of AI work inside Microsoft's Azure cloud runs on Nvidia GPUs, so Maia is a way to claw back some of that spending over time, not a wholesale replacement.
The company's spending, meanwhile, is staggering: Microsoft expects to invest roughly $190 billion in capital expenditures during calendar 2026, even as Azure -- where revenue grew 40% in its fiscal third quarter (the period ended March 31, 2026) -- appears capacity-constrained through year-end.
NASDAQ: NVDA
Key Data Points
What it all means for Nvidia
Add it up, and these three, plus Meta Platforms, are on track to spend roughly $725 billion on capital expenditures in 2026 -- up about 77% from last year.
That's the bear case for Nvidia: a growing share of that spending will flow to chips its largest customers design themselves, and they have a significant incentive to do their best to depend less on a single supplier.
The bull case, however, showed up in Nvidia's own latest results. In its fiscal first quarter of 2027 (the period ended April 26, 2026), revenue rose 85% year over year to $81.6 billion, with data center revenue up 92%. And hyperscalers still made up about half of that data center business.
"Demand has gone parabolic," said Nvidia founder and CEO Jensen Huang in the company's fiscal first-quarter earnings call. He also pointed to a fast-growing tier of buyers -- AI start-ups, enterprises, and governments -- that "do not build chips, do not design their own chips."
So, which way does this resolve? Probably both at once, for a while. Custom silicon is real and growing, and it could erode Nvidia's pricing power over time. But the overall pool of AI spending is still expanding fast enough that Nvidia could keep growing even as it loses share at the margin. At a price-to-earnings ratio of about 32 as of this writing, the bigger risk may not be that the in-house chips fail -- it's that they succeed slowly while the market keeps pricing Nvidia for permanent dominance.
