Breaking
Loading news...

Nvidia's Grip on the AI Labs Is Slipping: Just 39% of Future Deals, JPMorgan Says

Nvidia's grip on AI labs slipping, 39% of future deals

For three years, Nvidia owned the AI chip market so completely that the question was never whether the big AI labs would buy its chips. The question was how many they could get. A new note from JPMorgan suggests that era is ending. Nvidia holds just 39 percent of future compute deals with frontier AI labs, according to the note, down from a market share of more than 80 percent between 2023 and 2025. The labs are not just shopping around. They are signing deals with almost everybody else, and in many cases they are building their own chips instead.

The numbers behind the slide

The figures come from JPMorgan's Eye on the Market note, written by Michael Cembalest and dated 14 November 2026. The note charts future compute agreements from frontier labs, and the total comes to roughly 41 gigawatts. Broadcom custom chips take 24 percent of those deals, Google TPUs take 16 percent, AMD takes 15 percent and Amazon Trainium takes 6 percent. Add the custom chips from Broadcom, Google and Amazon together and they hold 46 percent of future deals, seven points more than Nvidia. The note also points out that Nvidia earned 83 percent of AI chip revenue from 2023 to 2025, with gross margins of 75 percent and operating margins of 66 percent. The shift is not just about market share. It is about a business that has been printing money on near-total dominance, and that dominance is now visibly cracking. The note is Cembalest's own view, and its disclaimer says it is not JPMorgan research. The timing adds to the sting: it landed two days after Jensen Huang resigned as chief executive on 12 November.

Three days to switch

The most eye-catching part of the note is not the chart. It is a single anecdote about how fast the labs can now move away from Nvidia. Cembalest writes that he heard from one frontier lab last week that a year ago it would have taken six to 12 months to get its models running on AMD GPUs. Now, using its own latest AI model to do the engineering, the lab can do it in just three days. The lab is not named, so treat that as one lab's account rather than an industry benchmark. But the direction is the point. Nvidia has long argued that its software ecosystem, CUDA, made switching uneconomic. The note says AMD's MI455X results, which put AMD at the top of an Epoch AI inference index, may have eroded at least part of that advantage. Nvidia's Rubin Ultra series may change the picture when it arrives in 2027, but that is a year away and the labs are signing deals now.

The note also makes the point that the biggest buyers are no longer just buyers. Of the seven largest purchasers of AI accelerators, the five hyperscalers and the two frontier labs, only SpaceX is not trying to build its own chips. When your customers are also your competitors, every future deal is a negotiation with someone who has a credible alternative, and that is exactly what the 39 percent figure shows. The labs are voting with their feet, as Cembalest puts it.

The bill is getting bigger too

There is a second thread in the note that has nothing to do with market share. Nvidia spent 58 percent of its operating cash flow on equity investments and acquisitions over the last three quarters, according to figures cited in the note. Its off-balance-sheet commitments jumped from 214 billion dollars to 422 billion dollars in a single quarter, and to 530 billion dollars once land, power and shell guarantees are included. The note labels those figures as debt and lease liabilities, but Nvidia's own filings call them commitments and guarantees, not debt, and much of the rise is supply commitments the filings say may be cancelable or adjustable before firm orders. The numbers are real and they are large. Whether they turn out to be a smart bet on demand or a pile of obligations that outlives the boom is the question hanging over everything else in the note.

Nvidia is not finished, and nobody in the note says it is. It still holds the largest single share of future deals, it still has the software ecosystem the labs know best, and its next chip line arrives in 2027. But JPMorgan's picture is of a company that went from owning more than four-fifths of the market to holding barely more than a third of the future deals, while its biggest customers learned to build their own hardware and switch suppliers in days instead of months. That is not a collapse. It is a market growing up, and the company that defined the AI boom now has to compete in it like everybody else.

---------------
Author: Ryan Gardner
Silicon Valley News Desk