
Nvidia’s grip on the biggest AI labs is loosening — fast. A new J.P. Morgan note puts Nvidia at just 39% of future frontier-lab compute agreements, a stunning fall from the 80%+ market share it held from 2023 to 2025, as the labs building the world’s most powerful models spread their bets across custom silicon.
The note, Eye on the Market by Michael Cembalest of J.P. Morgan Asset Management and dated November 14, says Nvidia “has only a 39% share of future frontier lab compute agreements compared to its 80%+ market share from 2023-2025.”
Custom chips now add up to more than Nvidia
The rest of the chart is where the story gets uncomfortable for Nvidia. Broadcom’s custom chips take 24%, Google’s TPUs take 16%, AMD’s GPUs take 15% and Amazon’s Trainium takes 6%. Added together, the Broadcom, Google and Amazon custom chips alone hold 46% — more than Nvidia’s entire share.
The note also underlines why this is happening, noting that “Of the seven large users and purchasers of accelerators (the five hyperscalers and the two frontier labs), only SpaceX is not trying to build their own”. In other words, almost every giant buyer is now also a chip designer.
That shift is already visible in the models themselves — read how Meituan’s LongCat-2.5 just went open after training on non-Nvidia hardware.
Switching to AMD now takes 3 days, not a year
The lock-in argument is cracking too. The note relates what it “heard from one of the frontier labs last week”: “a year ago it would have taken 6-12 months to onboard and start running models on AMD GPUs”. Now, it says, “using their latest AI model to do the engineering, they can onboard new models on AMD GPUs in just 3 days”.
Cembalest calls this “the most important section of this note”, and concludes: “the economics (i.e., total cost of ownership normalized to output per GW) may be shifting as frontier labs vote with their feet.”
Why it matters
This lands just days after Jensen Huang’s resignation as CEO on November 12 — the leadership shock that already has more than 117,000 people petitioning for Huang to return to Nvidia. Nvidia still earned, in the note’s words, 83% of AI chip revenues from 2023-2025, and Cembalest still says Nvidia “looks interesting for long term holders” at 32% below its all-time high. But the next wave of deals — roughly 41 gigawatts of future capacity in the chart — is no longer Nvidia’s by default. Rubin Ultra in 2027 will now have to win that share back.
Source: FourWeekMBA / Business Engineer reporting on J.P. Morgan’s Eye on the Market note.
- Nvidia Down to Just 39%? Custom Chips Now Beat It in Future AI Lab Deals - November 15, 2026
- 1.6 Trillion Parameters, Zero Nvidia GPUs: Meituan’s LongCat-2.5 Just Went Open - November 15, 2026
- GTA 6 Has Already Made $1.18 Billion — And It Hasn’t Even Launched Yet - November 15, 2026