JPMorgan Sees Nvidia’s CUDA Moat Starting to Crack

Of the 44 gigawatts of computing capacity that frontier AI laboratories have contracted for 2026 and beyond, Nvidia accounts for 12.1 gigawatts, or 28%. The tally comes from Michael Cembalest, the JPMorgan strategist, and it is the number investors should start with.

The rest of the book is spread across Broadcom at 23%, AMD at 18%, Amazon’s Trainium at 16% and Google’s TPU at 14%, according to the same analysis. Read together, the five figures describe a market that has already been divided, not one that is still being won.

Those contracts are the order book for 2027 revenue. A supplier holding 28% of committed capacity is not a monopolist, and multiples built on monopoly assumptions tend to compress before the earnings do, because the market reprices the tail of the forecast first. Contract share today is revenue in eighteen months.

CUDA’s grip was never about the chips alone. It was the cost of rewriting millions of lines of code for another instruction set, six to twelve months of engineering that was enough to kill most projects before they started. With AI coding assistants doing the translation, engineers report the same migration now takes about three days.

If that holds, the moat becomes a toll bridge with a bypass beside it. The switching cost that protected Nvidia’s pricing power falls by an order of magnitude, and the customer’s question changes from whether it can afford to move to whether there is any reason to stay.

Epoch AI’s composite testing puts AMD’s MI455X at 2.9 times the score of Nvidia’s GB300. Composite benchmarks are not production workloads, and buyers still weigh networking, software maturity and supply. The direction of the gap matters more than the decimal point, and the direction has been consistent across the last three generations of parts. Nvidia’s lead in raw compute is still real; its lead in the cost of switching is what has narrowed.

The second support for Nvidia’s growth has been demand routed through Asia. Revenue billed to Singapore rose from 8% of the total to 22%, and then the company stopped disclosing the figure. No comparable line has replaced it in the quarterly filing.

A billing address is not a destination, and the company has said as much. But when a disclosure disappears just as it becomes material, analysts treat the gap as a risk rather than a rounding issue, and the burden of proof shifts to management at exactly the wrong moment, when the growth they are defending is the growth the market is paying for.

Washington has also been tightening its approach to open-weight models, the free-to-download systems that let buyers run capable AI on cheaper hardware. Tighter rules would slow the customer base that has been absorbing supply from outside the United States, which is the base the Singapore number describes.

On the screen, Nvidia trades at a price-to-earnings-to-growth ratio of about 0.2. PEG divides the multiple by expected earnings growth, and a reading near 1.0 is the conventional marker of fair value, which is why 0.2 looks like a gift to anyone who trusts the trailing growth rate.

A 0.2 reading says one of two things. Either the market expects growth to collapse, or the growth in the denominator is less durable than the trailing numbers imply. Both readings are arguments about duration rather than about the current quarter, and duration is the one thing a moat is supposed to guarantee.

The bull case rests on arithmetic. 28% of a market expanding quickly still produces a rising absolute number every year, and networking, software and full systems carry margins that merchant chips do not. Customers who buy a rack rather than a card are harder to move than customers who buy a card.

The bear case is that the money does not leave the sector, it rotates. Broadcom and AMD collect the re-rating that Nvidia gives up, an index holder of the whole complex keeps the growth, and an owner of the single stock absorbs the multiple compression on his own.

The 0.2 PEG is therefore not a discount offered by a market that failed to notice. It is a price for a franchise whose two supports, switching costs and offshore demand, are being tested in the same quarter, and prices like that are usually right about something. The question for a holder is whether they are right about the right thing.

The test is measurable. If AMD’s data-center revenue grows faster than Nvidia’s for two consecutive quarters, and Broadcom’s AI guidance keeps stepping up while Nvidia’s contract share stays near 28%, then the low multiple was a warning rather than a bargain. If the reverse holds for another year, the bears have priced a transition that has not started.

Related Posts

  • November 20, 2026
  • 12 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 11 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…