JPMorgan Cuts Nvidia’s Share of AI Lab Compute Deals to 28%

J.P. Morgan’s Michael Cembalest has revised the Nvidia chart in his Eye on the Market note. The PDF now served says Nvidia “has only a 28% share of announced frontier lab compute” agreements; the version first posted said “has only a 39% share of future frontier lab compute agreements”. The note still carries its 14 November date; the new PDF was created on 16 November, its metadata shows.

The revised chart gives custom chips from Broadcom, Amazon and Google 53% of those agreements, or 23 GW, against 12.11 GW for Nvidia: 1.9 times as much, by our arithmetic. The note is Cembalest’s own view; its disclaimer says it “in no way constitutes J.P. Morgan Research”.

Business Pill · Switching costs explained

An evergreen Business Engineer explainer on switching costs, why it is hard to leave an AI supplier. It teaches the concept only and does not discuss Nvidia, J.P. Morgan’s note or any company in this piece.

The key insight: As we read the revision, the change is in what the chart counts rather than in the argument: it moved from future agreements with a bare JPMAM source line to announced agreements sourced to lab and accelerator disclosures, Nvidia’s share fell 11 points, and the CUDA passage stayed word for word, by our comparison.

Update (17 Nov): the onboarding figures match a claim Anthropic made to IPO investors. In its roadshow video, Anthropic said Claude Mythos Preview brought up prompt processing for one model on AMD MI355X GPUs in about three days, against six to 12 months by its traditional approach (our piece on the roadshow claim). Tom’s Hardware reported on 16 November that Anthropic co-founder Tom Brown made the comparison. The note quoted below still names no lab, so linking the two is our comparison of the figures.

What Changed in the Note

The first version set Nvidia’s 39% against “its 80%+ market share from 2023-2025”. The revised text reads: “For context, Nvidia had a 80-90% AI chip revenue share in 2024-2025 according to Silicon Analysts”.

The chart title changed from “Frontier Lab future compute agreements” to “Major Frontier Lab announced compute agreements”, with the unit now “GW, announced capacity with deployment 2026 or later”. The source line changed from “Source: JPMAM, 2026” to “Source: Frontier Lab and Accelerator disclosures, JPMAM, 2026”.

The revised bars read: Nvidia GPUs 12.11 GW (28%), Broadcom custom 10 GW (23%), AMD GPUs 8 GW (18%), Amazon Trainium 7 GW (16%), Google TPUs 6 GW (14%) and a new Cerebras bar at 0.75 GW (2%). The labelled bars sum to 43.86 GW, by our arithmetic.

Nvidia’s share is 11 points lower than in the first version. Custom chips rise from 46% to 53%, so their lead over Nvidia widens from 7 points to 25 points, by our arithmetic. The largest rise is Amazon Trainium’s, from 6% to 16%, by our arithmetic.

GPUs from Nvidia and AMD together fall from 54% to 46%, by our arithmetic. Our 15 November piece on the first version reported the 39% figure; it now carries an update.

Shares of frontier labs’ compute agreements by chip supplier in the first and revised versions of Michae
Shares of frontier labs’ compute agreements by chip supplier in the first and revised versions of Michael Cembalest’s Eye on the Market note; the GPU and custom-chip groupings are our arithmetic.
Bar chart: revised J.P. Morgan Eye on the Market chart of announced frontier-lab compute agreements: Nvidia GPUs 12.11 GW (28%), Broadcom custom 10 GW (23%), AMD GPUs 8 GW (18%), Amazon Trainium 7 GW (16%), Google TPUs 6 GW (14%), Cerebras 0.75 GW (2%)
Announced frontier-lab compute agreements with deployment 2026 or later, by chip supplier, from the revised chart in Michael Cembalest’s Eye on the Market note (PDF of 16 November 2026). The 43.86 GW total, the 23 GW custom-chip sum and the 1.9 times ratio are our arithmetic.

The CUDA Passage Did Not Change

Cembalest calls this section “the most important section of this note”. 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”.

It goes on: “Now, using their latest AI model to do the engineering, they can onboard new models on AMD GPUs in just 3 days”. The note sets this against Nvidia’s case that “its superior tokens per watt and superior software ecosystem (CUDA) made switching uneconomic”.

It says AMD’s MI455X results “may have eroded at least part of that advantage”. These passages read the same in both versions of the PDF, by our comparison of the two files.

The note cites an Epoch AI index on which inference performance per chip “actually shows AMD at the top”, and says Nvidia’s Rubin Ultra series “may change that when rolled out in 2027”. On a SemiAnalysis comparison of Nvidia GB300 and Google TPUs running Qwen 3.5 397B, it says: “Across all speeds except the slowest ones, Google’s TPU looks better”.

The software layer outside CUDA has its own moves on our desk: Amazon’s free PyTorch tool and Cerebras’s Tungsten for its wafer-scale chip. A box in the note adds that of the seven largest buyers of accelerators, “only SpaceX is not trying to build their own”.

Where the Note Still Puts Nvidia

The revision leaves the note’s profit figures as they were: Nvidia “earned 83% of AI chip revenues from 2023-2025 with gross margins of 75% and operating margins of 66%”, and trades at a PEG ratio of “just 0.2x”.

Cembalest writes that “the competitive landscape may be intensifying for Nvidia, which in some ways is just as notable a risk for investors as its management change announced this week”. Our TOP500 count shows the GPU split in supercomputers, and our results preview sets out the earnings bar; our market close report covers the stock before results.

The Structural Read

As we read the revised bars, the largest rise is Amazon Trainium’s, up from 6% to 16%, while Broadcom custom and Google TPUs are 1 and 2 points lower at 23% and 14%; custom chips reach 53%, by our arithmetic.

As we read the onboarding account, it remains one unnamed lab’s switching time on AMD GPUs, from 6-12 months to 3 days; the revision adds GW figures to the chart but no new evidence on CUDA.

Michael Cembalest, Eye on the Market, J.P. Morgan (revised PDF, 16 November 2026)

“Based on the bar chart above, the economics (i.e., total cost of ownership normalized to output per GW) may be shifting as frontier labs vote with their feet.”

Three Implications

FOR NVIDIA HOLDERS The revised note gives Nvidia 12.11 GW, 28% of announced frontier-lab agreements, against an 80-90% AI chip revenue share in 2024-2025 that it attributes to Silicon Analysts.

FOR CUSTOM-CHIP MAKERS Broadcom custom 10 GW, Amazon Trainium 7 GW and Google TPUs 6 GW sum to 23 GW, 1.9 times Nvidia’s 12.11 GW, by our arithmetic.

FOR AMD AMD GPUs rise from 15% to 18% of the chart, 8 GW, the only GPU supplier whose share rises between the two versions.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Supercycle.

The framework says: “Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”

As we read it, the revised chart now fills all three categories: Nvidia GPUs at 28%, hyperscaler TPUs and Trainium at 30% and Cerebras at 2%, by our arithmetic, with Broadcom custom chips a further 23%; the framework does not discuss this note.

What Is Not Established

The revised PDF does not flag the change, give a reason for it or list the agreements behind each bar; the landing page still shows the 14 November date. The bar-to-value matching is our reading of the PDF’s text layer, which also gives the 28% the note states for Nvidia.

The 3-day AMD onboarding figure comes from one unnamed lab, as the note relates it; we could not verify it. J.P. Morgan and Nvidia were not contacted for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer framework for the AI supercycle: the stack, the loops back into macro and who captures value at each layer.

Read the Map of AI →

The Bottom Line

The revised Eye on the Market note puts Nvidia at 28% of frontier labs’ announced compute agreements, down from 39% in the version first posted. Custom chips hold 53%, or 23 GW against Nvidia’s 12.11 GW, by our arithmetic, while the CUDA switching passage is unchanged.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The 28% share, the revised chart and the other quotations come from J.P. Morgan’s Eye on the Market note by Michael Cembalest, The Icarus Files, read in full as the public PDF served on 17 November 2026; its file metadata dates it 16 November, while the note’s masthead keeps 14 November. The 39% figures and the first version’s wording, chart title and source line come from the first version of the same PDF, saved by FourWeekMBA on 15 November.

The note says it is not J.P. Morgan Research. A ZeroHedge post that circulated the note was a lead only, not a source. Sums, shares and differences are our arithmetic. J.P. Morgan and Nvidia were not contacted for this piece.

Sources: J.P. Morgan, Eye on the Market: The Icarus Files (PDF, revised 16 Nov 2026) · J.P. Morgan Private Bank: Eye on the Market landing page

Scroll to Top
FourWeekMBA