Realtime AI News
J.P. Morgan note puts Nvidia at 39% of future frontier-lab compute deals, custom chips at 46%
A J.P. Morgan Asset Management note dated November 14 says Nvidia holds just 39% of frontier AI labs' future compute agreements, down sharply from the 80%-plus share it carried from 2023 to 2025. Broadcom custom chips, Google TPUs and Amazon Trainium together account for 46%, putting custom silicon ahead of Nvidia on paper.
J.P. Morgan Asset Management's Eye on the Market note, dated November 14, puts a striking number in front of anyone who assumes Nvidia still owns the frontier: the company holds just 39% of frontier AI labs' future compute agreements. Michael Cembalest's note contrasts that with the 80%-plus market share Nvidia carried from 2023 through 2025, a gap FourWeekMBA summarized on November 15 as "Nvidia 39%, custom chips 46%."
The chart breaks up the rest of the bar into Broadcom custom chips at 24%, Google TPUs at 16%, AMD GPUs at 15% and Amazon Trainium at 6%. Add Broadcom, Google and Amazon together and custom silicon comes to 46%, seven points more than Nvidia — arithmetic FourWeekMBA performed on the chart rather than a figure the note states outright.
The unit matters as much as the split. The chart is measured in gigawatts, and its single bar reads at roughly 41 GW, so Nvidia's 39% works out to about 16 GW on an approximate chart read. The note also says Nvidia earned 83% of AI chip revenues from 2023 to 2025, at gross margins of 75% and operating margins of 66%.
The explanation offered is supplier diversification. A box beside the chart notes that of the seven large buyers of accelerators — five hyperscalers and two frontier labs — only SpaceX is not trying to build its own silicon.
Read the source for what it is. The note is Cembalest's own view, and its disclaimer says it "in no way constitutes J.P. Morgan Research." The 39% and 46% figures are best treated as a directional signal about procurement, not as an audited share count.
The structural point is the one that matters. For three years a frontier lab's compute budget was close to a single Nvidia revenue line; now the same budget is split across GPUs, TPUs, Trainium and Broadcom ASICs. Buyers get leverage and redundancy, while Nvidia has to keep justifying its premium through interconnect, rack-scale systems and software, because scarcity at the chip level is fading.
What to watch is concrete: how the custom-versus-commercial split moves in the next round of capex guidance, whether Google and Amazon sell TPU and Trainium capacity more aggressively to outside customers, and whether Broadcom's custom ASIC backlog converts into shipped racks. Share shifts show up in analyst charts first and on earnings calls later.
Why it matters
Nvidia remains the default AI compute vendor, but falling below half of future frontier agreements means custom ASICs are now a genuine second leg for the largest labs. That shifts pricing power and pulls the cheaper-compute narrative closer to reality.
Nearby Updates
All11/15, 16:33
OpenAI Disbands Stand-Alone Robotics Team, Betting General Models Can Drive Robots
OpenAI has disbanded its stand-alone robotics team, reassigning its researchers to the company's main research group and its Applied organization, a move the company confirmed to Fortune. OpenAI framed the reorganization as a way to accelerate robotics work rather than retreat from it, reflecting a bet that general-purpose models can serve directly as a robot's brain.
11/15, 15:15
OpenAI launches ChatGPT Wallet for AI agent shopping
OpenAI has launched ChatGPT Wallet, a feature that lets its AI agents handle shopping and payments on the user's behalf inside ChatGPT. The release moves the assistant from recommending products toward completing transactions, putting authorization rules, liability and merchant integration at the center of the agentic commerce debate.
11/15, 15:07
NiCE launches next-generation Cognigy, betting on running agents rather than building them
NiCE has launched the next generation of its Cognigy agent platform, adding a redesigned Agentic Experience Studio, reusable skills, turn-by-turn tracing and native voice-to-voice support. The release folds default speech and language models into per-conversation pricing, framing governance and reliability as the real bottleneck for enterprise agents.
11/15, 14:41
Case filed against Anthropic's Claude over its use of the 'Mythos' name
A case has been filed against Anthropic's Claude over the model's use of the name 'Mythos', according to a report by Indian tech outlet Trak.in. The report does not identify the party behind the filing or the court involved, leaving the dispute centered on name usage rather than model capability or safety.