JPMorgan: NVIDIA's CUDA Moat Narrows as Frontier Lab Compute Procurement Share Drops to 28%
nashnova research
JPMorgan strategist Michael Cembalest reports that Nvidia GPUs account for just 28% of frontier compute contracts, down from 80–90% of AI chip revenue in 2024–2025. CUDA migration barriers have collapsed from months to 3 days — the moat is narrowing on both ends.
What does 28% really tell us?
In compute contracts deploying from 2026 onward, Nvidia GPUs hold just 28% (≈12.1 GW) of roughly 44 GW total. Broadcom custom chips take 23%, AMD 18%, Amazon Trainium 16%, Google TPU 14%, Cerebras 2%.
This means → compared with Nvidia's 80–90% share of AI chip revenue in 2024–2025, the procurement structure among frontier buyers has shifted from one dominant supplier to six competing alternatives.
In plain terms = buying compute used to mean buying Nvidia. Now every major buyer has at least five other cards in hand.
Why are the biggest buyers routing around Nvidia?
Cembalest names seven major accelerator buyers — five hyperscalers and two frontier labs. Only SpaceX has not begun developing its own chips. The rest are all working to avoid Nvidia's 75% gross margin.
This means → the core driver of diversification is not a technology preference — it is cost. A 75% gross margin means for every $1 spent on an Nvidia GPU, $0.75 is Nvidia's profit.
These buyers' compute demand is growing far faster than their budgets. Building in-house or switching to alternatives is a structural cost imperative.
How was the CUDA moat breached?
Cembalest cites a conversation with a frontier lab: a year ago, migrating a model to AMD GPUs took 6–12 months. Now, with the latest AI-assisted engineering tools, the same migration takes 3 days.
In plain terms = CUDA — Nvidia's proprietary programming platform that developers use to write AI software — was the strongest lock-in. The barrier was never the chip itself; it was the cost of rewriting software after switching chips. Now AI rewrites the code, and the lock breaks.
Epoch AI's latest inference index ranks AMD MI455X at roughly 2.9× the GB300 baseline; Nvidia's Vera Rubin VR200 scores about 2.2×. This reflects a phase where AMD leads on inference performance.
Where exactly is Google's TPU applying pressure?
SemiAnalysis inference modeling, benchmarked on Qwen 3.5 397B, shows Google TPU7x outperforms Nvidia GB300 NVL72 on cost-per-compute at every speed tier except the lowest.
This means → Nvidia's challenge is not just "others can do it too" — it is "others already offer better value." For buyers scaling deployment, that directly shifts procurement decisions.
Cembalest concedes Nvidia's Rubin Ultra "may change the picture after a 2027 launch," but that leaves at least a year-plus window for competitors to keep taking share.
Why is China the second front of narrowing?
Cembalest calls China's open-source community "one of Nvidia's core markets" — its models consume the majority of compute on OpenRouter, with capability indices tracking closed-source frontier levels.
The Institute for Progress reports that by end-2027, over a third (36%) of China's AI compute sits inside the country, sourced through both compliant imports and smuggling. "Nvidia is a key supplier, whether or not by choice."
Roughly two-thirds of China's remaining compute comes from renting U.S.-based silicon — another layer of Nvidia dependency. This means → the moat narrows on both ends: frontier labs diversify procurement, compressing the growth side; China faces tightening export controls, compressing the installed base. Whether Nvidia's next CEO can hold the ecosystem together is the key test ahead.
市场有风险,内容仅供研究参考,不构成投资建议。
