NVIDIA is confronting structural pressures from two directions, and these forces are just as significant as—perhaps even more consequential than—the departure of founder Jensen Huang.
JPMorgan strategist Michael Cembalest noted in a recent research report titled "NVIDIA and the Next Phase of AI" that frontier AI labs are increasingly diversifying their procurement across multiple suppliers including AMD, Google, and Amazon. NVIDIA's share of post-2026 computing agreements stands at only 28%.
At the same time, the open-source model ecosystem faces a regulatory crackdown from Washington. For investors, this means that the two core growth engines that previously supported NVIDIA's elevated valuation—AI chip pricing power and open-source market exposure—are coming under pressure simultaneously.
Cembalest's conclusion is that the shift in the competitive landscape is "somewhat as important as, or even more important than, management changes."
Frontier Labs Have Already Voted with Their Feet: NVIDIA's Share Drops to 28%
Cembalest tallied all computing agreements announced by major frontier labs for deployment after 2026, totaling approximately 44 gigawatts (GW). Of this, NVIDIA GPUs account for only 12.1 GW, representing a 28% share. The remaining computing capacity breaks down as follows:
Broadcom custom chips: 23%;
AMD: 18%;
Amazon Trainium: 16%;
Google TPU: 14%;
Cerebras: 2%.
This figure stands in stark contrast to NVIDIA's market position between 2024 and 2025, when it captured 80% to 90% of AI chip revenue.
Among the seven major accelerated chip buyers, only SpaceX has yet to advance its own chip development plans. Every other player is finding ways to circumvent NVIDIA's hefty 75% gross margin.
In October 2025, OpenAI signed a multi-gigawatt computing agreement with AMD, and AMD's stock posted its largest single-day gain in nearly a decade that day—an early sign of this trend.
CUDA Moat: Breached by AI in Three Days
NVIDIA's previous core competitive barrier was its CUDA software ecosystem—the engineering cost of migrating away from the hardware platform was high, and the switching cost made it difficult for customers to leave.
But Cembalest, citing conversations with a frontier lab, points out that this advantage is unraveling:
A year ago, completing model migration on AMD GPUs required 6 to 12 months, with high and cumbersome engineering costs. Now, with the help of the latest AI models to assist with engineering work, migrating new models to AMD GPUs takes only 3 days.
On the performance front, Epoch AI's inference benchmarks show that AMD MI455X's composite score is approximately 2.9 times that of NVIDIA GB300, higher than NVIDIA's own Vera Rubin VR200 at about 2.2 times.
Google TPU7x also leads NVIDIA GB300 NVL72 across the board on compute-per-dollar metrics, except in the lowest speed range.
Cembalest acknowledges that NVIDIA's Rubin Ultra "could change this picture when it launches in 2027," but at the current pace of the market, that is a long waiting period.
Open-Source Community: From Core Market to Regulatory Target
As frontier labs gradually diversify their procurement, Cembalest identifies the open-source weights community as "one of NVIDIA's core markets." This group consumes the majority of inference computing power on OpenRouter, and its model capabilities continue to close in on closed-source frontier levels.
SemiAnalysis estimates that 1.5 GW of restricted NVIDIA chips have already flowed into overseas markets, and "NVIDIA is a key link in the relevant supply chain, whether actively or not."
Cembalest also cites data showing that the proportion of NVIDIA's revenue from Singapore addresses rose from 8% to 22% after the Biden administration's export controls, and NVIDIA stopped separately disclosing this data starting in July 2025.
Goldman Sachs analyst Jim Schneider warned last Friday that NVIDIA "may face additional scrutiny in its operations in certain jurisdictions."
Cembalest cites a comment from former Stratechery contributor Andrew Sharp to pinpoint the crux of the issue:
NVIDIA, while enjoying all the benefits of the U.S. market, laws, and customers, does not act like an American company accountable to the U.S. government.
Valuation Is Cheap, But with Caveats
Cembalest is not predicting a collapse.
NVIDIA captured 83% of AI chip revenue between 2023 and 2025, with gross margins as high as 75% and long chip design cycles. He believes that for long-term holders, NVIDIA—trading 32% below its historical high with a PEG of just 0.2x—remains attractive and is the cheapest large-cap AI stock.
However, the 0.2x PEG is built on the assumption that consensus growth expectations will be realized, and both legs supporting that growth are wobbling: frontier labs are accelerating purchases of competitors' chips, while open-source buyers need to apply for licenses to continue procurement.
As Cembalest himself concludes: the shift in the competitive landscape is "somewhat a risk that deserves as much investor caution as management changes."
Comments