The 4.5 trillion dollar elephant in the room
Fear of potential NVIDIA retaliation is shaping important AI policy debates, whether NVIDIA means to or not
For months now, I’ve heard about widespread fear among think tank researchers and policy experts who publish work against NVIDIA’s interests.
NVIDIA is the most valuable company on Earth, the dominant supplier of AI chips, and is currently valued at 4.5 trillion dollars. I worry that fear of its potential retaliation—rumors of attempted firings and media hit pieces—is distorting some of the most important AI policy debates.
Based on what I’ve heard, there is no smoking gun of wrongdoing, but certainly a pattern of fear that impacts policy discussion, justified or not. This fear isn’t just limited to think tank staffers either: I’ve heard many concerning claims that the major AI companies disagree with NVIDIA’s policy stances and know that its talking points are flawed, but are unwilling to speak out because NVIDIA has the power to undermine their business.
To date, these details have mostly dripped out in isolated places online, and so I’ve done the legwork to document and consolidate them to one place, in addition to speaking with a wide range of sources for further substantiation.
NVIDIA, the AI boom, and chip sales to China
NVIDIA’s stock has surged as a consequence of the AI boom, way more than I think people realize.
At ChatGPT’s launch, NVIDIA was already roughly a top 30 company worldwide, valued at roughly $400B. Since ChatGPT’s launch, NVIDIA has grown in value by more than four trillion dollars. This growth is roughly like swallowing the entirety of Microsoft and Apple, circa 2022, which were themselves the two largest companies in the world at that time.
NVIDIA’s remarkable growth faces some threats, however: like the possibility that it might be restricted from selling its advanced AI chips to China.
This is one of today’s most important national security debates: Given that both US political parties want to avoid authoritarian countries surpassing the US’s lead on artificial intelligence, is it better to be willing to sell our leading chips to China, or to hold back from doing so?
The US has flipped back and forth on this policy position throughout 2025, and the outcome is meaningful for NVIDIA’s financials, with billions of dollars of revenue at stake and thus a significant incentive to shape the debate.1
Cards on the table, I tend to favor restrictions on sales to China, whereas NVIDIA does not, though the substantive arguments for and against are somewhat beside the point of this article.
Instead, I want to ask if we are able to have an open debate about these issues without participants fearing undue retaliation.
Think tank researchers fear retaliation for work related to NVIDIA’s China sales
To craft the best policy on export controls, there are many important empirical questions to study.
For instance, how common is illegal smuggling of NVIDIA’s chips under different types of regulatory regimes? How easily can Chinese companies sidestep intended export limitations to train powerful frontier AI systems anyway?
In the United States, a significant amount of this research and analysis happens within nonprofit think tanks. These organizations span a range of ideological affiliations—bipartisan, conservative, etc.—and their work is an important part of crafting good policy.
Unfortunately, dynamics related to NVIDIA’s influence appear to be hindering some research that can help the US to craft better policy.
One prominent case of NVIDIA allegedly going after a think tank researcher occurred back in 2023.
The New York Times reported on concerns that NVIDIA was trying to “sideline” a prominent analyst, Gregory Allen, who had studied the security issues related to China sales and favored more restrictions.
Thankfully, Allen’s boss said he had absolutely no intention of firing him. Another think tank researcher described this account of NVIDIA’s influence as “just the tip of the iceberg.”
It is noteworthy, of course, that CSIS, the bipartisan think tank where Allen works, did say to The New York Times that NVIDIA had not behaved “inappropriately.” The details of these cases are often circumstantial, and in this particular case, much of the relevant information is not public.2
Another prominent think tank researcher—Samuel Hammond, the Chief Economist at the Foundation for American Innovation, a conservative think tank—has been more public about how NVIDIA has reached out to “complain or worse” after his work critical of their China sales. Of course, the exact details of “or worse” are not spelled out here either.
Multiple sources at other institutions, however, described a similar concerning pattern to me: If NVIDIA dislikes your work’s implications, your bosses might hear about it.
In response, researchers try to avoid certain topics that might anger NVIDIA, or, as one source told me, only study export controls if they’re sure their boss is prepared to fully have their back.
Think tank staffers also appear to be concerned about retaliation other than just complaining to their bosses. For instance, last week one NVIDIA critic, Oren Cass, stopped just short of accusing NVIDIA of planting a “hit piece” against him. (Cass is the Chief Economist at the American Compass, a conservative think tank he founded in 2020.)
Cass’s name came up repeatedly in my research, as someone whom NVIDIA attempted to dismiss as a “doomer” on the heels of his analysis, like an August Op-Ed arguing the importance of export controls for certain NVIDIA chips.3
The piece about Cass last week appeared suspect to many on Twitter—though this is of course not strong evidence of NVIDIA’s involvement—with another think tank staffer, Michael Sobolik, calling its claims “laughable,” and another, Eric Sayers, asking “Who drove this? And why?” Matthew Yglesias added in as well that he has “many disagreements with Cass” but that he’s “right that this story is fishy.”4
More specifically, the article reads to me as dismissing Cass with innuendo. It claims in its headline that Cass “faces scrutiny over China corporate ties,” though in my review, I find no “scrutiny” sourced to anyone other than the author.
Instead, it mentions that Cass has expressed concern about NVIDIA’s chip sales to China, but that his think tank also counted Google and BlackRock among a dozen sponsors of a recent gala. According to the piece, these are “companies whose global operations remain deeply intertwined with China’s economy.”5
Another piece of evidence given to “complicate Cass’ image” is that one of 26 members of American Compass’s advisory board is “also a professor at Peking University’s Guanghua School of Management, a leading business institution with close ties to China’s Communist Party.”
Again, many sources pointed this story out to me as being suspicious, but suspicion is quite different from having hard evidence of NVIDIA’s involvement. It is totally possible that Cass is on alert for seeing NVIDIA’s influence even where none is present, though I do agree with commentators that the article seems strange.6
Sources also pointed me toward a separate unreported incident, which they likewise described as suspicious but not firmly tied to NVIDIA.
In September 2025, Janet Egan of the bipartisan think tank CNAS (Center for a New American Security), wrote an article about export controls, arguing that “Selling AI Chips Won’t Keep China Hooked on U.S. Technology.”
One common argument for continuing to sell NVIDIA’s advanced AI chips to China, Egan wrote, is the belief that these sales will keep China hooked on NVIDIA and the US’s AI ecosystem.
These arguments have some merit, she said, but are complicated by the fact that some US AI companies, like Anthropic and Google, have already managed to route around dependence on NVIDIA; buying NVIDIA’s chips is no guarantee of staying dependent on NVIDIA:
The United States’ own AI labs show that reducing reliance on Nvidia’s proprietary ecosystem is not only possible, but already happening. …. Nvidia’s chips are less of an addiction than a gateway drug. There is clear proof that the Nvidia moat is surmountable. If U.S. AI labs with no political mandate to leave the Nvidia stack can accomplish this transition, then Chinese labs facing clear national incentives can likely do the same.
One week after Egan published this piece, she was described in a POLITICO newsletter under the header of “HIGH-TECH CONFLICT”.7
POLITICO claimed that Egan wrote an “op-ed that praised a significant technological shift by artificial intelligence company Anthropic” without disclosing “several close connections her think tank has to the company.” They continued on, describing Egan as “applauding that Anthropic and Google had moved away from relying exclusively on Nvidia’s chips.”
You should draw your own conclusions, but I’ve read Egan’s full article, and I do not see the praise or applause that POLITICO alleges.
Not only do I not agree that Egan had a meaningful conflict of interest here, but if she did, it might even be the opposite conflict of what is alleged: Her employer is in fact partially funded by NVIDIA, which would seem to give Egan a potential reason to support NVIDIA’s continued China sales rather than to explain a flaw in the argument.
But this context is not mentioned in POLITICO’s piece, even though it is publicly available on her think tank’s website, which strikes me as a noteworthy omission.8
To be fair, one of the POLITICO journalists has also broken stories of undisclosed pro-NVIDIA conflicts of interest. The dynamics of media coverage are complicated, and I don’t think a relationship to NVIDIA is especially clear or straightforward.
Still, analysts shouldn’t have to consider at all whether they’ll face retaliation from speaking their policy perspectives publicly, and the pattern of these incidents in aggregate seems concerning.
It is difficult for me to imagine that, given this climate, we are having the most robust policy debates possible.
(These incidents are also only about influence on people employed at think tanks; I won’t discuss allegations of influence on government employees in detail in this piece.)9
NVIDIA’s market power gives it influence over corporate partners
Beyond concerns of think tank staffers feeling reluctant to voice certain policies, NVIDIA has another major means of influence for policy positions: that it is one of the most important AI suppliers in the world, with nearly all AI companies dependent upon it.10
Because there is much more demand for NVIDIA’s chips than there is supply,11 NVIDIA has market power to decide who should receive their scarce computer chips, in what order. Other companies thus have reason to stay in NVIDIA’s good graces.
Of course, NVIDIA has said publicly that it allocates GPUs “fairly”, based on factors such as whether a customer would be capable of deploying more chips were it to receive them.12
Still, considerable reporting, especially from The Information, has dripped out over time that paints a more complicated picture: NVIDIA’s commercial partners appear to fear retaliation from NVIDIA if they take actions contrary to its interests. Based on what I’ve heard, these dynamics extend to an unwillingness to criticize NVIDIA’s policy stances and arguments it advances in its lobbying.
As one example of NVIDIA having large influence, The Information recently reported that both Google and Amazon give NVIDIA’s CEO, Jensen Huang, a personal heads-up before announcing updates to their products intended to compete with NVIDIA:13
At the center of it all is Huang, to whom other leaders in the industry show unusual forms of deference. For example, when Amazon and Google have news to announce about their in-house AI chip efforts—which they’re developing to lessen their dependence on Nvidia— they’ve learned it’s best to first give a heads-up to Huang, say several people involved in these communications.
The cloud computing businesses of Amazon and Google can’t survive without access to Nvidia’s GPUs. And while Huang tolerates their chip efforts, his partners know he doesn’t like surprises about what they’re up to, said those people.
In many industries it would be strange to inform a competitor before revealing your new form of competition, but NVIDIA is not a pure competitor to these companies; it is also an important supplier, who provides them many of its own chips for their large cloud computing businesses. It would be quite harmful to receive fewer chips than wanted or to receive them more slowly, and so the companies need to be careful.
I should be clear that companies acting proactively to avoid angering NVIDIA does not necessarily mean that NVIDIA has done something wrong. For instance, The Information’s reporting on this topic does not allege that NVIDIA made any explicit threats or demands, only that Google and Amazon acted preemptively for fear of angering NVIDIA.
(One friend made a similar point to me about his own company: They worry a ton about a large partner getting angry with them, not because the partner has made any threats or done anything explicitly aggressive, but just because the partner is very powerful and important to their business. The same could be true of NVIDIA.)
The Information also reports that Microsoft has made certain NVIDIA partnership decisions from a fear of retaliation as well, and that NVIDIA is alleged to have made potential customer prioritization dynamics explicit in this case: Specifically, they report that an NVIDIA vice president told Microsoft that they would be lower in line for receiving GPUs if they didn’t also purchase a specially designed server rack, which Microsoft feared would further restrict their choice of suppliers in the future. Reportedly Microsoft’s decision reached the desk of CEO Satya Nadella before NVIDIA ultimately backed off this demand.14
Fears of retaliation extend to NVIDIA’s media partners as well. In May 2025, one of the largest online reviewers of GPUs, with over 2.5 million subscribers on YouTube, accused NVIDIA of “Dirty Manipulation of Reviews.” Specifically, they allege that NVIDIA told them that NVIDIA would have difficulty making spokespeople available to the reviewer unless the reviewer agreed to present certain specific comparisons of NVIDIA’s new chip models, which the reviewer deemed unreasonable.15 NVIDIA, of course, is likely not under an obligation to make its employees available, but I can understand why partners might be wary of running afoul of NVIDIA.
NVIDIA’s influence on AI company policy advocacy
If you speak with people in-the-know at AI companies, fear of retaliation from NVIDIA is a major concern when staking out policy positions. None were willing to be named in this article, for understandable reasons, but many sources familiar, including within multiple frontier AI companies, told me that this fear affects what policies frontier AI companies are willing to advocate for.
Some fear that this trend of self-censorship will get worse as OpenAI and NVIDIA deepen their relationship: Recently, NVIDIA and OpenAI announced the intention for NVIDIA to gain partial ownership of OpenAI by investing up to $100B, which is 20% of OpenAI’s latest reported valuation.
It isn’t yet clear what influence this would give NVIDIA over OpenAI’s decision-making around policy advocacy. OpenAI’s former Senior Advisor for AGI Readiness (and my former boss) Miles Brundage worried on Twitter about the impact on OpenAI’s policy stances of deepening their relationship with NVIDIA:
The technology analyst M.G. Siegler recently observed that OpenAI is so dependent upon NVIDIA that—even when making deals with a rival chip company, as OpenAI did with AMD in recent weeks—they still need to “do the public fealty and show who’s really in charge” by, for instance, mentioning NVIDIA more times than AMD in Sam Altman’s tweet about the deal.16
The only prominent AI company I’m aware of opposing NVIDIA’s position on sales restrictions is Anthropic, which continues to prominently support export controls even despite frosty public fights with NVIDIA on the heels of their support.17
Why is Anthropic comfortable to come out against NVIDIA’s policy position when other companies have not? From the sources I spoke with, the answer is straightforward: Anthropic is more insulated from NVIDIA and so has less to lose from provoking their ire.
More specifically, Anthropic’s AI systems are now run across chips from a few different providers, and Anthropic doesn’t have a large cloud computing business that needs NVIDIA chips for serving third-party customers.18 Without this dependence on NVIDIA, Anthropic is more comfortable to name the dynamic it sees: that selling powerful chips into China will reduce the US’s edge on one of the most important technologies of the future.
As recently as March 2025, companies like OpenAI were openly in favor of continuing strict export controls on advanced AI chips to China, as they documented in their submission to the US’s Office of Science and Technology Policy. The consensus appears to have changed in Washington since this timeframe, however, with the White House now backing NVIDIA’s desire to sell its advanced chips to China.19 Given these positions, expressing a contrary position could prove more costly than is worthwhile, and OpenAI CEO Sam Altman has recently been quoted as saying, “My instinct is that doesn’t work,” about attempts to have fewer powerful chips reach China.20
A wide range of the AI companies would be benefitted by NVIDIA not selling its powerful chips into China: OpenAI doesn’t benefit from Chinese AI companies buying up leading chips that they could otherwise buy, potentially helping them to catch up to the US frontier. Microsoft doesn’t want Chinese companies to build out burgeoning cloud businesses to rival Microsoft’s own Azure. So why aren’t they speaking out as well?
Many of the companies—not just Anthropic—have reason to want NVIDIA not selling to China without restriction, but these interests are outweighed by the 4.5 trillion dollar elephant in the room: that these companies are still hugely dependent upon NVIDIA and can’t afford to risk fallout in their relationship.21
Where do we go from here?
When writing this piece, I heard a wide range of reactions: Some contacts, including people who work in AI but tend to stay focused on their own lane of work, were shocked by the pattern of NVIDIA’s influence that they hadn’t seen laid out in one place. Others felt more that there wasn’t anything too novel here—that of course big businesses are going to lobby and push people hard in pursuit of their interests, and why should one expect anything else.
I also think it’s important that we separate out a few types of questions: Whether conduct is normal, legal, and desirable are of course three different questions.
I have the strongest point-of-view on whether it is desirable for so many in the AI industry to be afraid of retaliation from NVIDIA: No, this is bad.
NVIDIA is an extremely important American company, and their sales—which chips go to which countries—matter a ton. It’s of course fine for people to disagree about the best way to regulate their products, if at all. But we need to be able to have an open and honest debate about the issues without any fear of retaliation.
My hope is that NVIDIA, too, is disheartened to hear that some policy analysts and corporate partners feel this way, and would perhaps take steps to discharge the fear. As I’ve mentioned throughout the piece, I don’t take the examples I cite as firm proof that NVIDIA is intentionally driving some stifling agenda. I think the examples are compatible with many other explanations that have NVIDIA behaving benignly.
Still, I would be interested in learning more about whether NVIDIA has asked any AI companies to adopt a particular stance on export controls. I’d be interested too in understanding more about the dynamics of major companies like Google and Amazon fearing retaliation from NVIDIA: When they have these fears, is this merely their employees acting defensively, or have they heard particular sentiments from NVIDIA that make them cautious?
There’s much more that I hope to understand about how NVIDIA’s influence is affecting various AI policy debates, whether NVIDIA means to have this influence or not. Most of all, I hope that this consolidation of the stories can help to broaden the conversation.
Acknowledgments: The views expressed here are my own and do not imply endorsement by any other party. If any person or institution named in this article would like to provide comment, you may share your institutional email with me here, and I will email you for comment. If you enjoyed the article, please share it around; I’d appreciate it a lot.
NVIDIA’s most recent annual report lists $17B of revenue attributable to China (including Hong Kong but excluding Taiwan), though only a portion of this would be at-risk, depending on where the line is drawn on what chip sales are restricted. Beyond direct sales to the Chinese market, however, export controls might also require NVIDIA to tighten sales to other geographies that could be sources of smuggling chips into China. For more detail on chip smuggling, see “Countering AI Chip Smuggling Has Become a National Security Priority.”
Several other think tank executives also added their own commentary that the piece appeared suspicious, including Zach Graves, who runs the think tank where Hammond is employed. Graves wrote with apparent sarcasm: “American Compass, a famously soft on China think tank. Good journalism here. 👍”. See tweets for Sobolik, Yglesias, and Sayers.
For the article in question, see “Think tank founder faces scrutiny over China corporate ties despite decoupling advocacy”
For two articles exploring the political dynamics of various chip sale deals, see “Anatomy of Two Giant Deals: The U.A.E. Got Chips. The Trump Team Got Crypto Riches.” and “Nvidia Comes Out Swinging as Congress Weighs Limits on China Chip Sales.”
NVIDIA’s market share depends on the exact market segment one is considering, but it is not uncommon to see claims that it commands north of 90% of the datacenter GPU market.
It is a common refrain at this point for OpenAI executives to tweet about being out of GPUs, for instance. Limited GPUs are a bottleneck both for how many customers a company can serve with its products, as well as how many researchers they can hire because of the need to apportion each researcher enough GPUs that they can productively do research.
For discussion of NVIDIA handling allocation “fairly,” see Kylie Robison’s coverage here. NVIDIA could alternately just increase the price of its GPUs until demand equals supply, but this doesn’t seem to be its approach.
On the Big Technology Podcast, Siegler said, “Right now, if you want to be a main player in this game, you’ve obviously got to work with NVIDIA. And as everyone knows, while they say they don’t play favorites, if they’re gonna invest … $100B into you, you’re going to get access to their chips, like they’re not going to screw you over. But, if you did a deal with AMD behind their back, maybe you’re not first in line for those chips anymore, and so I think you … have to do the public fealty and show who’s really in charge.”
OpenAI’s complete corporate announcement does mention AMD many times and NVIDIA none at all. Additionally, Altman said on Ben Thompson’s Stratechery that he didn’t know whether NVIDIA’s Huang was aware of the deal’s details before it was announced. More recently, OpenAI has also announced efforts that will involve “designing its own chips and systems,” which did not refer to NVIDIA.
After Anthropic released a pro-export control report back in late April 2025, NVIDIA came out with statements implying that Anthropic was “tell[ing] tall tales”, and saying that NVIDIA CEO Jensen Huang disagreed with Anthropic CEO Dario Amodei on “almost everything.”
For one brief overview of the White House’s change in position, see “Nvidia and AMD to pay 15% of China chip sales to US”
It would be quite useful for someone to assemble a full timeline of AI companies’ public statements about export controls, mapped against the reporting of NVIDIA’s relationship-building with the Trump administration. CNBC has good contemporaneous reporting on this topic from August.
As an example of how these companies might otherwise weigh in, the large AI companies employ many of the technical experts with the ability to speak with authority on export controls’ efficacy—how plausible is it to implement certain forms of verification on chips, how plausible is it to stop various forms of export control circumvention. But given fear of relationship fallout, I do not expect these experts to speak freely if it were to contradict NVIDIA.
Wow, the part about major AI companies being unwilling to speak out realley struck a chord. Your observation that 'NVIDIA has the power to undermine their business' is so important. This kind of diligent documentation is crucial for healthy AI policy debates. Thanks for shedding light on this.