Scoring the Jensen-Dwarkesh debate
Export controls are holding back Chinese AI. Should we keep them?
I’m a little late to the party on this one, but two weeks ago, Dwarkesh Patel had a really excellent episode in which he interviewed Nvidia founder and CEO Jensen Huang:
Zvi Mowshowitz had what I thought was a very good in-depth breakdown and analysis of the discussion, which covered Nvidia’s technology, their business moat, and the question of chip sales to China:
Dwarkesh is rightfully gaining recognition as one of the podcast world’s best interviewers. He’s not an adversarial interviewer like Isaac Chotiner; his goal is not to get you to slip up, or to expose the contradictions in your thinking. Instead, he tries to draw his subjects out and help them explain their worldviews to the audience.
As someone who also prefers this style of interview, I can attest that it’s actually very difficult to pull off. It’s all too easy to slip into doing a softball puff piece — fawning all over your guests and treating them like gurus dispensing wisdom from a mountain. This is an even easier trap for someone like Dwarkesh, who is very young and who is primarily known for interviewing people instead of for dispensing his own thoughts. So it’s extremely impressive that he consistently avoids this trap — he always manages to challenge and provoke his subjects, rather than just letting them spout their usual talking points.
Rarely, though, do we see Dwarkesh actually debate his subjects. In his interview with Jensen, they really get into it on the subject of chip export controls to China. Those export controls — which Trump has significantly loosened — are preventing China from purchasing the best AI chips. Jensen, whose company sells those chips, wants to sell more of them to China. Dwarkesh thinks that’s not a great idea, and pushes back hard.
I actually wrote a post on export controls not too long ago, and I was wondering whether to write another:
But Jensen is one of the premier industrialists of our time, and Dwarkesh really managed to create some interesting dialogue in this interview, so I thought I’d go ahead and score their debate.
Before I get started, though, it’s important to make one distinction. There are actually two types of American semiconductor export controls on China:
Prohibitions on the sale of chipmaking equipment (for example, ASML’s EUV machines) to the Chinese semiconductor manufacturing industry
Prohibitions on the sale of AI chips (for example, Nvidia’s Blackwell chips) to China’s AI industry
There is very little debate about the first of these two types of controls — the controls on chipmaking equipment. There probably are a few people within the Trump administration who would love to sell EUV machines and such to China, but they’re being silenced. The entire debate is about the second type of controls — about whether to sell American-designed AI chips to China. That’s what Dwarkesh and Jensen are arguing about. (In fact, as I’ll talk about in a bit, the stunning success of the equipment controls is the only reason we’re even having a debate about the chip controls in the first place.)
Also, keep in mind that I’m only covering the part of the Jensen-Dwarkesh conversation that’s about export controls. They actually covered more than just that, and for analysis of the other pieces, I recommend Zvi’s breakdown (though be warned, Zvi is very focused on the concept of superintelligence).
So anyway, let’s get to it. Here are the most important points that jumped out at me while watching1 Jensen and Dwarkesh go at it. Overall, I thought Jensen didn’t do very well in this interview — he made a lot of incoherent, self-contradictory arguments, and ignored or waved away some of Dwarkesh’s most important points. He did make some interesting arguments and important points, but didn’t articulate them especially well. Dwarkesh, meanwhile, did a great job pressing Jensen on specific points while also giving him the space to talk.
Jensen’s argument that China already has enough compute is not coherent
Dwarkesh’s main argument for export controls — which is also Dario Amodei’s argument — is that America needs to stay ahead of China in terms of critical security capabilities. Anthropic’s new Mythos model, with its reportedly superior hacking abilities, could represent a powerful weapon. China’s models are improving fast, but if the U.S. maintains an edge, it’ll maintain a military edge as well — one that could help balance out China’s superiority in manufacturing physical weapons like drones.
That’s not necessarily a slam-dunk argument — it relies on a lot of assumptions — but it’s a coherent one. Jensen’s counter to this argument is less coherent. He argues that China already has the compute necessary to train models like Mythos, because they can just use a larger number of older, slower chips:
Mythos was trained on fairly mundane capacity, and a fairly mundane amount of it…The amount of capacity and the type of compute it was trained on is abundantly available in China. So you just have to first realize that chips exist in China…They manufacture 60% of the world’s mainstream chips, maybe more…AI is a parallel computing problem, isn’t it? Why can’t they just put 4x, 10x, as many chips together…If they wanted to, they just gang up more chips, even if they’re 7nm…Huawei just had the largest single year in the history of their company…They have plenty of logic, and they have plenty of HBM2 memory.
I am not an AI researcher, so I can’t evaluate Huang’s claims about being able to train and run AI models just as easily by wiring together older chips as by using newer chips. But if he’s right, it raises a question: Why does Nvidia make so much money in the first place? Nvidia now makes most of its money — now up to $120 billion a year in profit, and growing fast — selling chips for AI models. If AI companies could train and run their models just as easily by wiring together a bunch of dirt-cheap slower older Chinese chips in parallel, why are they shelling out such huge premiums for Nvidia’s chips? And if China has all the compute they need for AI, why would they need Nvidia?
Presumably, Nvidia’s advanced chips confer some sort of very important advantage for AI companies — it’s cheaper and/or faster to train models on Nvidia’s chips. And if that’s true, then having exclusive access to Nvidia’s chips must confer some kind of important advantage for American AI companies over Chinese ones.
In fact, Jensen seems to admit this, when he talks about the heroic lengths Chinese AI researchers have gone to in order to make up for their lack of computing power:
The fact of the matter is, [China’s] AI development is going just fine. The best AI researchers in the world, because they’re limited in compute, they also come up with extremely smart algorithms.
OK, so if China is “limited in compute”, and being forced to invent all of these workarounds, then doesn’t that pretty much invalidate Huang’s earlier argument? For what it’s worth, some of China’s own leading AI companies have also publicly declared that the country’s shortage of compute is holding back their models. Wouldn’t selling China all the compute they want allow them to catch up to American models, thus eliminating the U.S. advantage in cyberwarfare?
Dwarkesh presses Huang on this question, but Jensen never gives him a straight answer.
