Scaling Laws: Export Controls: Janet Egan, Sam Winter-Levy, and Peter Harrell on the White House's Semiconductor Decision

[Intro]

Kevin Frazier: It is the Lawfare Podcast. I'm Kevin Frazier, the AI Innovation and Law Fellow at the University of Texas School of Law, and a senior editor at Lawfare. Today we're bringing you something a little different. It's an episode from our new podcast series, Scaling Laws. Scaling Laws is a creation of Lawfare and Texas law. It has a pretty simple aim, but a huge mission. We cover the most important AI and law policy questions that are top of mind for everyone from Sam Altman to Senators on The Hill, to folks like you, we dive deep into the weeds of new laws, various proposals, and what the labs are up to, to make sure you're up to date on the rules and regulations, standards, and ideas that are shaping the future of this pivotal technology.

If that sounds like something you're gonna be interested in and are hunches, it is. You can find scaling laws wherever you subscribe to podcasts. You can also follow us on X and bluesky. Thank you.

Alan Rozenshtein: When the AI overlords takeover, what are you most excited about?

Kevin Frazier: It's, it's not crazy. It's just smart.

Alan Rozenshtein: And just this year, in the first six months, there have been something like a thousand laws.

Kevin Frazier: Who's actually building the scaffolding around how it's gonna work, how everyday folks are gonna use it.

Alan Rozenshtein: AI only works if society lets it work.

Kevin Frazier: There are so many questions have to be figured out and nobody came to my bonus class. Let's enforce the rules of the road.

Alan Rozenshtein: Welcome to Scaling Laws, a podcast from Lawfare and the University of Texas School of Law that explores the intersection of AI law and policy. I'm Alan Rozenshtein, associate professor of Law at the University of Minnesota, and research director at Lawfare. Today I'm talking to Sam Winter Levy, a fellow at the Carnegie Endowment for International Peace, Janet Egan, a senior fellow at the Center for New American Security, and Peter Harrell, a non-resident fellow at Carnegie and a former Biden White House official. About the Trump administration's unprecedented deal to allow Nvidia to sell advanced AI chips to China in exchange for a 15% cut of the revenue.

We discussed the deal strategic implications for the US-China Tech race, the potential damage to America's international partnerships, and the serious legal and constitutional questions raised by what might amount to a direct tax on exports.

[Main Podcast]

Sam Winter-Levy, Janet Egan and Peter Harrell. Thanks for joining the podcast.

Thanks for having me. So we have a lot to get through. But before we get into the specifics of the Nvidia export deal, I wanna step back and provide a little background for, for the audience on what export controls are and, and what this regime of export controls has been in particular with respect to semiconductors and, and AI chips you know, through the Biden administration and the early Trump administration. So Sam kind of set the stage and then get us right up to what happened last week, and then I'll turn it to, to, to Janet for that.

Sam Winter-Levy:  Sure. Yeah, so there's a, there's a lot of backstory here. So starting in the first Trump administration and then really kind of escalating throughout the Biden administration, the U.S. government introduces a kind of series of export controls.

On some of the kind of core inputs for advanced AI systems today. So in particular, semiconductor manufacturing equipment, and then in October 2022 during the Biden administration, advanced AI chips which are the kind of the key computing power that you need to run these kinda increasingly powerful, increasingly popular AI systems because the Biden administration can view these systems as likely to have kind of major national security implications.

When the Biden administration launched these controls in October 2022, at the time, the best chip you could buy were these kind of very advanced Nvidia chips. And at the time it was the Nvidia A100. Now the kinda most advanced Nvidia chip is the H100. These are the kind of really powerful chips that you train charge GBT and other sorts of you know, AI models on, in response to that, Nvidia wanted to produce a chip that was legal to sell in China. And so they produced a kind of degraded version of these chips called the one, one of which was the H20, which is the chip that I think we're gonna go on to discuss in a, in a little bit. And just to give you a little bit of context on what, what the H20 is, it's not the kind of most cutting edge chip that remains banned for now.

It's a chip that. Is less good for training powerful AI systems on, but it's chip that's still very, very good for running an AI system for inference. So if you think about the difference between kind of training GPT-4 0.5 GPT five, and running it, when you actually ask it a query, the H 20 chip is very good for that kind of second stage of of the process.

So that ship was allowed to be sold throughout the Biden administration and during the kind of last month of the Biden administration, they released a series of new export controls that kind of tightened, cracked down on various elements of the, the semiconductor supply chain. But the H 20 remained unbound.

You were still allowed to sell the H 20 Nvidia was still allowed to sell the H 20 to China. This was a kind of topic of fierce internal debate within the Biden administration. They probably would've banned the H 20 had they had a few more you know, had, they had a, hypothetically another year in office.

But when they left office, the H 20 was still legal for Nvidia to sell in China.

Alan Rozenshtein: And, and let me, let me jump in there before we then get to what the Trump administration did, what was the effect of this? Presumably the idea was that it was gonna hold our competitors, in particular China back did it in fact do that?

I mean, I remember when Deep Seek R one came out and everyone had a complete freak out. And, and there were some concerns whether export controls really were effective. So just, just, just talk about whether this was in fact all that effective in the stated goal.

Sam Winter-Levy: Sure. Yeah. I think just kind of to take one step back, just, you know, before, before answering that directly, I think one thing to the central debate in this whole export control conversation to understand is.

The overriding premise when the Biden administration first introduces export controls is that they're trying to thread this needle. On the one hand, they want to hurt. They want to kind of cripple China's advanced AI industry because they think that industry has kind of key national security implications.

So they don't wanna sell China the most advanced AI chips. On the other hand, they also want to kind minimize the incentive. For Chinese companies to develop their own alternatives to the U.S. kinda semiconductor ecosystem. So they want to allow to continue to allow us companies to sell all the less powerful, less capable chips to China to try to kind of thread this needle between hurting China's AI companies but not accelerating the indigenization of the semiconductor stack by, know, by helping Chinese, you know, helping Chinese domestic semiconductor industry. And throughout the Biden administration, there's this kind of raging debate of how you thread that needle.

Alan Rozenshtein: And that kind is, is that needle thread? I mean, I, I, I guess the, the reason I, I, I ask and, and Janet and Peter free, free to jump in as well here.

You know, China's a very large country. It's very sophisticated, it has ambitions right. You know, this is, this is not North Korea. I know they, they can just do things. And presumably if you are keeping the most advanced, the most cutting edge. Chips from them, won't that incentivize them to try to develop those chips, right.

Meet TSMC at the cutting edge or, you know, re replicate the A SML lithography machines. You know what, whatever the, the, the cutting edge bottleneck is whether or not you give them a bunch of b rated chips to, I don't know, run toasters on.

Sam Winter-Levy:  Sure I, we can get into this. You know, least, I think it's kind of the, one of the central kind of cruxes in this debate.

I think critics of the decision to allow the US to sell the H 20 would say China is already all in on indigenizing the, the semiconductor supply chain. They're already all in on this. They're, they're not gonna be, they're not gonna spend less effort on this just because they get access to a few more Nvidia chips.

They're already committed. They're spending hundreds of billions of dollars a year. They're pumping their best talent into trying to kind of, figure out how to do this themselves. That would be the argument, you know, against that would be the argument that would suggest that China is not, China's incentive is not gonna change based on selling additional chips to them.

Just to kind of go back to, to your original question about whether the controls have actually worked, I mean, I think the evidence there suggests that China really is. Compute constrained. They really are struggling to get access to the scale of advanced chips that they need. So according to the administration itself, Huawei, at the moment, the kinda main Chinese company in this domain is producing, is likely to be able to produce 200,000 chips a year.

For context of video producing 10 million, 50 million. Significantly more chips. There are, every single Chinese AI company is pretty much every single Chinese AI company is on record as saying the main thing holding back their progress is access to advanced chips at scale. US companies also need access to more compute.

That's like the key bottleneck for a lot of these, for a lot of these players. But I think a lot of the evidence suggests that China really is struggling to get access to advanced computing power, and that is probably the kind of single most important. Obstacle to their AI ambitions. They have the talent.

They have the energy. They have plenty of capital. The thing they don't have right now is advanced AI chips, which is why this debate has become so heated because for critics of the administration's decision, we are now kind of giving them, we are now removing one of the kind of key obstacles that are holding back China's AI companies.

Alan Rozenshtein: Okay. But before we get to Jenna, why don't you just sort of finish the, the chronology here. So, so we have the Biden administration. They do these controls. There's this diffusion rule, which goes quite beyond the China export bans, and it bans chips to a whole variety of nations. And the issue limits them many to our allies because of concerns about, you know, re-export and things of that nature.

So Trump comes in. And what is the overall tenor bin of the Trump export control approach to AI hardware?

Sam Winter-Levy:  Yeah, so I say here there are kind of. Mixed signals, I think is, is one way you could put it. So the Biden administration has not banned the H 20 when they leave office. They are kind of discussing it, they're thinking about it, but they, they, they have not followed through on it.

They have introduced a bunch of measures that kind of crack down on chip smuggling. They try to, to try to enforce the export control regime more intensely. The Trump administration enters office talks a lot about how they kind of agree with many elements of this, many of which had started in the first Trump administration, cracking down on chip smuggling, et cetera.

They also have a rollback, the diffusion rule, which was one of the Biden administration's kind of big moves to try to crack down on chip smuggling. And then on the H 20 in, I think it's April they the reports emerge that millions of H 20 chips are likely, are likely going to be sold to China, which is kinda a really significant number.

At the same time, Trump is escalating this kind of trade war with China and in response to Chinese, retaliation to those tariffs, to the trade war on critical minerals and right Earth, the Trump administration then sends a letter to Nvidia saying, if you want to export the H 20 chip, you now need a license, which was not the case previously.

And implicitly saying that that license would be denied. So it's adding a whole a, a new, a new set of controls on the H 20 chip that was previously legal to sell. And then I think, you know, Janet can get more into the details of this, but last month the Trump administration then says. We will now, you will now be allowed to sell the H 20.

Those licenses will be granted. H 20 can flow to China and there's various kinda conflicting accounts over whether this was wrapped up in the kind of broader trade discussions or whether this was just a decision by the US government that when we, when it, when it, when it thinks about threading the needle over where to draw the line in these kind of chip controls between.

Cracking down on China's AI companies versus not helping their semiconductor companies by incentivizing increased purchases of Chinese domestic semiconductors where, you know, whether the US government just decided that a different way to thread that needle would be by allowing the, the sales of the H 20.

And we can get into the exact timeline of, you know, which of those accounts seems more, more plausible.

Alan Rozenshtein: So Janet, let's get into that. Let's get into that timeline. So the Financial Times reported last week that Nvidia and a MD will now be able to export a, a whole range of advanced AI semiconductors to China as long as they pay the US government.

15%.

Janet Egan: Yeah. It's a very unusual, unprecedented deal, I'd say. So essentially,

Alan Rozenshtein: which, which, which, which is a very polite way of saying it's nuts, right? I mean, this is. Cuckoo pants, right? This me, this is the Lawfare podcast. I'm, I'm gonna be loose. My jaw dropped when I read this reporting.

Janet Egan: It was really interesting to me to see President Trump own up directly to his negotiations with Jensen on this to say that, oh, I asked for 20%.

We settled on 15% and, hey, now, and, and

Alan Rozenshtein: Jensen, this is Jensen Wong, the CEO of Nvidia.

Janet Egan: That's right. Famous leather

Alan Rozenshtein: jacket wearing globe trotting, CEO of Nvidia.

Janet Egan: And I think what even struck me more was then flagging that there's even discussions open about then exporting the latest generation of Chip Blackwell and Unenhanced version as well.

So just since April to come to August and have this complete roundabout face to change the strategy seems quite remarkable. I think there's three key reasons. This deal seems like a pretty big deal to me. And the first Sam's touched on this, but. It seems to erode American AI dominance just at a time when you want to preserve the strongest possible lead over China.

So a common misconception is that this is an old chip. This is not a good chip. The H 20 is one that China does not have. And cannot make indigenously. And having more and better chips is the key differentiator that keeps the US ahead of Chinese AI development.

Alan Rozenshtein: And, and sorry, just to clarify, when you say that the HH 20 China does not have it, you mean, you mean like China has whatever it has gotten right?

Mm-hmm. Under the Biden administration before the Trump administration clamped down on it, but they, they cannot, as of now, purchase more or build something with equivalent capabilities.

Janet Egan: That's right. There's other export controls on some of the key components that go inside that chip, which is the high bandwidth memory.

So China can't import the key elements that it would need to make these chips. It can't make those key elements domestically. And so now they're able to import these H 20 from the US in exchange for revenue. And I just wanna, I just wanna

Alan Rozenshtein: stay on that point for a second because I, I think sometimes I think may be helpful, and I'd love for you to get into this, some of the details on what makes these advanced chips so, so advanced.

And my understanding is that. Obviously chip design is very complicated. So I'm oversimplifying immensely here. But, but then you can think of the capabilities across a number of dimensions, one of which is clock speed, right? So, these chips have some thousands of numbers of cores, each of which can do mathematical operations, you know, some millions of times per per second or, or billions of times per second.

But there's also a huge amount of data that has to transfer between different parts of these chips and this memory bandwidth. Is as important of a bottleneck and is as difficult for China to reproduce in its own cutting edge chips as, as the, as the as the clock speed. And, and, and so I, I, someone please correct me, my understanding was that, you know, the, the, the, the chips that, for example, deeps seq was training on, were they slower but with more memory bandwidth, or they were faster but with less memory bandwidth and the, you know, the, the, the, the geniuses at deep seek managed to, to, to deal with whatever problem that they had by, you know, working on the other thing.

Janet Egan: I think it's generally understood that Deep Seq was trained on H one hundreds that were provided to China before export controls really bit, and or through smuggling as well. But so those ones are the ones where they're optimizing for computational performance. So clock speed as you put it. What the HBM does differently is it, so Nvidia made them to fall just below the threshold when it came to, you know, the computational performance, but then they maxed it out on this higher bandwidth memory and why that matters.

So since about 20, since the 2023 export controls came in there was an emergence of a new AI learning paradigm. So it's called reasoning models or test time compute. And what this does is it uses a bunch more of computational power not in the training phase, but at the time of deployment. So this is after you've trained a model and, and you've enhanced it.

When you are actually using it, you're giving the model time to think and you're allowing it to connect a greater degree of information across longer time horizons. Now, high bandwidth memory is key for this, and so we're providing, it looks like we're going to be be providing China chips that allow them to deploy models.

Better. And that allows models to think more cleverly, but it also allows you to generate more synthetic data for the next training run. And it allows you to run more experiments on how to optimize the design for the next training run. And so these are key things that are gonna help accelerate China's progress at Frontier AI models as well, as well as just deploying it more across society.

Alan Rozenshtein: So, so you, you, you, so you, you mentioned a a number of reasons. You thought this was a big deal. So obviously you have the capabilities. What are some other reasons that you mind?

Janet Egan: Yeah. The second one, it challenges the partnership with our allies. So for the us, the US cannot successfully impose export controls alone.

All of the controls to date on semiconductor manufacturing equipment have been done in really close partnership with countries like the Netherlands, Japan South Korea to some extent. And that's because a lot of these components and products aren't actually made in the US but they're made overseas via partners, so we can put controls on US chips.

Sure. But that's pretty redundant. If the Netherlands sell their extreme ultraviolet light machines to China, and then China just makes it own. And so to date, these countries have been bound together by in-depth conversations around the strategic risk and what's at stake for national security. And now it seems there's a really strong shift in the US government.

They've said allies and partners impose an economic costs on your businesses to not supply China. And now the US is saying, Hey China, we'll give you some, but we're gonna take some additional revenue for us. And so I can't imagine this is gonna go down well in this sort of semiconductor supply chain alliance.

And I'm concerned about what this means for the longevity of export controls. I.

Alan Rozenshtein: So I'm actually, I'm actually, can I wanna follow up on that? So, so as our listeners may may notice your lovely Australian accent and you of course worked in, in the Australian government for a number of years working on sort of these really important tech policy issues.

I'm just curious from the perspective you know, not just of. Countries, for example, that, you know, make some of these bottleneck components. You know, the, the Netherlands, Germany does optics. Obviously Japan and South Korea have a lot of semiconductor manufacturing. But a country like Australia, and perhaps it's just my ignorance, but my, I my understanding is, is not an impor is not a key part of the, let's say, like semiconductor supply chain for advanced AI chips, but is obviously a super important partner for the us you know, a five eyes.

Member important intelligence partner. Like if, if you're, if you're sitting in Canberra and, and you're reading the Financial Times and you know, Jensen and Donald have come over your coffee, you know, Jensen and Donald, you're reading over your flat white Right. Have come to some agreement. Like, what are you thinking in terms of the US as, as a, as a strategic diplomatic partner?

Janet Egan: I think across the world it's raising questions as to hang on. If we're asked to be tough on China, or if we're asked to take strong steps to, in our relationships with China that come at economic cost to us, are we, can we be guaranteed that the US is gonna be consistent and by our side in those decisions?

And so I think countries around the world are probably asking that question as to, well, do we need to be hedging against the risk that there might be a grand deal struck and suddenly, you know, it'll be us and China collaborating on a lot of things that we've cut ourselves off from China from.

And so I think these are they're probably conversations happening around capitals, around the world. And then

Alan Rozenshtein: just to finish up, what, why, what is the last reason you think this is such a big deal?

Janet Egan: I think the use of an export tax is pretty unprecedented. And there's been, I won't go deeply on this, I think we've got other people on the call who can speak much more clearly on this, but I have heard questions as to its legality and whether it's in violation of export control legislation.

And so how durable this will be is an interesting question.

Alan Rozenshtein: Alright, so there's a great tee up for the, that was a great tee up for for the legal questions. So lemme, lemme turn to you. Lemme turn to you, Peter. Is any of this, is any of this legal? Let, let's start with the statutory issues and, and then we will, we'll talk about the constitutional ones somewhat separately.

Peter Harrel: Well, I think the question of is any of this legal is gonna depend first and foremost on what this deal actually is, right? We've seen press reporting and we've seen the president say that he's gonna charge or that there is an agreement to charge a 15% kind of rev share fee if you will, of Nvidia H 20 sales to China.

But, you know, we don't actually know how this is. Structured. Is this actually like a condition of the license? Is this maybe actually not a payment by Nvidia to the US government, but is Nvidia now gonna announce it's dedicating 15% of sales to, you know, building factories named after the president, you know, and 30 new American states?

Like, we don't actually know what exactly Nvidia has promised here. And I keep that in mind because, you know, I do a lot of work on, on trade policy and we've seen trade deals. Or sort of trade term sheets over the last couple of months with the EU and Japan, where details seem very hazy and where you actually see quite different views between what Washington says is gonna happen and what a foreign capital is saying.

So, so I, I, I think we don't actually know what, what this deal actually comprises of as of as of today. And I, I just make that as a caveat, but let, but let's, let's posit what the press at least seems to suggest, which is that Nvidia. Is being charged 15% of revenue. Don't know whether that's gross, don't know whether that's net, you know, but 15% of something related on their their export of chips to to to, to China.

I think that it is unlikely that if that is a condition of the license or related to the license, it's legal. There's actually a specific statutory. Provision codified at 50 USC, 48, 15 C, which is included in the 2018 Export Control Reform Act, kind of the current legislative basis for us export controls that explicitly prohibits the Commerce Department from demanding or collecting a fee related to processing or issuing a license or.

To, you know, any other activity related to the export control regulation. So it's very hard for me to see, you know, if this is in fact a fee that was demanded of Nvidia as a condition of getting a license, how to square that with the statutory provision. Now, of course, it's a question who can actually sue here?

I'm, I'm assuming that Nvidia, even if this is clearly illegal as a statutory matter. I'm assuming that Nvidia, you know, at least seeing Jensen, having just agreed to this probably isn't gonna turn around and sue over it. And I kind of that, I mean, I kind of get that from his perspective, right? 'cause he is looking at this as a choice of I can either sell these chips in China and pocket 85% of the revenue, or I can not sell the chips in China and get zero China revenue.

So I, I get why he made the decision. He is, and sometimes it's not his interest

Alan Rozenshtein: to sue. And presumably Nvidia can also just jack up the 15% right? Like presumably Yeah, they probably think they redo it on. I mean, but also as,

Peter Harrel: as, as, as Sam, and, I mean, yes. A, they probably think they can reco it from their customers and B as as Sam and Janet said, I mean these are actually 3-year-old chips.

Right. Actually, the cost of production of these things is they're marginal cost for a unit of chip here. Given these are highly, you know, fully depreciated is quite low, so even if they eat the 15%, their margin on these chips is gonna be very high. Janet, do you wanna jump in?

Janet Egan: I think it's, it's interesting to think about NVIDIA's approach to date in terms of how much they're marking down their chips to China compared to what they're selling in the US because I've heard, I've read it in some reporting that they've actually made significant discounts on chips to China.

Already, and it seems like they're looking to capture more market share there than they are to like price up or include the costs in the chip themselves. I also think that as we move to a world with more inference rather than just training AI models, so in the us it makes sense to be purchasing the very latest.

Cut of the chip, like the blackwells because you're using them for training before they then turn into use for inference. But if you're starting to build out inference only data centers around the world, it could be that the H 20 is actually the most price competitive and best in the market now.

Alan Rozenshtein: And, and from a, from a balance of, of, of kind of AI power, how big of a deal is that?

And, and, and the, the, the kind of point I'm trying to get at is I, I think for a lot of folks what's, what's most important if you're trying to kind of keep score and who is quote unquote winning? Right. And we should. We could have a whole conversation about whether that eve is even a coherent way of talking about this, but to the extent that we're gonna buy into the sort of US China race, right?

You know, you, you really think about who has the biggest models, who has the smartest models, who got another, you know, 1% on this benchmark, you know, who got another 2% on that benchmark? And that's all about, or that's mostly, at least about, in addition to architectural advances just kind of raw, raw compute.

But if you want to actually deploy these. That's actually inference. And, and so, whether we're thinking now or 2, 3, 4 years from now what is your sense of the relative importance of inference compute versus training compute when it comes to, you know, the question that a normie, you know, might care about, which is who is, who is again, quote unquote winning China or the United States?

Janet Egan: Yeah, just a small question today. I love it.

Alan Rozenshtein: Welcome to the LA podcast. What are the scaling laws?

Janet Egan: I think, I think inference will continue to play a really key role here and like to take a step back when we talk about this AI race, I mean, it means different things to different people, but looking at the AI action plan that came out from the Trump administration part of how they define that was exporting us technology around the world and having the world.

Be based on US models and so. I think what's what's really relevant here is like it's not just around building out the best models. It's about diffusing them around the world. And if you go and talk to other capitals or folks in other countries, they're less a GI pilled than in the us. They're not always just focused on, wow, what's the absolute frontier of capability they're asking for?

Drug discovery, they're asking for use in medicine. They're asking for scientific advancement, and that's where the inference will really play a role. So those general purpose models trained up that, that training compute will still be so important and only a few companies will be able to afford compute at the threshold to train that at the frontier.

But I think the inference is a really key part of diffusion.

Sam Winter-Levy:  Yeah, I, I mean, I think Peter should, should, should you know, weigh in more on the, on the legal details of this, but I think it's worth just putting some. Numbers on the sorts of revenue gains that the US government might be getting out of this.

Just to give you some sense of, you know, the merits of this decision relative to, you know, the cost that Janet was just talking about in terms of, you know, the, the amount of compute that that, you know, a US competitor might be getting access to here. So Nvidia, when they originally had to write, when, when the H 20 was originally banned by the Trump administration.

Earlier this year, Nvidia had to write off a bunch of revenue that they would've made from this, and I think it was something on the order of 50, 15 or $16 billion that they had to kind of write off. If you think about kind of 15% of that, that's something like. $600 million in revenue, the US government would be captured.

I mean, again, it depends on how it's structured. We don't know a lot of the details of this, but this is something on the order of hundreds of millions of dollars likely. If you think about the kind of, so that's in the, in the kind of benefit side of the thing, in the benefit side of the ledger in terms of the kind of costs, in terms of handing over kind of millions of chips that are gonna be crucial for kind of deploying these models that may well have kind of central national security implications.

I think a lot of the reaction from a kind of a lot of national security professionals has been that. The co, you know, maybe there is some dollar value that you could put on this that would, that, that might make it make sense. But it's gonna be a lot more than a few hundred million dollars in, in US revenue, the, the, the, the US government is gonna capture from this deal.

So just, just having some of those kinda numbers in mind might help kinda assess the kind of merits of, of a decision like this from the administration.

Peter Harrel: Let me let me jump in on, on that point. You know, it, it seems to me it is basically impossible to understand this rev share as serving a national security purpose.

So Sam says the money is kind of. Almost immaterial to the US government, given the scale of the, the, the US government budget. And it also like if, if there is a national security threat giving the US government 15%, a 15% cut, there's absolutely nothing to address the national security threat, right?

You could envision a way that you tailor an export control to address a national security threat, you know? Require further downgrades of the chip, or you try to do end user controls or things like that. But 15% just doesn't serve a national security a national security purpose. And it's, it's not actually gonna bring in a huge amount of money that that said, like, I, I think so.

So I guess what I would say is, I, I think the way to understand this is not a national security. Issue, but rather Trump having decided that he thought it was okay from a national security perspective to sell the H 20 to China. You know, he bought into the arguments. These are old chips. It's better to keep the Chinese hooked on.

The US chip ecosystem then developed their own, like he just bought the, the substantive arguments there and then saw an opportunity to get a cut, right? So it's not, the 15% doesn't advance national security. He thought there's no national security and I issue and I can extract some, you know, revenue for the public fist here.

I think that's what's going on.

Alan Rozenshtein: I, I, I, I wanna, I wanna pause on this 'cause like this is the thing that's been driving me nuts. And then we'll finally get to the sort of con law part of this, which is also quite, I think, quite interesting. So when you say Trump getting a cut, this is a very interesting question, right?

Because, you know, when I first saw some reporting around this there was a talk about, you know, Trump is, you know, they're gonna have to pay Trump 15%. This is corruption. And you know, we could talk about the merits of this, but that's of course not what's happening. Like, I, I, at least, at least it's not been reported that Nvidia is gonna like, give Trump a, you know, a 15% check or endow the Trump Memorial Library in, in, you know, Palm Beach or, or, or whatever of this, it's going to the Fisk.

So, but then the question becomes why, like, what does Trump think this is useful and, and. My sense is that Trump just likes extracting things from people. Like this is just almost a purely psychological benefit to him, that, you know, the biggest, most important billionaires and industrialists, you know, come hat in hand to the White House or Mar-a-Lago or whatever, and, and he can dominate them.

Right. And he's a businessman, right? He's a, he's a real estate developer from Queens. And the way you dominate people is you get them to give money, even if. This is all kind of pointless. Like that's how I understand this, which doesn't really make me feel better. That's the only way I can get around this though.

This, it makes me uncomfortable. 'cause I do not like to psychoanalyze world leaders, especially Donald Trump. Am am my, let's start, start, start with you, Peter, and then I kinda wanna hear other, other folks, thoughts on whether I'm reading this psychodrama accurately.

Peter Harrel: So I am, I am not gonna be in the business of psychoanalyzing our president.

'cause it's not something I feel I have any expertise in. But I, I would make one maybe related point not to a psychology, but to policy. And one, one thing we have seen, and I, I, I, you know, I, I agree. Trump seems to be a very transactional president. I do think looking across his policy agenda, if you will, he does have.

Idiosyncratic views of where the US government should get tax revenue from. He doesn't actually think the US government doesn't need tax revenue. Right. But he actually wants to see, I think, a relatively lower tax burden on incomes and a relatively higher tax burden on trade flows. Right. And we see that on his.

Tariff front where he talks about it as a way of raising money for the Fisk. Now we can have all tax debate about why that's bad tax policy, but I think that Trump in his own head is in his own weird idiosyncratic way, looking for sources of revenue that he views as kind of free for the United States, and much as he.

Wrongly believes that tariff revenue is free for the United States from his trade war. He may also hear, kind of wrongly believe that this is a sort of free source of revenue for the United States. The the one point I would make, and I I just, it was an interesting point 'cause I saw mark Cuban, the, you know, former Dallas Mavericks owner making this point online, which, you know, is a fair point.

He pointed out, you know, progressive Democrats have tried completely unsuccessfully for the last 20 years to extract additional tax revenue out of big tech companies. Right. Completely unsuccessfully getting tax revenue out of big tech. And like, you know, it's not a ton of money as Sam argued, but like actually he is maybe extracting some tax revenue outta big tech here.

Janet Egan: Okay. I have an idea where we can get more tax revenue and not relinquish. The advantage to China and I think what we're missing is the concept of accessing chips through the cloud. So there's this whole idea of like, why don't we just rent and not sell these chips to China? So once you've exported these physical chips, they're often to China's mainland.

There's zero ability for the US to actually control who's accessing them, how they're being accumulated and amalgamated, and to what end purposes they're being used. But instead, if you just have an arrangement where you have trusted us, hyperscalers like already happens, owning that compute and Chinese firms given the assurance that yes, you can remote in and access.

Potentially up to a threshold that would serve many of these strategic purposes, like, undermining demand signals for domestic semiconductor industry in China. It also gives the US leverage looking forward as to look if geopolitical conditions change and yeah, the relationship might sour or there's risks that emerge.

It can then be sort of shut off at any point as well. And I don't, I think if you're thinking about revenue streams, there are ways to sort of have your cake and eat it too. And sure, China obviously won't be as happy and neither will Nvidia at that, but at the same time. There's still opportunities here to find these middle grounds and to ensure that, you know, you get more revenue for the US government.

That's great. But it also stops this comparative advantage being eroded.

Alan Rozenshtein: Well, well, well, let's actually talk about that for a second because I, I do think this question of chips versus clouds or I, I think miles Brundage, who used to be at OpenAI. I, I think on Expost it something like this is the Netflix versus DVDs model which I, I thought was a, a, a nice a nice comparison.

The, the, the, the cloud question's a very interesting one. I have a couple questions about that. First, you, you know, you said I totally under understand why. The Chinese government would not be as delighted with this. I mean, it's better than nothing. 'cause you can run the sort of non-national security models and whatever.

But why wouldn't, why would Nvidia care, right? I mean, Nvidia can get into cloud computing itself if it wants, right? I mean, it already does that obviously on a much, much smaller scale with with this like Nvidia GForce game video game cloud, cloud gaming system. So knows how to do that, even if it doesn't want to get into that.

Obviously. It kind of doesn't, you know, money's, money's fungible, right? It doesn't really matter if it's getting paid by China directly, or if it's getting paid by a US cloud provider to serve to China. So actually why does, why is Jensen Wong so gung-ho about selling to China other than, you know, or, or, or is it just a matter of Well, he's the CEO of a publicly traded company.

His obligation is to his shareholders, and if he can get a half a percentage more of profit, who cares about the US national interest?

Janet Egan: I think there's a lot of people in the ecosystem who have different takes on this and have different assessments of what the motivations are and the key drivers behind the strategy that NVIDIA's engaging in at the moment.

I do think that they are aiming to get more customers rather than consolidate into just hyperscaler customers. And to keep them as sort of a direct seller to many entities around the world, not just the key hyperscalers, but I don't know. Sam, did you have to add on that?

Sam Winter-Levy:  Well, I think a kind of additional, additional part of this puzzle and argument the kinda critics of this deal have made is that Nvidia is right now capacity constrained.

So every chip that they're selling, there's a buyer for. So one argument that people will make that kind of get to this point of, you know, what exactly are NVIDIA's incentives here? They'll argue that. Nvidia can't ev every chip Nvidia is making already has a, already has the one who wants to buy it in the US and elsewhere.

Like, why care so much about the Chinese market when you can't even address the full US market right now? And I think there the answer is it's more of a kind of short term, long term issue where in the short term it's true that Nvidia at least this may be changing now, but at least through 2024, first half of 2025, they were, they were, they were capacity constrained you know, enormous demand for these chips.

If that starts to ease in the future, then you really want to kind of block in as many additional markets as you can. And China will be a very big market for them to lose in a few years time if you know, if, if, if demand starts to ease off from other parts of the market. So I think that that then that kind of long run desire not to kind of cut themselves off from a, you know, potentially huge additional market, I think is another part of the, the incentives here for Nvidia.

Alan Rozenshtein: Jenna, I, I wanna go back to, to you to kind of ask my second question about the, the cloud versus chips debate. You know, you said that and this seems intuitive, that once you sell chips to China, you lose control of them. Now China gets to plug them into their domestic energy grid. You know, there's a, obviously a huge story right now about how much electricity generation they are, they're producing, they're completely putting the United States to shame.

Part of it's, you know, coal, but it's also a lot of renewable. So like, again, like once you get the, they are, they are. Chip constraint, not energy constrained. I, I have read some interesting proposals, some kind of thought experiments about whether in fact it is the case that it's all or nothing from a control perspective once you export the chips.

Is, is there, is there any possibility for on chip surveillance, monitoring? You know, I hate the word kill switch 'cause policymakers seem to like just saying kill switch, assuming that you can create kill switches, but. Kill switch or is that just all a pipe treatment? It really is the case that, look, if you're gonna give someone magic sand, they're gonna do whatever they want with it.

And you've just lost control over it at that point.

Janet Egan: Yeah, great question. I think at the moment, doing something like location verification where you have a chip ping to a trusted server nearby, and by the time like calculating that time it takes for that signal to be received, let, gives you within sort of a.

A bound of estimates, the location of the chip, that's already doable and it's used in other technologies and it's, it's not particularly novel. And I've seen reports that that could actually be done without major upgrades to chips. So that's something that is like low hanging fruit and can be done, but that isn't so useful if these chips are already being sent to China.

That's particularly useful if you are sending chips elsewhere and saying, do not send these chips to China. Because it lets you track smuggling routes and make sure that when chips are being diverted at a high rate from certain locations, you can increase enforcement there in a much more targeted way.

But I think the, the broader avenue of what you could do on a chip and there's a lot of folk exploring. There's, I think there's, many people for good reason, who push back against this idea of kill switches. Traditionally when you put back doors and things, people can find ways to exploit that for unintended purposes.

And so there's still a lot more research and development needed to actually think about this in greater detail in terms of what are the compromises you're willing to make in terms of the assurance of the chip or like, how can you make this tamperproof and robust that these things work when they're needed to.

So those things are a bit further off. I think and potentially weren't some investigation. But I think it's very different location verification on one side, doable now and then other things like more enabled control of a chip remotely. That's something that needs further work.

Alan Rozenshtein: Let's, let's go back to the legal questions.

'cause 'cause I, I do think there's still some, some useful stuff there. So we talked about the, the statutory questions. Peter you started talking about, then the question is of who can sue? Let's just talk about the, who can sue for a quest for, for a moment? I, I assume the, the sort of who can sue question is about standing, which is to say, you know, I can't, you know, just sue as a, as an enraged US citizen, right?

I have to be someone that's actually harmed by this. Presumably Nvidia is not gonna sue 'cause it's benefited by this. What about a what about a economic competitor to Nvidia, either in the, the chip space? I mean, maybe Intel, though of course Intel probably needs a big bailer from the government.

I, I, as I'm saying it, I'm hearing the problem. I mean, do you, do you see any realistic possibility for anyone to be able to sue given sort of pretty stringent standing requirements?

Peter Harrel: I mean, I'm not sure that anyone would have standing to sue unless they were a similarly situated and B you know, applied for a license and got told you two have to pay 15%.

Right? Because otherwise I'm just not sure who would show a harm. I mean, conceivably to me. You know, maybe a, a a a, you know, a chip reseller or a competitor that wanted chips to sell comparable spec chips to China and was denied a license, could sue. But I think the courts would insist they go and get, you know, see if they could get a license first.

And then if they were told they had to pay 15%, maybe they could, they could sue. But I'm not sure that anyone's gonna have standing, you know, unless they, they can show they're directly armed here.

Janet Egan: Can I ask a follow up on that? I've heard folks talk about a potential situation in which, say Nvidia is the one that could sue, but is choosing not to yet imagine policies change down the line.

Could there be a case of Nvidia retrospectively suing to recoup lost? Revenue. I don't,

Peter Harrel: I, I, you know, I think that would a, depend on the nature of the deal. And B I don't know. I haven't, I haven't thought that through or what the statute of limitations on any of this. Like, I just haven't thought through the, that kind of option.

Alan Rozenshtein: Alright. To, to close out the, the law conversation. Peter, explain to me why this might be constitutional and, and I'll just say and maybe I should not admit this, and if any of my students are listening, please pause. Please pause now. I teach constitutional law for a living. Like quite literally, that is what the University of Minnesota Law School pays me to do.

And I had no idea that the Constitution contained a. Relevant to this which again is shame on me, but I think it does show just how bizarre all of this is. And, and I say this not only as a a con law professor, but as someone who at a level, much less interesting than what Peter did in the White House.

But when I was at Department of Justice, did export controls, did trade stuff, did cfius and no one ever mentioned this constitutional issue. So, I, this was a learning experience for, for, for, for me as I suspect it was for for many of us. But Peter, what, what is this constitutional issue that is floating around here?

Peter Harrel: Yeah. So, article one, section nine of the Constitution, which I think probably Alan, when you're teaching it, you talk about, you know, this is a part of the Constitution that has kind of specific prohibitions on government authority. So this is a part of the Constitution that says, for example, you know, you, you, you, there, there can't be bills of ater and things like, it says the government can't issue titles of nobility.

Right. So, so sort of, sort of some of these things that come up on these kind of ne this, this list of specifically enumerated things the US government cannot do. And one of the specifically enumerated things that the government cannot do in. Article one, section nine of the Constitution. It says, no tax or duty shall be laid on articles exported from any state.

And this of course is coming from, you know, colonial era where one of the many tax related gripes, the revolutionaries in the early Americans had was when Britain was trying to tax American exports to other colonies and to the, the rest of the world. So there's a very clear. What I think in 1789 appeared to be a very clear cut constitutional prohibition on export taxes.

And, and I I should say before getting into the analysis here, like that, has by and large been true, like part of the reason this Nvidia deal seems so unprecedented is there haven't been previous efforts to tax, to the best of my knowledge, having done some of the historical research to tax exports at scale, where you do see.

You know, sort of fees associated with exports. They are quite small and clearly limited to like the government's. You know, expenses processing something, it's like the State Department, which governs export controls of military weapons, like fighter jets and munitions. If you wanna export fighter jets or munitions, you have to register with the State Department.

I think it's now $4,000 a year. You know, so it's like a, a fairly small fee. And in fact, you can get reductions in that if you're a small exporter. And it's very clearly tied to, you know, you got somebody at the State Department reading this document over there, so, you know. We've just never seen an effort to sort of substantively tax exports previously.

And, and, and part of that is this constitutional prohibition. Now, does this apply in this case again, let's stipulate that we don't know the details of this deal. Clearly the intent of that provision is to cover this kind of a tax on what we all think of as an export, but I think that if I were the government.

Arguing that this is constitutional, even if maybe violating a statute, I would note that first, I believe the age twenties are made in, or actually fbd in Taiwan. So I think you'd be arguing question of, is this an export from. Any state. Right. I, I think that's kind of question number one that you would, you know, litigate here.

The other one, I mean, what has been exported from, from the United States is intellectual property, right? I mean, what's happened is you've had Nvidia design a chip here and then Nvidia send the chip somewhere out in California, send the chip design over to Taiwan to be to be exported to be fbd in Taiwan.

The prohibition is by its own terms, an article. And I think you'd also, if you were defending this on behalf of the government, you'd say, well, insofar as there is an export of the United States, it's intellectual property. And that is not an article under under, you know, as, as sort of intended by the Constitution.

I don't know. I, I, I think. I, I would enjoy litigating this and arguing it is unconstitutional. I think you probably you know, should be able to prevail that. But I do see the governments, I do see the gov what the arguments the government would make given the export from the US' ip. And this is actually an article that is being exported from Taiwan, not a US

Alan Rozenshtein: state.

And, and, and again, all of this is the merits question. There's still this lingering. Well, yes, the first Nvidia

Peter Harrel: has to decide they wanna sue here, right? I mean, somebody's gotta decide they wanna sue. And you know, if I think about how this is likely to play out, I am, I take the point, you know, Sam and, and, and Janet, that maybe down the road Nvidia could change its mind.

And Sue, maybe I has then sort of thought that through. My assumption is putting that aside. If this is just kind of a one-off that applies to only Nvidia and a MD. We may not see a lawsuit here if, as the Treasury Secretary Scott Besson suggested in a media interview yesterday, the government might apply this concept of an export tax, or at least what the headline in Bloomberg characterize as an export tax covering.

Best essence RI. I think then we will at some point see somebody decide they wanna sue rather than pay an export tax.

Alan Rozenshtein: So I, I wanna, I wanna close out by asking what you all are gonna be looking. At sort of going forward right in, in terms of how this deal works out, sort of what the kind of key key details are, and also in particular, how much of a big deal this is, this is going to be, right.

So how will we know if the export control hawks, let's call them, are right. And this does in fact meaningfully advance China's. Capabilities either on training or inference and therefore harms us national security or perhaps the doves are correct. And, and sort of in the counterfactual world where we kept sort of, high export controls that simply created the conditions for the Huawei's and other Chinese tech companies to, to, to go.

So, why don't I, I, I start with Sam?

Sam Winter-Levy:  Yes. I think the two. I think two things to watch moving forward. One would be the reaction from Congress. So we've already seen kind of quite a. You know, a fairly significant amount of pushback to this deal from several kind of, you know, influential Republican national security voices, both kind of former Trump administration officials, but also you know, senators and and, and folks in Congress.

So, continuing to see whether Congress decides it wants to kind of flex its muscles here, how much of, how much leverage they actually have. I think that's one, one thing to watch moving forward. I think the second big thing to watch moving forward is whether the administration. You know, it seems like they've been persuaded on the age 20 that there's a national security case for, you know, easing our phone controls for the age 20, and then trying to kind of capture some of this revenue at the same time, whether they, whether we see them applying a similar argument elsewhere in the semiconductor supply chain.

So, do we start to see them potentially approving the exports of even more advanced chips? Maybe they get deprecated slightly, but you know. You know, the, some of NVIDIA's us more powerful AI chips, maybe the administration also proves the sale of those chips for on a, on a similar rationale to, to the case of the H 20, maybe even on semiconductor manufacturing equipment.

And I think that would be a bigger, that would be a much bigger move for them to take. I think that would really, that would really be it, a pretty dramatic step. But I think those would be the kind of two big things to watch. Is the H 20 just a one-off? You know, it was a kind of borderline case that the Biden administration itself kinda went backwards and forwards on.

And all of us extend further in terms of this kinda final question of how will we know who's right. I mean, I think that partly hinges on how big a deal, AI. Is in the coming years, if AI kind of fizzles out, it's kind of basically a normal technology, you know, a lot of the kinda national security hype around it is overrated, then maybe these sorts of deals make much more sense to do.

If it turns out that AI really is the kind of core strategic technology of the coming year, of the coming years, then I think. The administration's kind of calculus that it's worth trading away a massive lead in compute or kind eating into a massive lead in compute in exchange for the preservation of media's market share and some additional kind of marginal improvements to the us to, to, you know, to the treasury's revenue.

If that, if that you know, that calculus will look much less much less appealing if, if we start to see the kinda real national security stakes of, of, of these AI technologies in the kinda next you know, three to five years.

Peter Harrel: I guess I'd just say I, I defer to Sam and Janet on all things ai. So what I'll be watching out of this is whether the administration does try to apply these you know, export taxes slash rev share agreements on exports more broadly, and if so.

What kinds of goods you know, do they, do they apply that to, so I, I kind of come at it from, from that perspective, you know, and against a backdrop where we have a president who has been over the last seven, eight months, kind of remarkably interventionist in the American economies, especially for a Republican president.

Right. I mean, he's, we have the tariffs, we have this, we have, they may take a stake in, in Intel they've obviously expressed views on who should and shouldn't be CEOs of companies, right? He is a very interventionist president, and so I'll be interested to see if he views this as a new tool to be applied more broadly.

Or if this does end up kind of being a one-off, that fades away.

Janet Egan: I think for me Sam covered a lot of what I'll be watching, but an additional thing I'll watch is the number of leading AI companies coming out of China. Both in terms of applications and like frontier models. I think it's, it's interesting to see that.

You know, there's some reporting that Deep seek tried to use Huawei chips and then had real problems with them. And are we gonna see some upticks as there's more compute injected into that ecosystem than China can make itself? That's something I watch. And then I just to end on a nice hawkish note, I.

I just wanted to gently push back on the high export controls, creating conditions for tech companies like Huawei approach, because I guess I look at critical Minerals as an example where China's state sponsored approach and you know, centrally directed ecosystem is able to make in whole industries. Non-competitive for others to engage in and then wipe out competition to become that sort of really central player in the supply chain.

And so I think it's as we look at AI chips further, like at the moment, I'm not particularly optimistic that, well, not particularly pessimistic, that China can build out massive amounts of their own chips in the near term. Longer term. It'll be interesting to see if they're doing large subsidies on inferior chips to try and bring others into their ecosystem.

Alan Rozenshtein: I think it's a good place to end it. Peter, Janet. Sam, thanks so much for coming on Scaling Laws.

Janet Egan: Thanks so much to having me.

Alan Rozenshtein: Yeah, thank you.

Janet Egan: Thanks.

Kevin Frazier: Scaling Laws is a joint production of Lawfare and the University of Texas School of Law. You can get an ad free version of this and other Lawfare podcasts by becoming a Lawfare material supporter at our website, lawfare media.org/support. You'll also get access to special events and other content available only to our supporters. Please rate and review us wherever you get your podcasts. Check out our written work at lawfaremedia.org. You can also follow us on X and bluesky and email us at scalinglaws@lawfaremedia.org. This podcast was edited by Jay Venables from Goat Rodeo. Our theme song is from Alibi Music. As always, thank you for listening.