← The AI Policy Podcast
The AI Policy Podcast
What Selling Nvidia's Blackwell Chips to China Would Mean for the AI Race
01:04:29
Georgia Adamson and Saif Khan join Greg to unpack the geopolitical context of the B30A chip and the implications of US exports.
Loading summary
Transcript72 lines
- [00:00]
A
Foreign.
- [00:11]
B
Hey, it's Greg. On this episode, we've got a really exciting overview of the current state of the Nvidia chip exports to China debate. We are bringing on two guests who just wrote a paper about the B30A chip, which is Nvidia's latest and greatest chip that is marketed for the China market. But as we were recording this episode on November 3rd, a really bombshell exclusive reported story came out of the Wall Street Journal with some really important information about the state of the debate within the Trump administration on exporting Blackwell chips to China. So according to this report, Marco Rubio, the Secretary of State, pushed back on Jensen Huang and others regarding the idea of selling the more advanced Blackwell chips to China. And because of that, President Donald Trump decided not to raise that with Chinese leader Xi Jinping during the talks in South Korea. And notably, Marco Rubio was not alone among administrative officials. And here I'm going to quote from the Wall Street Journal piece. Others against the approval, the official said, included U.S. trade Representative Jameson Greer and Commerce Secretary Howard Lutnick, who helped lead trade talks. Faced with nearly unified opposition from his top advisors, Trump decided not to discuss the advanced Nvidia chips during his October 30 meeting with Xi in Busan, South Korea. The said, so that's where we are in the debate today. Of course, Nvidia is going to keep pushing the Trump administration to allow the export of the more advanced chips, but this is an important development and one that I think confirms a lot of the analysis that we give you on this episode about why selling the B30A or Blackwell generation chips in general would be such a game changer for China's AI industry and the US China AI race. So with that, I encourage you to listen to the whole episode. It's got a lot of really great data, and I'll leave you there. Welcome back to the AI Policy Podcast. On this week's episode, we're going to do a deep dive on Nvidia's B30 chips and proposals to sell such chips to China. And we've got two lovely guests who are in extremely good position to talk about this because they just finished a paper that they co authored along with two other folks titled should the US Sell Blackwell Chips to China? This is an incredibly hot topic in the exact moment of AI policymaking. It's something where articles are literally coming out every single day. So our guests are Georgia Adamson, who is a technology fellow at the Institute for Progress IFP, and more delightfully before that, spent about two years here at the Wadhwani AI center at CSIS with myself and the team. So, Georgia, welcome back to the podcast. And welcome back.
- [03:03]
C
Thanks very much.
- [03:04]
B
Great. And then our other guest is Saif Khan, who spent a number of years working in the National Security Council and at the Department of Commerce, spanning both the Biden administration and the second Trump administration, and is now a distinguished technology fellow at ifp. So, safe, thanks for coming on the pod.
- [03:24]
A
Thanks for having me.
- [03:25]
B
So let's just start with the basics here about the B30 chip and where it kind of fits into the story. I think almost everyone is familiar with the idea that there are these export controls that are restricting the sale of advanced AI chips to China. The Trump administration famously strengthened those controls by limiting the sale of H20 chips to China early in its term, and then later reversed itself and said that it would allow those sales of H20 chips. But at that point, China had blocked the importing of H20 chips. And so that's where the B30 chip arrives. It is of the Blackwell generation of chip. And so, Georgia, why don't you just sort of take us to the B30 as an item and where it fits into this overall policy debate?
- [04:17]
C
Yeah, I'm happy to. I think let's start a little bit, going back a little bit before we start with the kind of actual report that we wrote coming out late last month, but starting more early August. Actually, just where this story really begins, with Trump suggesting that he may allow a downgraded version of the Nvidia Blackwell chip to China. This is the famous, like, enhanced in a negative way Blackwell quote which made circles in the news a few months ago. And in that he says, you know, enhanced in another way Blackwell, in other words, take 30% to 50% off of it. This is kind of after reports that Jensen Huang Nvidia had been sort of lobbying the Trump administration to export B30A or what was what is now called the B30A, a Blackwell chip to China at the time. And then just about a week later in August 19, roughly, Reuters reported that this chip, which was reportedly then called the B30A, would have only a single die design of the flagship B300. So that's a kind of cutting edge Nvidia chip out there right now, part of the same Blackwell architecture family, but it would have compared to the kind of eight HBM stacks of the B300, it would only have four. This is a B30A and compared to the two Grace Blackwell processors, it would only have one. Moving forward from that. So we kind of had this sort of announcement of what was potentially speculation. We use these speculation, these speculated reports in our sort of assessment of what the B38 is for our paper, as well as announcements that the chip would be only 20 to $25,000, which is about half of the B300 price that was released around early September. So leading up to the whole XI Trump trade deal that happened last week on Thursday, October 30th in Seoul, there were kind of these like small leakages of what was actually going to come out from Nvidia that were all kind of rumored reports and rumored specifications, but which we've taken in our report to say, okay, if these speculations are true, if any kind of version of this 50% downgraded Blackwell from the B300 as Reuters reported and as Trump himself suggested are true, what does this mean for US China AI competition if we were to export these in either small quantities or at scale, and what does that mean for like overall U.S. competitiveness on the AI race and the U.S. aI advantage? We had heard rumors in like around mid October that there was some amount of scope for this chip to be on the table for deals. And then as it was kind of ramping up towards the actual trade deal that happened last week, it was becoming increasingly plausible. And as Trump himself said that we'll be speaking about a Blackwell deal in Seoul, it was becoming very clear that this was very much on the table. And so we wanted to kind of get this analysis out. Think about like, okay, if these are the actual specifications of this chip according to various media report, media reporting at the time, what does this actually really look like? So that's the super high level kind of pitch of our paper. We look at kind of US China competition from a kind of total compute perspective. We look at Chinese capabilities to produce their own state of the art AI chips and how those sort of compare to the B30A and other US AI chips and then also just kind of Chinese manufacturing capabilities, looking at kind of the HBM side of the equation, which is very important for AI production writ large, but then also on the kind of logic processing die side of things. So cool. We have all sorts of numbers. This, this report thanks to SAFE and also my other colleagues as well, Tim and Tao, who we should give a shout out to. Also IFP is full of numbers, it's full of estimates. We have lengthy appendices of methodology and everything like that. But happy to get into the specifics as we go on.
- [08:40]
B
Great. So I definitely want to go there because that's what I thought was so interesting about the paper is you actually sort of modeled the scenarios for if this is true, if the chip is this good, how much of a improvement in the overall computing environment of China could it possibly make under different kinds of procurement and deployment scenarios? So we're going to get into that. But let me just rewind for one second and say, you know, what is an H20 chip? And there are kind of three things that you could hear in everyday language called an H20. There is the silicon die. So that is you have a silicon wafer which is like a big circle. That's what comes out of the fab. You cut that up into little rectangles and one of those is a silicon die. Then you integrate that into what's called a module, which has the all surrounding HBM high bandwidth memory chips on a silicon interposer so that all these chips are collaborating in a very efficient manner. And that module you could also call the H20. And then the third thing are the servers, which is groups of modules integrated together. So you have the silicon die, you have the module and you have the server. And you might hear anyone refer to any one of these three things as an H20. What's very interesting is that the H100, this is the hopper generation of chips that was state of the art a year ago. The H100 and the H20 have the same silicon die, which means they come off the same production line at TSMC and are identical except for Nvidia can blow fuses to degrade the performance of the H100 die. And that degraded performance version then becomes the H20 die, even though they start out as the same. Now the module level is different. You have different arrangements of hbm, you have different arrangements of sort of other components. So that's where they start differing and that results in the differing performance. In the case of the H20, that's a case where the US government told Nvidia, this is what we're going to allow you to sell performance wise. And Nvidia then created a chip that was right up to that line of allowable performance in order to be exported the B30. Correct me if I'm wrong, here is a kind of a different story, which is to say this is Nvidia's pitch to the US government. They know that this chip is currently illegal under existing export regulations to sell without a license, and that license applications to export these chips would be reviewed with a presumption of denial given the level of performance. So Whereas with the H20, you know, the government told Nvidia what they were going to be allowed to sell. This is Nvidia coming up with a product and says this is what we think you should allow us to sell. And let's just, you know, put, let's start putting some numbers to this. Like how good is a B30 compared to an H100 compared to a B300 compared to an H20? Like how good is this chip?
- [11:52]
C
Yeah, absolutely. And safe. Feel free to jump in here if I'm like messing anything but tons of numbers to add to here. So I guess we'll start with the sort of B30A compared to the B300 because I think that's the closest comparison. How we.
- [12:06]
B
That's like the best thing you can buy in America versus what Nvidia is proposing to let be the best thing you can sell in China.
- [12:12]
C
Exactly. So the flagship Blackwell, the best thing you can buy not just in America but in the entire world on AI chips versus the one the downgraded or enhanced in a negative way. 30 to 50% here we're taking 50% based on reported specs from Reuters and others. So let's. So starting with that kind of the B30A versus B300 because this is a sort of major comparison, you need to know the B300 has two pieces of silicon, these AI processor dies and eight stacks of the high bandwidth memory. It's HBM3E that create this B300 that we already know. The specs are out there, largely well reported and well covered. If it is true that the B30A is half of this, so it has half the silicon, has half the HBM available, then we know that it is approximately half the performance generally. So as a rule, kind of if you have half the silicon available, half the bandwidth memory, you have kind of half the memory bandwidth capabilities and also the processing performance. So these are the things needed for the memory bandwidth on the kind of high data transfer requirements for AI and then for the processing, this is kind of the large amount of calculations needed for AI performance generally. In thinking about kind of what this means in terms of overall compute metrics, we in our report use the BIS metric for kind of assessing overall chip capabilities. So this is Bureau of Industry and Security over at Commerce who are involved in all of the chip export controls. They have something called total processing performance which is how they sort of account for different, account for different chip capabilities and compare them on a sort of apples to apples basis. So for the B300 we assess the kind of B300 chip has about 60,000 total processing performance, TPP taking about half of that. Assuming that kind of 1/2, 1/2 capabilities, we infer that B38 has about 30,000 TPP. What this really means to say overall is it has essentially just half on par of this high bandwidth memory, so inferencing capabilities largely and then the processing performance, so training capabilities. Now where this becomes really interesting is kind of comparing it to what's already available on the Chinese market. So this gets to the H20 that you were alluding to earlier and then also the kind of Chinese state of the art which is the Huawei 910C. So on the H20 side, this is what you said was kind of Trump Admin has flip flopped on whether to allow exports, whether to not allow exports of the H20 because it's seen as a kind of semi national security risk, but something that maybe the United States is willing to take a bet on given that it is fairly downgraded compared to other flagship or prior flagship Nvidia chips like the H100. So according to the TPP metrics alone, we estimate that the B38 is about 12 times, it's technically 12.6 times better than the H20. So that's 12.6 times, almost 13 times the next best US chip on the Chinese market. Similarly for the 910C which is Chinese own chip, that's the Huawei Ascend chip, that is something around kind of 2x, the processing performance has about 25% of the memory bandwidth capabilities. And overall sort of a price performance, which is really what matters at scale here when we're thinking about data center level stuff, as we usually are, around 3x a better price performance than China's own state of the art. So when you start to like you're.
- [16:04]
B
Saying the B30A would have 3x the performance of the Huawei 910C in terms of price performance.
- [16:14]
C
So this is kind of on like performance per purchase. So when you're thinking about like for every dollar that you put towards this chip, this is the kind of performance that you're getting out of it versus a dollar spent on a different chip.
- [16:26]
B
And then what is the price to performance comparison between the B30A and the H20?
- [16:33]
C
Yeah, so B38, it's pretty, generally pretty good. So when we think about price performance, we're kind of looking at TPP divided by a hundred dollars, just kind of a standard thing. So B38 we estimate if you take TPP divided by 100 it's around 133. This is in measured in TPP, something like the H20 is actually only like 20. So you have almost close to a 10x, maybe not quite 10, but 7x.
- [17:05]
B
Correct me if I'm wrong here, right, but that your analysis would suggest that the Huawei 910C is better than the H20.
- [17:13]
C
I mean, I think in terms of training. Yes.
- [17:14]
B
Yeah, okay, that's really interesting. So I think. And you're putting it in a sort of processing power. I mean TPP is total processing power, which is flops, you know, the amount of computations it can do per second. And then also related to the area on the silicon die that is occupied in order to generate that computational power. But I think there's other things that go into computational performance that are worth pointing out here and that comes down to like how easy the chips are to use, whether or not there is a software ecosystem that can make effective use of them. And I think there was some really interesting reporting that came out in scishin, which is sort of like China's Wall Street Journal. So it's the number one business newspaper and magazine in China and they reported on October 27, 2025 that at AI computing centers in China, chip utilization is below 30%. Now why is that? Right. If the chips actually have, you know, decent processing performance, it's because they have very, very Huawei chips have very high failure rates. They break a lot more often than Nvidia chips. So there's a reliability issue. The surrounding software ecosystem. So like, can you actually use this hardware to run useful software is terrible compared to Nvidia. And so the point basically being that yes, like the apples to apples hardware performance things matters, but there's also other variables in the equation that matters. And a lot of the way in which Huawei is unattractive is separate from its computing performance. And just to give you a point of comparison, like this 30% utilization in late October, Amin Vadat, who is the VP of AI Infrastructure at Google, he said that at Google, 7 and 8 year old TPUs, which is the Google proprietary chip, have 100% utilization. So even these Stone Age era TPUs at Google are still getting 100% utilization because the demand is so incredible. So what's going on in China where they can't get above 30% utilization? It's because the chips stink. And they stink in multiple ways, not just the hardware things. But I do think it's worth just calling out for our audience here. You are not disputing that on a pure hardware specification basis, that Huawei's latest Grace chips, which is the 910C, not the 910B, which is far more widely deployed in China, actually is better than the H20, which is of course something that should concern Nvidia. And I think one thing that we should say on the hardware specifications about the Huawei Ascend 910C versus the H20 is that like, even though it is better in TPP, which is like one of the metrics that US export controls care about and that matter to customers of AI data centers, there are other hardware metrics where the H20 is better. One that's really important is memory bandwidth. So basically how fast, you know, the information can be made available for processing on the chip. And in that regard the H20 is better than the Ascend. And it's better in a way that matters a lot to specific use cases like AI inference. The H20 is way better in terms of a software stack reliability, the ability for the chips to integrate and work as a team on the sort of software part of the equation. And then it's also better on this key performance parameter of hardware. Even though, you know, Huawei is not totally useless on some hardware metrics and can look good, but in a way that I would argue is kind of misleading.
- [21:08]
A
Yeah, I would add one thing on top of kind of the point that you made, Greg, about, you know, the specifications on the tin, you know, that is, at least for processing performance, the Ascend does quite well. I think we would totally agree with you that the software ecosystem is quite immature. And especially as you're scaling to a cluster level, you know, these chips, they'll fail at some kind of rate. And once you aggregate this many GPUs into a cluster with thousands of chips, the chances that you will have one or two GPUs going offline at a given time and causing instability in the overall cluster increases quite a bit. And I think that contributes to the relatively low utilization rate that you get. So that is actually right.
- [21:54]
B
Just to help people understand what SAFE is talking about there. The mean time between failures for a Hopper Nvidia Hopper chip is six years. So on average a Hopper chip will just die and become non functional after six years of continuous use. So if you have 365 times six chips, you should expect one to fail every day in your cluster. And that's what state of the art reliability looks like. Huawei chips have much lower reliability than that. And what it means is you're constantly facing chip failures. And Nvidia has built a lot of redundancy into its overall data center stack to make sure that training can go on even if there are chip level disruptions. I think the best available public discussion of this topic said that it was extremely difficult to network more more than 2,000 Huawei chips together as of about a year ago. And that is to have these chips working as a team where the whole is more than the sum of the parts. And by contrast, we now know of Nvidia chip clusters that are 200,000, 400,000. So they're really good at working together as a team, which is another thing that sort of makes them attractive. Sepafat I think Huawei is making progress on getting more and more chips to be able to network together. They're probably upwards of 5,000 or 10,000 now, but still far below where Nvidia is. And this is yet another sort of metric of performance other than processing power that really matters.
- [23:23]
A
Okay, just to add to that, Greg.
- [23:26]
C
I was just going to say just to point out also what this looks like in practice because you can have all these numbers in theory and stuff, but thinking about actually models that are trained on US versus Chinese chips, I know in 2024 there are about 100 roughly models Chinese models trained on US trips versus just two models trained on Chinese chips. For this reason, this is kind of citing epoch data and similar kind of ratios in 2025, although the average model number has gone down. And then similarly, I think, you know, when you're looking at so as a.
- [24:01]
B
Revealed preference, Chinese model developers still prefer Nvidia.
- [24:05]
C
Still prefer Nvidia. And then thinking about actually even just for the H20 bans that were coming into effect in the spring, these are just kind of anecdotal evidence but you can sort of it points to this larger point of in early spring when the H20 was banned was being announced, there was a lot of stockpiling of H20 chips by Chinese companies. So I think it was Tencent Baidu and Deep Seq or some kind of combination of that. I probably missing one of them, but it was around $16 billion worth of stockpiled of H20 chips. Because of this superiority, even when the H20 is not the best in class of US chips anymore, and it wasn't even at that time, but it was still better than what a lot of Chinese companies were able to get their hands on. And similarly, I mean, just like at a scale thing that we were talking about before, I know like Tencent Baidu and Deep SEQ execs have all said that their number one bottleneck to kind of AI development and inferencing is AI chips. And this is a.
- [25:07]
B
So let's talk about that as the bottleneck for China because you have modeled in this paper what would happen to China's overall computing ecosystem and the sort of total aggregate amount of compute that is available in China. In the case of we do allow B30 exports and we don't allow B30A exports. But just because I want to give the. Before we go there, before we give the, I want to give the other side of this argument a fair hearing and so safe. Could I ask you to please sort of Steel man, the argument for allowing B30A exports, like one of the strongest arguments in favor of allowing these exports. Obviously, I know you came down on the other side, that's why you co authored this paper. But let's just, you know, give, give the other side a fair hearing here first.
- [25:58]
A
Yeah, absolutely. So you know that one of the key arguments that you hear is Huawei. We shouldn't underestimate them. Obviously. You know, there were export controls put in place many years ago on Huawei and they're still surviving. And so maybe we should take seriously that they're actually going to scale up chip production at a level that will be quite surprising. And if it is the case that Huawei is going to scale up chip production anyway, and the choice we have is either to have them hooked on US Technology, US Chips, or Huawei chips, is much better for the United States to have China be dependent on US Made chips. That kind of has benefits to the overall US tech ecosystem. You know, you hear arguments that, you know, China has a very large percentage of all of the AI developers. And the way that these software ecosystems become mature, you know, to the point that we discussed earlier, is that you have a developer community that is helping to troubleshoot and improve the software and the whole overall ecosystem and that provides a real benefit.
- [27:01]
B
So the argument, if I could sort of highly oversimplify it, is you're going to help Alibaba and Baidu. Yes. By selling Nvidia, but you're going to hurt Huawei. And it's probably worth more to hurt Huawei than it is to, you know, the, the consequence in the, in this steel man argument you're presenting.
- [27:21]
A
Yeah, I think that's right. I mean, there's two different scenarios here. I would say there's actually one scenario where we're so dramatically underestimating China that they're actually just going to satisfy domestic demand, in which case, you know, it's not even the case that we actually have the leverage to harm Alibaba Baidu Frontier AI development in a long term sustainable way. But I agree with you that there are also kind of intermediate worlds where Huawei has a reasonable amount of production but is not able to satisfy demand. And that there just is this kind of trade off that you described. And then the other element is for selling the B30A specifically, you know, China's banned, at least temporarily. We don't know how long term the ban is going to be. Some less powerful chips from being used in China, including the H20, as well as the less powerful Nvidia chip called the RTX 6000D. And so the argument is, if you want to be able to achieve dependency on China, China's dependency on the US tech stack, maybe you're going to have to sell something that they actually want.
- [28:20]
B
And so that here I have to interject because I've heard some information that I think is, is worth just sharing. So obviously, you know, when you're talking about what's going on in the world of smuggling of chips, there are multiple constituencies that have an incentive to lie. But I'll just say that the Financial Times reported that Chinese regulatory crackdown on H20s is genuine. Which is to say Chinese regulators actually are going around and saying like, no, we don't want you to buy any more H20s. No, we don't want you to install any more H20s. NoO, we don't want you to buy any more RTX. The Financial Times also reported that Chinese authorities are now working to crack down on smuggling of Blackwell chips. And I'll just say I've talked to people who have talked to smuggling communities in China and they say that that Financial Times report is false and that China is still happy to allow smuggling of the most advanced chips. So in terms of, you know, what is China's incentive for shutting down the imports of H20s? Obviously there's different constituencies in China. Huawei is in favor of banning all Nvidia forever. I think that's pretty clear what their incentive structure is. Alibaba and Baidu would probably be in favor of allowing all Nvidia imports forever. And then the Chinese government has to thread the needle. And at least from what I understand the data that we have, they are threading the needle by blocking H20 and allowing smuggling of the more advanced chips, which suggests that they still want them and they still see a benefit to their AI ecosystem from having them. So that would lend credibility to the hypothesis that this ban was part of a negotiating gambit to basically persuade Trump to allow sales of something like the B30A, which is considerably more attractive than the H20 for all the reasons that you've said, and I should say I'm not the first person to hypothesize that this is just a negotiating gambit. Bill Bishop of Cynicism had a nice write up on this right around the time it happened. But I think that based on the data that we have, that's the most credible interpretation of events thus far.
- [30:30]
C
And not to give away like the end of the story, just to jump ahead briefly, because it matters here a lot. Part of the trade off of not selling the Blackwell and the exporting the Blackwell is part of the US China trade deal. On Thursday was Trump, President Trump saying, you know, we are not, we did discuss the chips. We are not exporting the Blackwell. We're not talking about the Blackwell that just came out yesterday. But we will be exporting a lot of chips. And he says a lot of chips that will be good for us. So the question is kind of what is this chip that they're going to allow exports? Likely, in one version of events, it's that Chinese regulators roll back these bans on the H20 and RTX 6000D, which is the other kind of Chinese specific Nvidia chip. And that's the chip that we are saying, okay, this will be the large volumes of exports because anything after that is either already banned to Chinese markets. So kind of the H100 sort of side of the story. And then we've already just discussed the Black bells. So I think you can expect that to change as well.
- [31:29]
B
So now I think we can get into the meat of your paper and a lot of the number crunching that you and the team did here. And I want to first say that I do encourage folks to go read this paper. It's got a lot of good data, it's got a lot of good charts. Obviously, there's some uncertainty with the numbers that they're inputting into this modeling. But the authors have been very transparent about, you know, where, where they're making assumptions and what those assumptions are grounded in. And if you differ with those, you can, you know, change the numbers and get different results. So let's go to the four major implications that you have from this paper. First, you said that such a policy would be a substantial departure from the administration's current export control strategy. Georgia, I think you've already explained to us how that is true, because you're saying that these B30 chips would be 12.6 times better than the H20 in terms of raw processing performance. And I forget what number you're using, but it's something like 6x performance per dollar, is that right?
- [32:32]
C
Yes. Along those lines, I think here what's also really worth pointing out is just kind of repeating back the Trump administration's own stated priorities for export controls. So in the first day of office, President Trump signed executive order on the trade policy, which was to in part tighten and close export control loopholes. And then in later in the action plan, which came out at the end of July, there is a whole part on export controls and tightening these kind of loopholes as well. And I think this is kind of worth actually just reading out loud as a quote from the AI Action plan, which is that advanced AI compute is essential to the AI era, enabling both economic dynamism and novel military capabilities. Denying our foreign adversaries access to this resource then is a matter of both geostrategic competition and national security.
- [33:22]
B
So that's a direct quote from the Trump administration's flagship AI policy. They say they want to limit compute exports to China, at least by allowing the H20 sales. Trump is not weaker than the Biden administration because the Biden administration also allowed H20 sales. It's just the Trump administration reversed itself on that point. But allowing B30A sales would unambiguously be a weakening of the restrictions compared to the Biden administration and even the first part of the Trump administration. Great. So the second thing that you claim is that fewer AI chips would be sold to U.S. customers. What do you mean by that?
- [34:04]
A
So on this, what we specifically note is when there are conditions of supply inelasticity. And so what that means is there will be times, number one in the world where due to the really rapid kind of increase in demand for AI chips, there have been times when manufacturing capacity for AI chips has not been able to keep up the supply has not been able to keep up with the enormous, increasing, exponentially increasing demand for AI chips. And so we have seen past instances where there have been supply shortages and we can expect those to continue into the future as long as we're in a world of kind of rapid scaling of compute.
- [34:42]
B
And I think just to understand where we are in that story, I think the quote that I read earlier from the VP of AI Infrastructure at Google, where he said even 8 year old TPUs are 100% subscribed, you know, that suggests that we're very much chip constrained right now. But I do want to flag that Microsoft CEO Satya Nadella said just the other day that on a podcast that they actually have GPUs that they have procured that they cannot install because they're waiting for available power generation capacity. So we're in this bizarre kind of twilight zone moment where end user demand for AI tokens is insatiable and leads to, you know, skyrocketing GPU chip demand. However, GPU chip demand in a practical sense can be bottlenecked by energy buildout. And so there's kind of a oversupply and an undersupply at the same time. But why, why does Nvidia as a company care about having the Chinese customers if Microsoft or Google or whomever, OpenAI anthropic is willing to buy all the chips that Nvidia can make? Like, what's Nvidia's incentive structure to want to have more customers, even if ultimately it's just going to sell every single chip TSMC can make?
- [36:07]
C
I'm happy to jump in here and then go. Go ahead, Steve. I think, I mean, customer diversification is like the name of the game. So when you have, I can't remember exactly the statistic it's escaping right now, but something like, is it 60% of your total kind of supply is sold to top three hyperscalers, roughly which are all American. You know, obviously there are kind of great risks with just concentrating so much of your sales on so few customers and, and just looking for other sources of diverse revenue is super important to any business, as it would be for a $5 trillion company. And, you know, if any one of these hyperscalers kind of go bust in some version of an AI bubble or they're just, you know, get other sources for AI chips, whether that's kind of an ASIC owned brew or something else, you know, that's, that's major revenue losses for Nvidia. So just just looking for other areas to make money is a huge.
- [37:04]
B
Yeah, I think, let's just even go to a hypothetical extreme, right, where there's only one customer. OpenAI is the only customer. Well, then Nvidia maybe is a monopoly supplier, but OpenAI is also a monopoly buyer. And so the negotiating leverage between them is relatively equal. If by contrast, there are a thousand AI companies and they're all roughly of equal size, then Nvidia has all the negotiating leverage. None of those a thousand companies companies have really any leverage. So adding China to the customer pool basically bids up the price of chips and also diversifies the customer base to decrease the negotiating leverage of any one individual customer, including all the American customers.
- [37:44]
A
Yeah, to that is, you know, you also have certain customers. Even if there's not a kind of a monopsonic condition here where you have like a single buyer or a really powerful buyer that is distorting the market.
- [37:56]
B
It.
- [37:56]
A
It's also just the case that once certain consumers get to a large enough scale, it starts to make economic sense for them to use in house asics. We see this kind of across the board with Google and its TPUs, Amazon and Trainium. And you have other hyperscalers that are also developing in house silicon as well. I would also add one other factor here in terms of kind of the argument that fewer AI chips would be sold to US customers and the rest of the world if China got more chips, which is if you give B30 as to Chinese companies, that will actually kind of empower Chinese cloud companies to service customers around the world. Suppose there's a customer somewhere in the Middle east or Europe or the like, and a Chinese cloud provider using B30 as is getting AI business from these consumers, they will actually decrease the demand that is available to US cloud providers to serve those same customers. So they'll then buy fewer chips.
- [38:53]
B
Oh, right, yeah, yeah. Because if we sell a bunch of chips to Alibaba, then Alibaba Cloud will go compete with Microsoft Azure, Amazon, AWS in Africa, in Indonesia, in wherever, using those chips that they bought. And then if Microsoft and Amazon and their revenue goes down, then maybe they'll order fewer chips the next time around. And yeah, so there's a clear line there to me. Okay, the third big implication that you discuss in this paper, Chinese AI labs would have access to equally powerful supercomputers as the U.S. georgie, we want to unpack this one for us.
- [39:29]
C
Yeah, absolutely. This comes back to kind of the 1/2 times plus 1/2 times 2 sort of argument that we were making earlier. So really on the kind of chip specification level. But then we can talk larger about the kind of cluster size. I mean, most simply like when you have half, half a chip, essentially. So B300 sort of cut in half and being overly simplistic here, but you have kind of half the inputs, you get half the kind of performance at the end of the story from that. And then you're pricing it at half the kind of price of a B300. You can simply buy two times as many. This is kind of what Chinese customers have reportedly called this, a good deal. And it Certainly is because you can achieve in theory kind of similar AI capabilities as leading US AI labs who have access to the full B300 at around a similar cost to USAI Labs when you get this like half price kind of deal.
- [40:25]
B
So this is a really, this is a really important claim that you're making. You're saying that effectively Even though the B30A is a quote unquote 50% performance degradation versus the B300, if China and its data center customers are sophisticated, they can actually take two B30As and put them together and get something that is almost as good as a B300. So it's not a 50% performance degradation. I think you said in the paper that it is going to be more expensive to have to put these two chips together and make them work as a team and blah, blah, blah, blah, but that it's actually only a 20% performance penalty in a performance per dollar basis. So actually it's, it's, it's a, it's a good deal as, as you said, Chinese customers are openly talking about on the Internet.
- [41:17]
C
Yeah, absolutely. I mean, I think, yeah, just to reiterate, these are obviously speculated things as well. Again, just want to like huge caveat to all of that but like this.
- [41:26]
B
Is, Nvidia has not yet released the specifications for the B3A.
- [41:29]
C
Yeah, this is not, this is not official from Nvidia. This is kind of through Reuters reporting and similar, you know, trusted news outlets. But if any of this is speculated is true, which we can sort of assume some level of it is, or at least this is the best available information that we have at the time to be making informed decisions about U.S. export control policy ahead of, you know, the Trump Xi deal last week. This is exactly a good deal. There's, we estimate about a 20% more cost to Chinese consumers just because of kind of networking scaling that you have to do. So for every chip that you buy, if you're buying two of them, the kind of networking price that you have to, you pay to network these chips together. And also the server cost just kind of scales linearly to that. It's a bit of a fixed cost.
- [42:18]
B
Per chip, but in overall terms.
- [42:20]
C
Right, like overall that's great. Yeah, yeah, that's something that China will, will certainly cough up is my, my non expert view. Even just.
- [42:29]
B
And, and when Huawei, you know, talks about how their servers are competitive with Nvidia, you know, they're talking about it in terms of flops. But in terms of flops per watt, Huawei is like nowhere near competitive with Nvidia. And that really matters when you talk about having to build a gigawatt, you know, power plant next to this data center to feed it. Right. So the overall construction cost of the data center and its associated power infrastructure might be much, much, much higher if you're buying Chinese chips versus if you're buying Nvidia chips. And then there's the secondary phenomenon we're talking about, which is that Nvidia software stack, Nvidia's networking stack is so good that you can make a cluster of 400,000 chips or a million chips. And at least right now you just can't do that with Huawei, full stop. And so that gets to your conclusion about how at the highest level of clusters they would be able to match XAI Colossus, OpenAI Stargate. And that is just, at least for right now, not on the table. Yeah, yeah.
- [43:35]
A
And one thing to add to that, we actually have a couple of numbers noting that the Huawei cloud matrix server system that includes hundreds of Huawei Ascend GPUs, comparing that on a kind of flops per dollar metric with kind of some of the reported numbers for the Nvidia servers that use B200s, which is the chip that came before the B300. And you actually just see something like a 2.9x better price performance on purchase price per flop and then also on energy consumption as well, you see kind of a similar premium for Nvidia chip. So yeah, you will just have to build very large power plants in order to.
- [44:17]
B
Yeah. This reminds me of a Jensen Huang quote from, I think it was last year when he said, if you think about the total cost of ownership of our chips, you know how many tokens they're going to get for you, how much you have to spend on the data center, how much you have to spend on the power infrastructure. Our total cost of ownership advantage is so good that even if our competitors chips were free, it would not be cheap enough for them to be competitive with us. And I think it's kind of interesting that you have that kind of confidence coming out of Jensen Huang and then also at the same time, oh my gosh, Huawei is going to destroy us kind of fear coming out of Jensen Huang. It's an interesting juxtaposition of those two positions. Okay, now let's go to the fourth major conclusion you have here, which I think is really, really valuable for the readers to understand. So US AI compute advantage over China would shrink dramatically. Georgia, what does that mean? And can you put Some numbers behind it?
- [45:15]
C
Yeah, absolutely. I'll kind of start by priming like, what is compute advantage? What is it good for? And then maybe see as the kind of wizard behind numbers can add some color to this. When we talk about compute advantage in total, I think this is where it kind of matters to think sort of zoom out from these individual specs on individual chips and everything is like 2x or related to something else or 3x worse than something else. It's just important to think about like, okay, what do all these numbers actually aggregate to? And why is that sort of important? So why is kind of compute the one thing that like the US AI compute advantage when we're talking about that, why is that what we're focusing on? And why is it chips in terms of just total, you know, compute production globally? The US is undoubtedly in the lead here. We have some numbers on what Chinese compute production is, which we can go into later. But I think it's like thinking about what are the kind of capabilities that you get out of this. At the end of the day, you have, you know, certainly according to scaling laws that are well documented, you get better AI sort of training performance. You also just get better inference. So you're able to run more models for longer and have for more users, for more users. And with kind of the rise of reasoning models, the longer you have this kind of test time compute, the better your performance is in many ways. And then I think the third one that's very important to think about, especially when we think about kind of export control debate is on the innovation side. There's this kind of argument that US export controls have fueled Chinese innovation and it's actually made them be more creative. And Deep Seek showed that it had similar efficiencies to US models not too long ago, about a year ago now actually. But in terms of the actual compute that USAI Labs are using, there's a great table from Epoch as well on showing kind of the just vast quantities of AI compute that are actually being used on AI R&D. So that's kind of looking at new innovative ways for algorithmic developments and efficiencies and ways to kind of essentially make AI models just better. So Deep SEQ showed that it could kind of do a sort of jump change in efficiencies last year. But there is a vast amount of compute that USAI Labs are using way more than they're using to actually train models or inference them on just developing new kind of paradigms and new ways to make more efficient models.
- [47:38]
B
Yeah, this is kind of, I mean, what you're sort of saying is that, that having a lot of compute is very helpful to accelerate the pace at which you discover compute efficiency. And this is an advantage that American companies have right now. So let's talk about that advantage because you have different scenarios that you modeled here. Under no AI chip export to China, conservative B30A exports, baseline B30A exports, and aggressive B30A and other AI chip exports. And to sort of get to the baseline, I think it's very, very interesting that in your scenario, even with smuggling, which you're taking into account as a real factor and a pretty high amount of smuggling, you're saying in one of your scenarios, let's say that 4% of all US AI chips are smuggled to China, the United States still has a compute advantage over China of 11.7x. And if we manage to crack down on smuggling and stop it entirely, the US compute advantage over China is 31x. By contrast, in your most aggressive scenario where you suggest that China would buy 31% of all US AI chips and have an inelastic supply condition, China would actually have a compute advantage over the United States of 1.1x. So safe. I think this is a pretty good remarkable claim. I mean, you're basically saying the difference between selling B30As and not selling B30As could be the difference between a world in which America has 31 times as much AI compute, at least produced in 2026, and a different world in which China in 2026 actually gets more AI compute added to their overall system than us. I mean, am I interpreting these numbers correctly? Is there anything else you want to unpack from this table?
- [49:32]
A
Yeah, happy to. So first, before getting into that specific scenario analysis, it's worth just noting that a lot of this rides on relative US and China production numbers for GPUs. And so we have a few charts where we kind of aggregated analysis from various sources, including crunching some of our own numbers to kind of get a sense of what the US advantage in production actually is. And so kind of going back for a second, I think we've discussed this a little bit, but it's worth noting there are kind of two key components that go into a high end AI chip. One is the AI processor die, and then the second is high bandwidth memory. And the US has and partners have kind of an enormous manufacturing advantage that is the result of U.S. and allied export controls on semiconductor manufacturing equipment, which has just created this production bottleneck and gets to the point where you know, even if compute was not available to China, there are just hard limits on how much they can make indigenously. And so number one on the AI processor die side, we have this kind of normalized metric for 7 nanometer and below production for US and partners versus China. And the result that we have is that the US has about a 35 to 38x advantage. And then if you further adjust that for the fact that China and SMIC in particular, which is their leading chip maker, has very low yields on making advanced chips.
- [50:57]
B
Yeah, I think this is worth just elaborating on because it's a really important constraint in the overall Chinese ecosystem. Right. If, if America sells no chips to China and we don't allow Taiwan to sell chips to China, we don't allow Korea to sell chips to China, right. Then to a first approximation, the chips that China will have are the chips that China can make. And they have been constrained by restricting exports of US semiconductor manufacturing equipment and their leading logic chip provider, their leading advanced logic chip manufacturer manufacturer smic, which is the, the partner that makes the Huawei chips, at least for now. The Huawei AI chips, they have been struggling with very low as safe said yields, which is on the big circle semiconductor wafer. What percentage of those chips actually work when they come out of the manufacturing process? And to give you some points of comparison, you know, when TSMC, the Taiwanese semiconductor giant, was making 7 nanometer phone chips like six years ago, before they introduced EUV technology, they were getting yields of 70 to 75%. That's with no EUV lithography technology, although that is phone chips, which is not quite apples to apples with GPU, according to the best reporting available, SMIC was at 20% yields last year. And what I've heard from folks who are talking to SMIC and their supply chain, they're still at 20% this year, which is meaning a dreadful outcome for SMIC because some of the machines that they have, they're struggling to get spare parts for them and so they're breaking and they're trying to introduce substitutes and the substitutes don't work very well. So their overall manufacturing capacity is very constrained. Which as, as your paper points out, means that if we don't sell them the chips, they won't have very many of them.
- [52:49]
C
Yeah, and just to point out also compared like 20% compared to what? I think the Nvidia case here is important. So they have about 60 to 80% yield. So that is a vastly superior number. And of those 60 to 80% of the kind of 40 to 20% that are not a perfect yield. They also get used on different other kind of like downgraded chips through kind of binning and stuff like that. So I think there's a huge jump. We're not comparing it to 100% yield because not even Nvidia has that. But these are important.
- [53:22]
B
Yeah, but I think it's just so revealing that SMIC now having been doing 7nm since I think it was July 2022, is when they were confirmed to be in the business of making 7 nanometer chips. It may have even been earlier than that. Their yields are still dreadful. Right. They are not making the kind of progress that TSMC was making when they were doing 7nm five and six years ago. And that tells you that a lot is riding on these export controls, on continuing to be effective in the manufacturing equipment export controls, on continuing to be effective in the, in the chip export controls. Because that kind of gets to the scenario so safe. I want to take us back to this sort of like 31x US advantage versus 1.1x China advantage. Obviously those are the extremes of the scenarios that you modeled. There's other scenarios that you model. But let's just talk about like what would happen to America's AI advantage over China in a world where we have a 31x advantage versus a world where China has a 1.1x advantage.
- [54:24]
A
Yeah, absolutely. And just to note, you know, you can look in the report and see a little bit more on the high bandwidth memory production as well as how many chips US companies are actually getting then getting as a result and how many chips that Huawei and other Chinese chip makers can make. The advantage is enormous, but check out the report to see kind of some of the specific numbers. But it's these numbers that we rely upon to then make the estimates of what the compute advantages would look like. One quick clarification I would note before I kind of get dive into the specific scenarios is in those charts we specifically focused on chips that were made that are expected to be made in 2026.
- [54:57]
B
Yeah.
- [54:58]
A
Where the US is going to make, we estimate around 6.9 million B300 equivalents. And that's kind of normalized for processing power. Whereas China, if you take the average of what they're expected to make, it's less than 200,000. And so the advantage is absolutely enormous. So we kind of take.
- [55:17]
B
And those 200,000 might be lousy.
- [55:19]
A
Yeah, yeah, yeah. So. So we're not considering the installed base that came before, but we don't have this in the paper. But our best estimates, which we'll be publishing more on later, is that the US advantage even today is very large, probably at least 7 or 8x or so. But we'll have more on that later. But then really the key point is that future production flows are really going to be most important for the long term advantage, just because the number of chips that will be made next year in 2026 will actually be more in terms of aggregate compute power than all of the compute that is installed by the end of 2025. And so quickly the numbers that you get for 2026 kind of represent where we're going to be headed depending on various scenarios. So it is in that way a simplified analysis, but it does show, given those very large differences in production numbers that, that, you know, if we don't sell anything to China, they're essentially just reliant on the domestic production, which as I mentioned is very low, then we would expect the rest of chip production from the US to be divvied up between the US and allies. We roughly assume that, you know, 75% of compute would go to the US and the rest of us produced chips would go to partners. And then you get something like that 31x advantage, which is absolutely enormous. And then smuggling, you know, we kind of got to that estimate based on a very recent CNAS analysis of, you know, smuggling that we saw in 2024 and taking that point estimate, which was 140,000 H100 equivalents, and then scaling it up to assume that a similar percentage of U.S. chips would be smuggled, that's the number we get for 2026 that ends up being, you know, 4% and quite a significant amount and actually more than what they can do, than they can produce indigenously, yet still you end up with that 11x advantage. And then we have these three kind of categories of scenarios where essentially the US is allowing exports of B30 as one key assumption here is that essentially what we're doing, and we've been very careful kind of not to describe this as the worst chip, but since it essentially, or a 50% worse chip than US chips, it essentially is, as we discussed, roughly a cost competitive chip. And so you can just expect that we are in a scenario that looks relatively unrestricted if you are just approving these licenses.
- [57:36]
B
Yeah, the 50% degraded is true, but also misleading because you can, you can take two halves and make something almost a whole.
- [57:46]
A
Yeah, right, exactly. So then what we can do then is Essentially look at this in terms of kind of the percentage of US chips that China would be interested in buying if the chips are essentially unrestricted. One metric that we used for the more conservative scenario is if you look pre October 2022 export controls, the percentage of revenue that Nvidia was getting from China was around 26% for its data center chips. And in that world also we are assuming that other US companies that have their own in house silicon are not selling to China. So the actual percentage that is going to China is more in the range of 16% of US chips. And in that scenario, as you can see, you already cut down the advantage to something like Forex. And that's kind of like a baseline of what you would expect from China. Like we're talking about a world before AI was even seen as something that strategic. It was pre chat GPT. Yet in that world, you know, you still had China buying a normal share just because it's such a large economy and it has demand, demand for these levels of chips.
- [58:53]
B
I mean, let's just use Jensen Huang's words here, right? He is frequently pointing out that half of all AI developers on planet Earth are located in China. So you could see a plausible scenario where they want to buy half the chips and we've been constraining them over the past few years so they have some built up demand because they weren't buying as much as they would have liked to over the past few years. So I don't think it's. I think your conservative assumption, which as you said is based on historical, historical pre ChatGPT trends, is way too conservative. I mean, I think your aggressive scenario is the most plausible scenario if we were to go and allow B30A exports.
- [59:29]
A
Yeah, and that's right. So when you go to the baseline scenario that actually kind of takes into account exactly what you're describing, that in reality China's demand is going to be quite enormous at this point. They see AI as a strategic technology. There are going to be some level of subsidy involved. And one of the kind of benchmarks that we used is, as you know, well, Greg, in the semiconductor manufacturing equipment industry, despite there being quite significant U.S. and allied export restrictions in place, it is still the case that many U.S. and allied equipment companies are deriving well over 40% of their revenue from China. And this is aggressive stockpiling behavior from China. They've already been recognizing for the last couple of years what a strategic need they have for.
- [60:13]
B
And when we say stockpiling, there's a couple things I want to say here, number one is, you know, it's not just me and Safe and Georgia hypothesizing that there is stockpiling. The equipment manufacturers are talking about this in their earnings calls with investors about the stockpiling phenomenon. And prior to this, that was almost an unheard of phenomenon in the semiconductor manufacturing equipment industry. When you buy a $100 million machine or a $300 million machine, you want that to be put to work and making money for you as fast as possible. But in the case of China, they're buying all the equipment they can possibly buy even faster than they can install the equipment because they're afraid of future export controls on other equipment and components, which, correct me if I'm wrong, but that's one of the key recommendations of your paper here. In addition to don't sell the B30A, you also make a recommendation of the equipment restrictions that we have right now, although they have had an important strategic impact in restricting the most advanced categories of machines, actually did not go far enough and we should have further restrictions on equipment. Do I understand what you wrote there correctly?
- [61:24]
A
Yeah, that's totally right. The more effective way to kind of limit Huawei's influence in China and its expansion in chip making is not to kind of sell chips to China in the hopes that, you know, that will decrease demand for Huawei chips. We've already talked about how both on the supply and demand side that isn't going to work because on the supply side, Huawei is already limited. They could be limited more, as you said, with further semiconductor manufacturing equipment restrictions. And then on the demand side, you could probably expect that China is going to create artificial demand by restricting imports to critical infrastructure markets. They've done this in other industries, They've.
- [62:04]
B
Done this in this industry. I think it's really worthwhile that there were no restrictions on the sale of Nvidia AI chips in 2020. And that's when Huawei was was preparing to fab a 7 nanometer chip at TSMC before Nvidia was going to have their 7 nanometer chip. And the Chinese government had the 3, 52 policy which directed state owned enterprises and government agencies, which are a much bigger part of their computing ecosystem than is the case over here, they were preparing to order them to buy local. So even when we allowed to sell the latest and greatest Nvidia chips, they were prepared to take measures to create a minimum demand pocket for Huawei. And so this is why I often say China is utterly committed to indigenization. They are utterly committed to what they call self reliance. They are committed to decoupling, frankly, at least in this, in this industry. And the question is, do we want to sell them everything that they need to build a convenient bridge to their decoupled future? Or if they're going to choose decoupling, do we want to make that as expensive and complicated as possible? And I think for me, it's, it's super clear, and I think your paper makes it helpfully super clear of just how large the stakes are and how the, the numbers that are out there, like half of a B30 can be misleading as to just how enormous the stakes are here. So we're gonna, we're gonna call it there just because I, I know you guys have to run to your next meeting and I want to be respectful of your time, but let me once again encourage the audience to actually go check out this paper, if for no other reason than it's got a lot of delightful charts to look at. And also it's a great example of great work. And so I was proud to see an alumni of the Wadwani AI Center, Georgia Adamson, produce such a high performing piece of work. So, Georgia and Saif, thank you so much for coming on the AI Policy Podcast.
- [63:59]
C
Thank you. Thanks for having us.
- [64:01]
A
Thanks, Greg.
- [64:05]
B
Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mand. See you next time.