← Sinica Podcast
Sinica Podcast
Paul Triolo on Nvidia H200s, Chinese EUV Breakthroughs, and the Collapse of the Sullivan Doctrine
01:24:50
Loading summary
Transcript146 lines
- [00:00]
A
Foreign.
- [00:09]
B
Welcome to the Seneca Podcast, a weekly discussion of current affairs in China. In this program we look at books, ideas, new research, intellectual currents, and cultural trends that can help us better understand what's happening in China's politics, foreign relations, economics and society. Join me each week for in depth conversations that shed more light on and bring less heat to how we think and talk about China. I'm Kaiser Guo, coming to you this week from my home in Chapel Hill, North Carolina. Sinica is supported this year by the center for East Asian Studies at the University of Wisconsin, Madison, a national resource center for the study of East Asia. The Sinica Podcast will remain free, but if you work for an organization that believes in what I'm doing with the show and with the newsletter, please do consider lending your support. You you can reach me@sinicapodmail.com and listeners. Please support my work by becoming a paying subscriber at. You'll enjoy, in addition to the podcast, the complete transcript of the show, essays from me, as well as writings and podcasts from some of your favorite China focused columnists and commentators. And of course, you will be able to enjoy the knowledge that you're helping me do what I honestly believe is important work, and so check out the page to see all that's on offer. Do consider helping me out for the last three years, US Export controls on advanced AI chips have been treated as a cornerstone really of America's China strategy. Export controls actually date back further than that, to 2018, in most tellings, when the first Trump administration forbade the sale of equipment used in the manufacture of advanced semiconductors to China and took actions against companies ZTE and Huawei. The Biden administration, of course, as all listeners are doubtless well aware, tightened the screws, imposing bans of sales of what it deemed to be more advanced semiconductors, as well as the machinery used to manufacture those, and forbade US Persons from working on advanced semiconductors in China. The following year, it expanded the restrictions to include even more classes of semiconductors. And so things stood. And so things stood when just weeks after the Busan summit in South Korea with Chinese leader Xi Jinping, Donald Trump took to Truth social on Monday, December 8th to announce what he would I'm sorry, I'm saying that again. And so things stood when just weeks after the Busan summit in South Korea with Chinese leader Xi Jinping. Don't Donald Trump took to Truth social on Monday, December 8th to announce that he would approve Nvidia's H200 sales to vetted Chinese customers. Many observers were quite surprised Though perhaps less so by the fact that he would be taxing the chips at 25%. That would of course go directly to the US Treasury. The lifting of the ban, though, cut directly against the logic of export controls. The though Trump had previously lifted such controls on other Nvidia chips or talked about it, namely the H20 chips. Immediately a heated debate broke out, making for what might look at first like strange bedfellows across partisan lines. You had Democratic China hawks decrying the decision and actually many Republicans, whose party has traditionally leaned much more hawkish when it comes to China, actually defending Trump's decision. The policy shift on chip exports raised many, many questions. Did this come out of the Busan summit somehow, even if it wasn't explicitly discussed there? Does this herald the beginning of the end of all the major export controls on tech? Would Chinese firms even want to buy the H200? Would Beijing allow them to import the H200s in significant quantities if they want them? Given that it's trying to achieve a kind of technology autarky? It struck me quickly that how one came down or Trump's decision depended very much on one's assessment of how far Chinese companies had come toward closing the gap since the export controls began. How far are they still from achieving genuine independence from US controlled supply chains? How they assess China's homegrown generative AI models, how much importance an observer accords to AI in this whole winning the future business. And even more fundamental questions about the bilateral relationship itself, about China's actual ambitions and intentions, about their explicit or implicit visions for America's place in the world, for American grand strategy. I realized just I was thinking about just so far I have not seen a single person change their minds after Trump's announcement. That is, if they were pro export controls before, they were still, you know, against what Trump did, and if they were against export controls, then they were still against these export controls and applauded what Trump did. So today, just a couple of days after Reuters has reported that China has now cracked ASML's EUV lithography code, I am joined by my good friend Paul Triolo to unpack why this decision on the Nvidia H2 hundreds was made, why it's provoked such strong reactions, and what it tells us about the future of technology export controls for China. Paul is Senior Vice President for China and Technology Policy Lead at DJA Albright Stonebridge Group. He's a non resident Honorary Senior Fellow on Technology at the Asia Society Policy Institute's center for China Analysis. Paul spent 25 years or more in the US government and he's got some good, strong opinions, well informed opinions, I should hasten to add. But like me, he's made no secret about how he felt about Trump won and Biden policies to try and limit tech exports to China, especially in semiconductors and related technologies. Anyone who reads his excellent substack, which I will link to the show notes here, or who's heard him on any of his many appearances on this program will already know that. Paul Triolo man, welcome back to Sinica.
- [06:02]
A
Hey, great to be back, Kaiser. And you've laid out quite a quite a table here for us to discuss here. The timing couldn't be better after the after the Reuters story on the euv, but it struck me as you, as you were in your intro, that December 8th may be a day that will live in infamy because of the Trump True social post. So great summary of where we are.
- [06:24]
B
Infamy is probably not a word either you or I would use though, if I read you correctly. But before we get into arguments and reactions, let's sort of set the stakes for us in concrete terms. What does this decision actually change in the real world and what do you think it mostly doesn't change?
- [06:45]
A
Well, it's a great question in the real world, what it changes is in the short term, the ability of some number of Chinese AI developers to continue to use U.S. technology to train their models and run inference on those models at some scale in China and potentially outside China, at data centers outside China. What it doesn't do, though, is give China, as many have asserted, some strategic advantage in AI writ large, or give China the most advanced US Technology in AI, as some have asserted. It gives them a generation of GPUs that is still widely used and is widely capable and is certainly better than what most companies in China have access to. But it's not giving them the Blackwell architecture, GPUs or beyond. So it's an important development. But I think some of the hyperbole around it, which we should discuss is really hyperbole. Right. I mean, and that's something we need to unpack a little bit.
- [07:55]
B
Right. So as you hinted, this is not their most bleeding edge technology. This is not their Blackwell. This is 18 months old. Right. This is how it's been described or I've seen it. Is that accurate?
- [08:08]
A
Yeah, I mean, roughly, the Harper architecture became widely used in 2024 timeframe. And so it's important to remember that these architectures have a sort of lifetime. Most people think that you know, these are going to be useful for roughly four to six years from the time they were developed before the subsequent architectures start being more widely used and companies eventually switch over. But they're still going to be very useful into the future. So they're not, they don't suddenly become obsolete because the next generation comes along.
- [08:39]
B
Right.
- [08:39]
A
So it's a big deal in that sense that these are, these are very capable chips that can be used for the next three to four to five years.
- [08:46]
B
So Paul, maybe before we head too much into this, let's rewind a bit and have you walk us through how US chip controls evolved from the entity list expansions under Trump's first term through the October 2022 rules, the 2023 revisions, and maybe the diffusion logic late in the Biden administration, the AI diffusion rule, maybe against that backdrop, how should we understand the H200 approval? Is it a reversal or refinement or a one off exception? But let's, let's first before we revisit that question, let's make sure we have the timeline correct.
- [09:24]
A
Yeah, I think that's important. And you mentioned the use of technology controls under the first Trump administration targeting Huawei and zte. So that was the first sort of widespread use of the entity list again, which is this dreaded tool of the Commerce Department which had heretofore been used primarily for companies engaged in supplying various types of inputs for weapons of mass destruction. And so the use of the, of this tool against ZTE and then Huawei was the first time it had been used against a nominally civilian private sector company. That, I mean, of course ZTE was at the time more state owned, more importantly, a company that had a broad and complicated global supply chain including many inputs from many, many U.S. companies. So that was a huge, a huge step and that was really predicated on the, the behavior of those companies. Huawei always business model was accused of theft of ip. And in the case of zte, actually the proximate cause was, was Iran sanctions. Breaking Iran sanctions.
- [10:26]
B
Right, right.
- [10:27]
A
So anyway, that was the first use of this. And then as we got into the Biden administration and again, remember 2018, 2020, nobody's talking about AI. It was all about 5G.
- [10:35]
B
Right.
- [10:36]
A
The real reason behind Huawei, for example, was the desire for the US to prevent Huawei from dominating 5G. Nobody was talking about AI. Then as we got into the 20s, AI started becoming an issue and the use of advanced compute became an issue because advanced compute in general high performance computers, some using GPUs at the time for acceleration became an issue after the, the Chinese tested this hypersonic missile. And that, that was one of the, one of the triggers. And so as we got into the 2022, the issue of what to do about advanced compute and China became a very salient issue. What should the US government do about this? At the same time we had Jake Sullivan finally in 2022, in the fall of 2022, articulate a US government policy on technology in China. And this is what I've called the Sullivan Doctrine. And this, this.
- [11:29]
B
Yeah, that name is stuck. Yeah, yeah, a lot of people call it that now.
- [11:32]
A
And this is important because it had these three pillars. One was the, the US would maintain an absolute lead over China in these key technologies, including advanced compute. And the other one was a small yard high fence which we all know and love. And the other was that these three sectors, advanced compute, biotech and green tech, were, you know, were of high concern to US national security. So that, that sort of set the stage then because that established critically this idea of an absolute lead. And that's what the export controls on October 2022, October 2023 and December 2024, the large packages of controls including on GPU' were designed to sort of freeze China in place at these nodes. For example, in manufacturing that were laid out the end use controls at 16 and 14 nanometers, 128 layer for NAND and 18 nanometer pitch for DRAM. So there was an attempt, so the Sullivan Dr. Was translated into these export controls that were designed to sort of freeze China in place in terms of GPU performance as we remember the threshold, the performance thresholds that were established and in October 2022 and then changed in 2023, which, which caught the, the H1 hundreds and the, and the H. The A1 hundreds which, which had been approved and then, and then of course Nvidia and other US companies, including AMD intel, designed up to those thresholds because the thresholds were drawn in a certain way. And so that, that was sort of the cat and mouse game. That's important background because that's one of the things that Trump essentially shot down in the Truth social post. He said that essentially that the policies that eventually forced Nvidia to downgrade or degrade its GPUs, that game was over. That era was over. That was that key part of his, of his tweet. And so the other. So the original question that the big difference now is that it looks like we're going back to a sliding scale in terms of this compute technology because Trump has essentially said we're going to allow the exports of this H200. And he was very careful to stress the national security issue in his true social post. So it's not a lifting of all controls. It's a return to a still quite, not quite defined period of a sliding scale with respect to this exports of advanced compute to China. And that's why it's important.
- [13:50]
B
We've seen no real clarification in the day since either. No clarification on what exactly will be and won't be allowed.
- [13:56]
A
Right. If we step back for a moment and maybe it's a good time to do that. The driver here, you know, how did we get to Trump making that true social tweet? It's important to understand that outline because it started at the very beginning of the Trump administration earlier this year. We're still in 2025. And it started with Trump bringing in people like David Sachs, the current White House Aizar, and other industry people into the administration. A sort of deliberate attempt to bring in industry. If you remember, the Biden administration was very hostile to industry and did and in fact vetted people, you know, and, and rejected people with industry experience. So the Trump administration made a sort of conscious effort to bring in people in industry. And of course, Trump had sort of cozied up to David Sachs, been to his home during the campaign. And so he created this new position of aizar. And the first time the US had had this. And, and Sachs is very well connected in the industry. He understands technology. And so he over time became more influential right off the bat. He didn't have an influence. And so if you remember, In April, the U.S. commerce Department issued letters to companies saying, hey, you need, you now need a license for the H20 GPU, which was the downgraded version that had been available in China for almost 18 months after the October 2023 controls. So that came out sort of a surprise to people, right? Like you're now you're re controlling H20. And that was largely driven by this idea that inference now was, was really important. In October of 2023, people were, were very focused on training models and the kind of compute you needed to train models. But then with the advent of deep seq, we were in an era suddenly where inference, now people realize inference is important. And the H20 turned out to be really, really good for inference and for sort of post training and post running of models. And so the US Decided to control that. But that was before Sachs sort of began asserting his influence. So then as we got into the trade talks and really I think the H20 and H200 issue was separate from the trade talks. Right. Because this was really driven by Sachs and Jensen Huang also becoming more active in Washington. And also this idea that Nvidia as a result of all this, particularly after the H20 issue in April, was essentially cut off from the Chinese market. They had gone from 95% of the data center market in China for compute to zero. And the H20s also forced them to take a $5 billion hit because they were sitting on a write off money which only could only be sold to China. Nobody else would buy that. Right. So that.
- [16:29]
B
Because it was deliberately downgraded.
- [16:30]
A
Right, right. And all of that got, got the attention of people like Sachs in Washington. And also again, Jensen Huang was able to sort of team up with Sachs to begin making the argument, which is critical to the getting to the H200, that the US shouldn't cede the market in China to Chinese companies and that US AI companies should be allowed to compete in China so that Huawei and other domestic Chinese companies didn't over time come to completely dominate the AI stack in China. And Jensen uses figures like the Chinese have more than half of the world's AI developers. And so that was a, that's a, was a market that he contended the US shouldn't concede to China. And so that argument was much discussed, coincided with the trade discussions, but it was a really, a parallel conversation and it didn't really intersect with the discussions. There was no negotiation with China over this. It was really Jensen and Sachs trying to convince Trump that this was a good idea and that that's what eventually led to then the, the H200 decision. There was a Blackwell diversion in the middle of that. This B30 diversion, which really was a chimera. There was no B30 chip. Nvidia never actually designed such a chip, but it was sort of out there as a concept. But they never actually designed it or fabricated it at TSMC as they did the H20 on the flight to Busan. Apparently Marco Rubio and Jameson Greer talked Trump out of doing anything on the Blackwells, but it was really sort of a, a non issue because there was no, there was nothing to really discuss there. So that also ended up resulting in this H200 decision because this was an actual GPU, it's being manufactured, it's not downgraded. And eventually Trump decided that the argument that US companies like Nvidia and he's, he's very Fond of Jensen should be able to continue to compete in China and not with downgraded GPUs, but with. With one generation back GPUs. And that's an important inflection point on December 8th that you started with.
- [18:28]
B
Paul, I wouldn't be wrong in assuming that were you in the room with Jensen Huang or with David Sacks and President Trump, you would have taken their side, Is that correct?
- [18:39]
A
Yes, I think I would have taken. I would have definitely taken their side because I think that again, the driving force behind this, which we. Again, stepping back even further, why were we controlling advanced computer? It began as this attempt justified by sort of military use. The idea that China would use these GPUs, would use advanced COMPUTE for military end uses, that was sort of the original justification.
- [19:06]
B
But, Paul, before you go on, let me make sure to do something here. The people who were pushing this idea, they're serious people, they're in D.C. they're in the Valley, and many places in between, they believed. They still believe now that approving H200 sales was a strategic mistake. So let's try to make sure to articulate their argument in good faith as they would recognize it, not as a caricature.
- [19:33]
A
And there are multiple elements to it. Okay, yeah, that's absolutely a good idea. And let me try to do that. If you look at the language in the rules that were issued by the Commerce Department, the original rules were predicated on things like civ Mil fusion, civilian military fusion. Right, right. And the idea that eventually the GPUs could be used to design models that could be used for military purposes. Right. And that's a valid concern. Right. But the bigger picture behind this was really the idea that the US Needed to slow down China's. The ability of Chinese companies to develop advanced AI, because the US Needed to get to AI first. Advanced AI, artificial general intelligence and artificial superintelligence. That was the real driving force. If you listen to people like Ben Buchanan, who was the White House AI czar under Biden. If you talk to think tanks in D.C. like at Rand, this was the real driving force. And this came out of. Of a process over time. This idea that advanced compute, so called compute governance, needed to be of high concern to the U.S. government. That concept started in 2017, 2018. Nick Bostrom. And it was really the idea that, hey, anybody who, as AI develops, if you can have a malicious actor, you know, a cyber actor or other bad actors, and they get a hold of AI, they could do bad things. Right. And so that's Again, a valid concern. But at some point the focus of that, the idea that you needed to have COMPUTE governance turned completely to China. That the idea that China shouldn't get access to advanced AI because China could misuse AI. And that became the eventually, if you look at people like Dairy Amadei, the CEO of Anthropic in that timeframe, 2017, 2018, he was for collaboration between the US and China and not having a technology cold war. But he has now completely done 180 degrees and, and now is for the export control.
- [21:35]
B
Not long before that he had actually worked at Baidu.
- [21:37]
A
Right, right. But also he has, he wrote in his, in his essay, his famous essay last year in October that he essentially said that democratic AI needs to win out over authoritarian AI, get there first and then that could, that should be used essentially to force regime change in authoritarian countries, meaning China. Right. So that is the sort of the overall, that's the argument that is that, that the proponents of the, of the choke point approach, not all of them.
- [22:08]
B
Go all the way to regime change.
- [22:09]
A
No, no, no, no, I'm not saying that that's the, that all of them are going there. But they all. But the, but most of them argue that the US should get to advanced artificial intelligence first before China. And so so, and in fact, Chris McGuire yesterday, a former Biden administration official and others have have contended that the purpose of the export controls was to sort of buy the US time to get ahead on AI because that's going to be such a decisive moment. And the argument there is that whoever gets there first will have a so called decisive strategic advantage. Again, this is a concept that's been around for a while, but it's been applied now here, very much so to China. So that's the argument is these controls are needed to slow China to enable the US to get there first on advanced artificial intelligence. And that's an interesting argument, but there are a lot of problems with that argument. And I can, I'll just note two quickly. Yeah, one is we don't really know when we're going to get to that level. There's a big dispute in the industry. There was a paper written last year called AI 2027 which laid out a scenario for the US and China going down the road of getting to AGI or ASI first. But they've now moved that out to 2028. And even that, you know, people in the industry are sort of skeptical that we're going to wake up in 2028 and be in the age of Artificial superintelligence. And the other thing is my personal. The big thing I objection I have to that idea is that there's this assumption that when we get there, governments are going to immediately use that capability to do something bad. And so, for example, Ben Buchanan in a podcast with Ezra Klein earlier this year, Ezra said, well, what does the world look like if China gets to AGI first? And Ben Buchanan went right away to they would take down our critical infrastructure using cyber means. And to me, that's where the argument breaks down is that I think governments would have a much more complicated decision and process around this idea of what to do and how to wield a really advanced AI capability. And there's not enough discussion, in my view, of what that means. There's an assumption that whoever gets there first would sort of rush to do something against the loser. And I just don't think that's the way that the world will unfold when we get closer to AGI or asi.
- [24:30]
B
Right, right, right. As I said in my introduction, it all seems to sort of rest on what one thinks about China's actual ambitions or intentions in the world. Right. It all really does come down to that. But, you know, they did see these constraints on Huawei, these real, meaningful constraints on Huawei as evidence that controls were working. Right. They saw spillover risks to national security. And so, I mean, by those assumptions, it makes sense that exporting H2 hundreds by compressing China's AI development timeline will, you know, bring about this. This possibility that China will arrive at ASI or AGI before we do. And that's inherently a bad thing. So by their own logic, that makes sense. Leave aside the fact that China has no real interest so far in the sprint toward ASI or AGI. I mean, it seems like they're much more interested in diffusing it as it is right now through enterprise and into all nooks and crannies of ordinary society. They're really into this whole diffusion argument that Jeff Ding argued, Right?
- [25:37]
A
Yeah, exactly. I think that now there's this idea out there, and again, I'm not sure I totally agree with all of the argument is that, okay, the U.S. is sort of, you know, AGI pilled. And so the goal of the major labs is to get the AGI first. And so they're rushing headlong towards that and building massive data centers with lots of compute capacity. And then China is sort of taking this more pragmatic approach. And when I've been to China, I've talked to many companies and they all talk about, they're really trying to address pain points and trying to use leverage AI to, you know, solve real world problems. I think that we have to be careful. I think that's also happening in the U.S. sure. U.S. companies are using AI coding tools, for example, are widely used. So it's not as though that every company in the US is driving towards AGI to the exclusion of deploying AI in practical terms. But in general, I think it's more the sort of the think tank world and the sort of strategic competition world where the issue of who gets the first to some advanced level of AI is being discussed in China. You just don't see that. You don't see the academic community or the safety community in China or the great power competition commentators in China talking in those terms. Right. It's just not, it's not something that I think is, is top of mind where, as in the west, there's a lot of more concern about doom, you know, about the existential risks around AI. And that translates also into a big discussion about who gets to AI first, advanced AI first, and that drags in China right away. But it's, it's definitely not that the level of discussion in China the same.
- [27:12]
B
Issue, not the same as we see here in the United States. Let's get technical, a little bit grounded in the real world implications. We've already got a pretty good idea of where the H200 sit in Nvidia's lineup. I mean, relative to the H20, the H100, the A100, Blackwell, what does it practically enable? So let's assume that there will be Chinese buyers. Is this mostly about training frontier models? Is it about improving inference efficiency? Is it about stabilizing existing large clusters? In your view, would relatively unfettered access to the H200 represent a lateral capability gain or something closer to a step change? This is something you addressed in that, in that paper.
- [27:54]
A
That's a great question. So, you know, again, thinking back up to how we got here, so we do have these. The example before of the H20, the H20 turned out to be really, really good for inference, as I noted. And so Chinese companies like Bytedance and Alibaba and Tencent all bought millions of the H20 because as demand in China went up for these models and use of these models and applications, the H20 proved to be really good for inference. And so that's what it was used for. Those companies continued to use older but non degraded GPUs for training things like the A100 because the. There was a time when the A100 was legal. There was a time when the H800 and the H and the A800 were legal. These other downgraded versions that were still pretty good. So Chinese companies have this very complicated mix now of hardware, including Nvidia hardware. Many of them are also experimenting with Huawei hardware and also with some of the new GPUs from companies like More Threads, Metax and Inflame, these sort of smaller players. And so when you go to China, I was. And I was there in September and also there in July. And you talk to companies, many of them have this sort of heterogeneous hardware layer, and they're working with all of this different hardware. So the H200, though, is going to be important because those companies are kind of running up against the lifetime, if you will, of training for systems like the A100. That's now pretty old. That's like more than. Almost five years old. More than five years old. And so. So they're at a sort of inflection point where the H2 hundreds will be a pretty big improvement in terms for training, basically, mostly for training. And that's where the focus will be, for example, the memory bandwidth and the sort of interconnectivity bandwidth, which is one of the things that was limited, of course, by the US export controls. The H20 was essentially a downgraded version of the H200, which degraded primarily that GPU to GPU transfer rate. And then it turns out that when you're training models, you want to have very high bandwidth between the GPUs, because you're doing a lot of parallel processing.
- [30:00]
B
Right, right.
- [30:01]
A
And so that's where the H1 will be particularly useful, will be in training more advanced models.
- [30:08]
B
Okay.
- [30:08]
A
I would expect Alibaba and Tencent and others to do this. Now, the situation is a little more complicated because as all this has been happening over the last year, as we know, and I'm sure you've talked about before on the show, there's been. There's been some smuggling or diversion of some GPUs to China, although I think it's probably been overstated a little bit just before. And again, I think the. One of the classic things is just before Trump's true Social was announced, that the Justice Department, you know, indicted these guys for smuggling H1 hundreds and H2 hundreds to China. And I found it really interesting that they. And I don't know where this actually came from, but the Justice Department indictment used this incredible quote. It said these chips are the building blocks of AI superiority and are integral to modern military applications, which is not really true at all. The country that controls these chips will control AI technology. The country that controls AI technology will control the future. Well, come on. I mean, I found that statement to be. I'm sorry, but it was really sort of like, really? These are not the most advanced chips, so they're not the building blocks of AI superiority. They're not really the building blocks of military applications. The US still controls these chips, and these are not the most advanced chips. And so it's a really way over the top assertion in this Justice Department indictment. It's just wrong. I mean, every. Every sentence in that is not accurate. And so who told them to put that in there? Right. And then, of course, the next day or two days later, Trump says, Here, we're going to ship H2 hundreds to China. Right?
- [31:42]
B
That very day. Yeah, that very day.
- [31:44]
A
So. So I think that's a really important thing. And people have pointed to that is like, wow, that's a real disconnect. But the problem is that that statement is wrong. And then the smuggling that was involved here was really, again, sort of small potatoes. These things are going to matter in large numbers, you know, large clusters. And so the smuggling is, you know, around the edges. There are probably some labs and other small players that are using the smuggled stuff, but now they can get access potentially to H2 hundreds, even though the Commerce Department will have a licensing problem. And then the last thing is that over that last year, it turns out there's no restriction, for example, for Chinese companies to use overseas data centers that have Blackwells in them to train their models. So there's some evidence that Chinese companies are taking advantage of that. You know, it's been described as a loophole. But it gets back to that AI diffusion rule that you mentioned earlier. Currently, there is no legal prohibition for Chinese companies to use overseas data centers to train their models, including on Blackwell architecture systems, like in Japan, in Southeast Asia, and even potentially in Europe. So that's another thing complicating the picture, because Chinese companies could. Their appetite for buying these H2 hundreds, for example, are going to be conditioned by the Chinese government's view of the situation, their own needs for training, and. And then their own sort of mix of hardware that they have access to, particularly for training their most advanced model.
- [33:09]
B
And that's the topic I want to get to next, because it is a very, very complicated picture, trying to suss out how much real demand there will be for H2 hundreds among these different companies because as you say, well, for one, they can train on data centers abroad that have the Blackwell capability that are dense with them. On the other hand, the clearly recognize that when it comes to their incumbent networks, their neural nets, they're at the limit of their capabilities right now and they need to upgrade. So these things seem to be pulling in opposite directions. Then of course there's the Chinese government which has these aspirations of achieving genuine sort of indigenous innovation. So, yeah, it's very complicated. So when you have all these vectors pulling in different directions, how does it all work, work out? Let's, let's add to this the fact that, you know, for these companies, it's not an abstract choice between, you know, Nvidia versus some Chinese champion. Because, okay, yes, there's inflame, there's Meta X, there's all these others, but for the most part, we know who the Chinese champion is that we're talking about. And that's going to be Huawei in their censorship set. And not everyone loves Huawei in China. Right? I mean, so as far as you can tell, how likely is it that, that, I mean, we saw what happened when we heard these early announcements about H20s coming back on. We had Howard Lutnick go out and say something which I thought was a little bit, you know, impolitic. Just talking about how the goal is now to addict China to Western GPUs so that they, they will become dependent on. And this didn't go down well in Beijing at all.
- [34:44]
A
Right. That angered a lot of people in Beijing. Yeah, so that's a great question. So this issue, and some have actually sort of, some of the conspiracy theories that I've seen online suggest that the Chinese decided to sort of, you know, restrict H20s, holding out the hope that Trump would approve, you know, the H200. I think that's sort of, that's too, that's too cute by a mile. Because I think that the Chinese government again, was not, this wasn't a negotiation and they were, of course they would like the, some of the controls lifted, but I don't think they were deliberately restricting H20s in the hope of getting something else. So anyway, so they, but, but, but now companies in China do have this complicated decision. And I think there's a couple things to, just to keep in mind. One is, you know, no company, even in the west wants to rely on one hardware platform. So you have companies like Google and they have their own TPUs that are used internally for some workloads. They're they're now potentially, potentially could sell those to other companies. They use Nvidia GPUs for some workloads. And they will also be using companies like Groq and 10 storen and Sambanova and Cerberus that are designing GPUs for inference. And so as we're also sort of in this weird transition between companies that are devoting most of their compute to training to a world where we're going to have much more inference. And so that's, the inference world can require a lot more different hardware. And so companies are going to have different sort of heterogeneous hardware stacks. But in China it's complicated because Huawei, as you note, Huawei has been trying to upgrade its manufacturing capabilities, working with SMIC to crank out more Ascend processors to meet the demand and to be a viable alternative to Nvidia. But they're not quite there yet for a couple of reasons. One is the Nvidia ascend design, the 910B and the 910C, which is just really two 910Bs, is not really a general purpose GPU graphics processor unit. It's an npu, a neural processing unit or a network processing unit. And that means the architecture is different. And so when you're trying to transition, for example as an AI developer from an Nvidia environment, which Nvidia GPUs are true GPGPU with CUDA which is optimized for the hardware that you're working on.
- [37:06]
B
Right.
- [37:06]
A
With Huawei you have this non GPGPU and then they've developed their own CUDA like software environment called Can C A N N And, and so when you're trying to, when the Chinese government is urging or encouraging, it adds friction to the development.
- [37:23]
B
I mean, it makes it difficult to transition to another system. Right, right, right, right.
- [37:27]
A
So, so the Huawei alternative is not really an equivalent alternative. Right. So therefore, if you're a company in China now and you, and you're, you're sort of hoping that over time now Huawei is redesigning the SEN to be a gpgpu, but that's a process that will take probably, you know, 18 months to two years.
- [37:45]
B
Just so people are clear, you mean general purpose gpu.
- [37:47]
A
General purpose gpu. Right. And when that, when that happens, and then by that time the software development environments in China will be better, then there eventually will be an alternative. So actually the way to look at this is the H2 hundreds now will provide this sort of bridge to a potential more viable domestic alternative. Future. And so that's why I think Chinese companies like, particularly the major model developers like Alibaba, ByteDance, Tencent, Baidu, even Deepseek, their model developers are all very familiar with Nvidia hardware already.
- [38:23]
B
Right, right, right.
- [38:25]
A
And they will be happy to train the models on those H2 hundreds. That doesn't mean, though it's important to note as some sort of allege that, that the, that those companies will abandon the alternatives. Right. They will continue to work with Huawei and do work to develop the Huawei software and hardware stack and potentially run inference workloads, for example, on the Huawei hardware, which is good for that. Right.
- [38:48]
B
What's their incentive to do that?
- [38:50]
A
Their incentive is that when the government is deciding what to do about allowing imports of Nvidia hardware, they visit the company. So when I was in China last time, I understood that Chinese ministries like MIT would visit companies and just make sure that the hardware stack included both Huawei and some of the other smaller GPU developers. And in fact, when I was at the World AI Conference in July, all the second tier cloud providers that I talked to, Silicon Flow and, and Infigenics are these two really good second tier cloud providers. And they, on their screen they showed their stack and on the hardware stack was Huawei, Nvidia and Metax and Bran and Hyggen of other Chinese companies. So the Chinese companies again are not going to abandon necessarily domestic competitors of Nvidia. So they're going to have to do both. But right now, and probably for the next 18 months to two years, there will not be a real equivalent hardware software stack that will make it easier to make a transition to a domestic focused AI development environment. And that's, that's sort of the reality, right? I mean Huawei is making great strides and we can talk about, you know, this, the recent EUV report.
- [40:06]
B
Yeah, we'll get to that in a second. But just, I mean, just to sum up, so you basically think that there will be a pretty eager market for the H200 then?
- [40:14]
A
Yes.
- [40:15]
B
Okay.
- [40:15]
A
And David Sachs, by the way, just real quick, David Sachs, about two or three days ago, there was an FT report that said that the government was going to somehow control, you know, or restrict companies from buying the H200s. David Sachs said, oh, they're refusing our, they're refusing our chips again, right? Because he, because again his argument is that we should allow sales of some of these GPUs so that China will use them and then they will not want to use Huawei. And again, the opponents of that decision, the decision to allow the H2 hundreds say, well, look, this is dumb because China is eventually going to not allow US hardware and they'll eventually go to a domestic, fully domestic stack. And so David Sachs I think was wrong because that story I think was premature. And talking to companies in China, as I do pretty regularly, there's clearly they will want to buy this. The government may restrict access to SOEs and some government, government entities may limit their ability to buy the H2 hundreds. But for the commercial companies and where the action is in China, Right, Alibaba, the usual suspects, they will want to buy large numbers of these. Yes. And I think Nvidia is already talking with TSMC on Taiwan because the H2 hundreds were being ramped down, the production was being ramped down and now they have to ramp back up. And that's the bottleneck there is packaging, advanced packaging chip on wafer on, substrate is everybody wants the, you know, the GPUs that TSMC is able to produce for companies. And so they need to figure out how to book some of that packaging capacity. So realistically, we're probably talking March, April, May, June before appreciable numbers of these H2 hundreds get to Chinese companies get in their hands and they start training models on them. So the first models we will see trained on the H2 hundreds will not be likely until next summer.
- [42:03]
B
There are some people who argue that if China were to allow imports of the H200 or of other chips, hopefully down the line, it could actually accelerate domestic innovation by setting a clear benchmark. They have done this often in the past. They've allowed foreign competitors in the market because they know that the old Tesla story, the Chinese EV game, wouldn't be what it is were Tesla not in the market for so long. Apple and the Chinese handset game. Right. Do you buy this idea?
- [42:36]
A
You know, I tend to operate from, you know, from discussions with companies and what the reality of sort of is on the ground here. And I think that here we need to just sort of use Occam's razor, you know, pretty liberally here. Right. These companies are not, and the Chinese government is also, you know, not this sort of going through these sophisticated calculations here trying to figure out how to do all this stuff. I mean, I think it's funny in some sense that one of the arguments that a former Trump and a former Biden administration administration recently made is that, you know, it's the, they say in the, in the, in a recent op ed it's the same Pattern that whether the product is solar panels, electric vehicles or telecom, China imports Western technology until its own production catches up, then it cuts out American companies. No, this is this AI and semiconductors are different. They're not solar panels, they're not electric vehicles, they're not telecommunications. They're much different in my view. Right. And so, for example, left to their own devices with no controls, Chinese companies would continue to be buying the most advanced tools to make semiconductors wherever they are manufactured. The us, Japan, the Netherlands. Right. Because the cost of building something from scratch is intense. I was in Shenzhen talking to Huawei just after they had really made a big investment in the Harmony operating system. And I asked a very senior Huawei official what about the Harmony operating system? He says, why would we have done this if we could just have built on Android? But we were forced to do that, of course, because the controls included Google Mobility Services making Huawei's Google version not very saleable. So they were forced into that, but they were forced to invest a lot of money with no return right into developing an operating system from scratch. And so again, the idea that is raised in this op ed that the plan all along was to import technology, learn how to do it and then kick companies out. That's just not the case in semiconductor manufacturing and arguably in GPUs and AI. If Nvidia had been allowed to sell its GPUs into China, then China would be, would 95% of China's data centers would be running Nvidia GPUs and Huawei would be, would be struggling to compete. The bigger issue is that China has done like with Tesla, was to invite Western companies in to then jumpstart their own industry and improve the quality of their domestic competitors. But again, they didn't, they haven't kicked out Tesla. Tesla's still there and Tesla's still doing quite well in China. So that, that, I think that model worked in the EV space because that was a particular sector. I don't think that model is applicable across the board. And neither is the, you know, the strategy that you raised there. It's, each sector is different and semiconductors and AI are very different than solar panels. And granted. Yeah, and so try to make a blanket statement like that. To me it's just, it misunderstands the technology and you know, what the Chinese government thinking is about these technologies.
- [45:36]
B
I have seen the argument raised though. Let's zoom out for a second. I mean one of the stronger claims that you make is about consistency of US policy. So if The US Is willing now, as they seem to be, to export 5 nanometer class GPUs with very advanced memory. Why do semiconductor manufacturing equipment controls remain frozen at much, much older definitions of what constitutes advanced? Is this in your mind just bureaucratic inertia or does it reflect a more deeper, unresolved strategic contradiction?
- [46:07]
A
Great question, great question. So, yes, I raised this in one of my latest substacks and this is not my argument. This is the argument of the broader industry, which is, okay, if you're going to set, as the US did in the Commerce Department did in October 2020, lines around the technology and define, for example, advanced node production as a problem and then define advanced node as 16 and 14 nanometers. You know, that's, that's a problem from the industry point of view. In October 2022, nobody in the industry thought that 16 and 14 nm was represented advanced node production. But again, under the Sullivan doctrine, the idea was to draw these lines in the sand and then say, okay, no matter what the technology does, that's going to be considered advanced node production. So that has created a whole host of problems because even at the time those end use controls were drawn, Chinese companies were already getting close or beyond those end use controls. And now every fab in China that's restricted because of those controls is in fact manufacturing beyond those end use controls. Right. And so that's, that's that. So the argument now is, okay, if you're going to, if the Trump administration is going to go back to a sliding scale here on the GPUs. And again, the H2 hundreds are manufactured at TSMC, a 5 nanometer process with HBM3.
- [47:33]
B
What does that mean?
- [47:34]
A
HBM3, high bandwidth memory. HBM3 is a pretty advanced, we're now on HBM4, but it's a pretty advanced process. So right now US companies, for example, can't sell tools that would enable Chinese companies to make high bandwidth memory. And also of course, their advanced node is still defined as 16 nanometers. Right. But we're selling China a 5 nanometer GPU. So what the industry is going to argue next year, and this will be a really interesting debate. And already those who support the export controls are worried about this and are saying, you know, is this next is the tool the tool we should. And actually many of them are saying we should double down because China is making progress and because we're selling them the H2 hundreds, we should be even more strict with the tool controls. But the argument of the tool industry is that you know, all you've done here is, you know, you've, you're selling China advanced GPUs, and then by setting these, these thresholds so high, you're restricting US Companies from access to more mature nodes in the China market. And in fact, for example, BIS is sitting on licenses for stuff that's 90 nanometers, which is nowhere even near advanced control. So the problem is you're going to have to have a whole strategic rethink on what the policy is here, because in the Biden administration, you had a group at the National Security Council that determined the policy and was very meticulous in trying to force the policy through and update the policy in October 2022, 23 and December 2024, and a lot of entity listings in between. So now you have no functioning similar body at the National Security Council. You have the White House and Trump wanting to make deals with China and wanting to avoid having new export controls mess up the deal and impact China's willingness to sell rare earths and magnets to the US So there's no clarity like who is going to decide, because Sachs, I think, was very, it was a little easier argument to make with Jensen Huang, which is, okay, we have these H2 hundreds, let's sell China this generation of GPUs now, when the tool thing is much more complicated, too, because there's a lot more tools, and then again, there'll be a lot more pushback in Washington from the proponents of the export controls on the tools. But again, the same argument would hold, right? You're disadvantaging Nvidia by making this absolute prohibition or forcing them to degrade their GPUs. And on the tool side, same thing, you're forcing US companies out of a market and also, by the way, creating their. Interestingly, Huawei may not be a competitor yet for Nvidia, but in the tool sector, for example, Chinese companies are already competing with US toolmakers in some cases inside China and eventually outside China, because every fab, for example, in China that's restricted has become a technology test bed for domestic Chinese technology. So, anyway, I lay this a lot of detail in my new paper that's going to come out in January.
- [50:32]
B
I'll look forward to it. I mean, since we're talking tools already, there's no tool that grabs more attention than. And then, of course, the ASML EUV lithography machine the size of a city bus. I mentioned that Reuters story about Chinese reported advances in euv. It's been seen as a major Perhaps the biggest technological hurdle for China in achieving its long cherished goal of technology independence.
- [51:02]
A
Yes.
- [51:03]
B
What's the scuttlebutt from what you hear, apart from this report about this effort? Did you have any issues with Reuters reporting on this? Did you? Yeah, I mean, I've heard talk of this for, for quite some time, but.
- [51:16]
A
Yeah, no, this is, this is, this is not unknown.
- [51:21]
B
Right.
- [51:22]
A
Within, within watchers, close watchers of the industry. But you know, the hard part is sort of getting inside the black box here. Right. So, yes, it was, it's been long known that Huawei has such a, and effort, I think. Yeah, I mean, it's a really good question about, about how I feel about this and it's sort of evolving. This has just, just come out a couple days ago. But basically one thing that, that's TR that jumped out at me was, okay, first of all, ASML gets all the attention here, but if you look at one of those buses, those advanced ASML EUV tools, there's multiple really critical subsystems there that are made by other companies. Right. So it's not just asml, it's, you're not just duplicating, you know, asml, it's.
- [52:06]
B
Also Zeiss lenses and. Yeah, you know.
- [52:08]
A
Yeah, optics. Yeah, there's, there's, there's this light source, the, and the optics, for example. And the, the piece mentions the, the light source and, and, and Zeiss, as you know, on the optics and how they're having trouble with optics. But, you know, China has some very capable optics research institutes. Again, the, the, the challenge here always is translating that R and D in the technology space into a working commercially viable thing. So first of all, the piece mentioned some of the critical subsystems, but it didn't mention a whole lot of other things. Right. That was one of my main issues with it. So again, a single EUV machine is useless, you know, if it's not within a broader ecosystem. So, for example, photoresists are really critical. These are the things you put on the wafer when you shine the light on it, Right, Sure. And those are, those are really hard to do for euv. Right. So if the, if photoresists aren't part of this, you know, this effort, which I assume they are, but it wasn't mentioned in the article how far they are, then, then it's, then it's tricky. And then the other thing is, are other critical things like computational lithography, which is, which is software that helps to align the light source, you know, so you, so because you're when you're talking about, you know, nanometer wavelengths, the accuracy is really critical to get all this to work together. So the problem is in the article, it described cannibalizing parts from other, other systems, which is also I'm. I'm a little bit dubious about, because there are no EUV systems to cannibalize cannibalized parts from. So they must have been cannibalizing parts from, from DUV systems from, from deep. And there's no secondary market for EUV as there are for other tools. So. So there's no way that they could have cannibalized part parts from EUV systems. So that means they cannibalize them from non EU systems, probably DUV systems. And the question then is how. There are definitely individual parts that you can do that for, but they're very limited. And then they talked about ASML engineers with experience. That's really hard to assess too, because there were people trained on the system, may not be the people who actually understand the inner workings and design the system who could really, you know, make. Make a difference here. So obviously they pulled together a lot of interesting resources on this, not surprisingly. And, and they're throwing at it other pieces of the system.
- [54:23]
B
The.
- [54:23]
A
The Shanghai Institute of Optics and Fine Mechanics. You know, they have, they have. Domestically, they have a lot of R and D research institutes that are really good in some of these, these specialties. But you have to realize these. There's like multiple, like 10 different, really advanced technologies that all have to work together perfectly.
- [54:40]
B
The upshot, though, the upshot of this is that it validates the claims that people like you and me and many others were making, that the whole endeavor of trying to starve China of key tech inputs would simply light a fire and would force Beijing to pour a lot of resources and a lot of effort into trying to bridge the gap. Right. And.
- [55:04]
A
Right, right. And that's a great point. And I think that, that, that's, that's probably the big takeaway is this idea of chokepoint technologies, which is, again, sort of the theoretical basis for this. This was a paper, A paper was written in 2021 on choke point technology at CSET. Right. And that ended up being the basis for a lot of the export controls. But the problem is, as I've pointed out and you pointed out in papers I've written on this, you know, this is an applied technology. And so there's always more than one way to do things. And you can't lock this technology in a box and then think that China isn't, you know, somebody couldn't invent the same technology. They might not do it the exact same way. But there's, they're, they're gonna, they're gonna figure it out. And so the, the, the, I think that that report highlights the sort of almost delusional nature of thinking that Chokepoint technologies are going to work. Now the, the proponents will argue that the idea was to slow China down. So they've slowed China down, but at the same time in slowing China down, they've done a lot of things. They've disadvantaged and undercut US Companies leadership in all these areas and potentially the Dutch over time, right. If they do develop a DUV and euv and you've, you know, you've disrupted supply chains, you've, you've created huge losses for companies in this space, a lot of which is not even known publicly. I can, I can assure you. And so the question of the costs and benefits, and I've written a lot about this, like did anybody look at the cost and benefits? No, they didn't look at that. Of this policy of trying to attack choke points. And again, you know, recently a former administration official sort of doubled down on the choke point thing. Right? We need to the idea that there are more choke points that we need to try to choke off more choke points. And again, as you adroitly point out, what this has done is forced companies like Huawei, working within what I call the Huawei semiconductor industrial complex, which will be featured in my paper is, you know, figure out how to do a lot of stuff, you know, again, working with a lot of different players here and that they never would have done. So it's inconceivable to people in the industry the progress that if you Woke up in October 2022 and looked at the domestic tool industry and now, you know, it's, it's night and day in terms of the capabilities they've developed. Because as I've noted, the Chinese front end manufacturers weren't willing to work with those, with those toolmakers. Now they have to and they are and that, that's a symbiotic relationship that drives the innovation. Right. The toolmakers design to the requirements of the customer.
- [57:43]
B
Right.
- [57:43]
A
And the customer helps to push the envelope on that making the companies have to design to meet those requirements. So that's how innovation happens. And when you're in that loop, you can make a lot of Progress. But before October 2022, Chinese companies domestically weren't in that loop. Now they are. Right. And so that's a flywheel, if you will. That's really, really important to understand because it has huge implications for companies like US companies and Japanese companies and Dutch companies that were the leaders in this space. Now you're creating competitors that will compete with you, not just in China, but outside of China too.
- [58:22]
B
Paul, before we move to wrap up with a couple of final questions, I do want to ask you how you think Taiwan fits into this. Is Taiwan safer if China is pushed toward a fully sanctions proof AI supply chain? Or does selective engagement actually reduce long term risk by actually maintaining the potency of the so called Silicon Shield?
- [58:44]
A
It's a great question and it turns out to be really complicated because yes, you could argue that if China has a fully domestic industry capable of doing everything, then that sort of makes Taiwan and Silicon Shield less valuable. But it's, it's a little more complicated than that because for example, right now there are lots of Chinese companies that can still use TSMC to manufacture their semiconductors. So Xiaomi, for example, can use TSMC for legacy chips? No, for advanced chips. For very advanced. Right. As long as they're consumer focused. The thrust of the US effort to cut off China from TSMC is for GPUs and data center GPUs, but for SoCs, for companies like Xiaomi, as long as they can do the design, there's nothing to prevent them from manufacturing those at tsmc. So Chinese companies are still using TSMC to manufacture consumer focused advanced.
- [59:51]
B
So for tablets and smartphones and computers. Yeah, okay.
- [59:55]
A
Yeah, absolutely. Right. And also Chinese companies use semiconductors manufactured on Taiwan. I'm pretty sure that the ZTE phone, for example, that ByteDance designed with ZTE, that the Nubia 153.
- [60:11]
B
Yeah. The agentic AI embedded.
- [60:14]
A
Yeah. That's using a Qualcomm ASIC that was manufactured on Taiwan. Right. So the linkages between China and Taiwan are still pretty complex. Right. And those would, those, I'm not sure those would go away. Even if Huawei gets at some point 2027, 2028, 2029, that they can manufacture some number of advanced semiconductors, including GPUs, including ASICs for, you know, it's, there's still going to be a lot of Chinese companies that would still want to use potentially tsmc. So there's not a point, I think in the near term where there's a complete, you know, sort of severing, severing of that relationship. And so that's, that's also a complicating factor because there's still quite a number of linkages there that that are hopefully.
- [61:01]
B
I've just given you a good topic for a substack to address in the future here that sounds like a substack. So let me zoom out a little bit here. Looking ahead, do you expect this decision of President Trump's to trigger a broader rethinking of export controls writ large, or are we more likely to see a kind of patchwork approach? It's exceptions here, resistance there, continued rhetorical escalation. I mean, does this whole episode with the H200 suggest that the post2022 doctrine is quietly collapsing or that it's being defended even as its internal logic is pretty obviously strained? Where do you think we are?
- [61:42]
A
It's a great, great question. And I think it's still sort of hard to tell. I think you're seeing a lot more people cite this and other things in the EUV report and say, look, it looks like the controls have failed, right? So you have these two camps, right? There are some pointing to the current developments and saying, look, the controls have failed. There are others who are pointing to the same events that are happening and saying, look, we need to do more on the export controls. We need to tighten them up. And of course, the H200, as I mentioned with that DOJ, you know, comment and has sort of become the poster boy there. The problem in assessing this is, as I noted, there was this sort of more coherent policy direction that was coming from the White House to Commerce to do, you know, to put in place these controls. That's completely gone now. So the question of what replaces that is still a little bit unclear to me as long as people like David Sacks and the other thing is people like David Sachs, as long as David Sachs is in the mix. Now remember, David is a special government employee like Elon was. And so it's not clear how long his tenure will last because he only has a certain number of days per year and, you know, nobody. I would like to FOIA his his time cars to see how he's keeping time because I would like to keep him in Washington because I think he's a good influence and because he's helping to. And again, he's very careful to say he's a China hawk, but he wants to have a more nuanced policy here. And that's why he supported the H200 decision here, because he thought that moving back to some kind of a sliding scale, being more nuanced about how this is approached is better for everybody. It's better for Nvidia it's better for the industry and it's better ultimately for US Dominance of the US Stack. Right. Which is, which is what his ultimate, you know, what he tends to tout. So if he stays around, you know, then, then there could be a broader rethink of this. The question is, who would do that? As I noted, the tool controls are very complicated. They're layered, right? They were, they were updated a couple times as the Biden administration was going out the door. They changed the end use controls on dram, which ended up catching more tools from some of the leading US Companies. So it's a very, it's a kind of a big ball of wax. Right. And to unpack that. And again, as I noted, I think you can make the same argument that Jensen and David Sachs have made on GPUs with the tools, but it's a little harder to make. Right. And the question is, who to make it to? Do you make it to David Sacks? Do you make it to Jacob Helberg at the State Department or Michael Kratzios at ostp? I mean, there's not a clear, there's not always a. Besides David Sacks, it's not clear who, what, what sort of sympathetic ears there would be in the, in the, in the administration. But as I noted, there will be this in, in January, a new interagency group meeting, I believe, under the leadership of Marco Rubio as, as National Security Advisor to look at the AI diffusion problem. Right. Because that's become so messy over the last year. There's a sense that, okay, we need, there needs to be a more orderly process to decide who gets licenses to ship GPUs to the middle east, right? To Saudi Arabia and the uae. And I think as part of that, maybe there will be included in that discussion a sort of rethink around China and GPUs in China, because as you asked earlier, there is no clarity in terms of what the new policy is. So when does Blackwell, for example, that BALCO architecture become available to sell to China? Is it 18 months? Is it 24 months? Is it 34 years? You know, nobody knows. Nobody. There's been no discussion of that at any sort of policy level that means anything. So, but I think in early 2026 that will change because there is sort of the sense that, you know, there needs to be some order returned to this issue because otherwise you're going to have people like Jensen lobbying Trump in Mar? A Lago in the White House and the China hawks and the export control proponents, you know, pushing back on this. And it's kind of a. It's kind of a disorderly process because there is no policy. And so there will be an attempt, I think, to impose some new policy regime around this in the first quarter or second quarter of 2026. But again, who the players are and how that comes out is tricky to.
- [65:57]
B
Say because they're so methodical and orderly. I mean.
- [66:00]
A
Right. And then you have, you know, Trump, you know, I mean, that, that true social post, I'll tell you that really, that really sort of blew my mind when I saw that, just because it was so interestingly worded and it was a culmination of a lot of things that I'd been following very closely. So I knew right away where it came from, what the thinking was behind it. But, you know, it was still a little bit. I was a little bit surprised that it happened so quickly. Right. Because after Busan, it was like, wow, you know, he's really pushing this. So anyway, so that's a, it's a fascinating issue, but there's, you know, but we are in this sort of disorderly state here where the policy processes are not clear, the interagency processes are not clear, and everything goes through the White House. Right. And that tends to be only the only thing that matters, as we saw from the truth social post. Right. I mean, to get. Can you imagine having policy in the Biden era determined by a tweet from Joe Biden? Right.
- [66:56]
B
God save us.
- [66:57]
A
It's inconceivable. Right.
- [66:58]
B
So, Paul, last question for you. So if we were to fast forward just two, three years, just for the group of people who have been very critical about Trump's age 200 decision, for them to have been right, for them to have been completely right to justify their criticism of this, what would have to be true? Looking at China, and then maybe on the flip side, what would have had to be true for us to be able to say they were wrong or for them to be able to admit they were wrong?
- [67:26]
A
Well, what would have to be true would be that the h2 hundreds enable deep seq and Alibaba to develop more advanced models than OpenAI, Anthropic, Google and Meta, and I just don't see that. Or an xai. Right. And if you look at what the resources that are being brought to bear on data center development in the US by all those companies, it's just mind staggering.
- [68:01]
B
Lots of Capex.
- [68:02]
A
Yeah. The numbers of billions of dollars. Now, the problem in the US Is going to be power and energy. We work with companies in that space. And if you extrapolate out those data center commitments to 2030, it's like how the heck are we going to provide power to those? There's not enough power in the US to do that. Right, right.
- [68:19]
B
Not an issue for China. Right.
- [68:20]
A
China's putting in place every year the equivalent of the whole US grid. Yeah. So power is really this huge thing. And so again, that doesn't mean power in itself, doesn't mean China wins the AI race. Right. It's still much more complicated combination of innovation and the model space. And you know, it's not. Power is not a determinant. Neither is compute either, as we've seen from Deep Seq and, and other Chinese companies that have been able to train really good models on much less compute than in the US So I think that what would have to be true is somehow the h2 hundreds enable deep seq to make some really unthought of innovation that leads suddenly to some really advanced capability, step function and capability. And again, I don't think that's going to happen because right now the beauty of open source and Chinese open source models and Deep Seq in particular, they're publishing a lot of data about how they're doing these models. Right. And really what's more important is the sort of open versus closed model development environment. And the H2 hundreds could for example mean that Chinese open source open weight models are better and are adopted more globally. Right. So that's, but that's an economic argument. Right. So and again, the proponents of the control sort of mix up the military and the more long term risk and the, and the superintelligence risk with the economic risk. Because I just want to quote you on that in the Matt Pottinger, he quotes Lika Chang as telling him Mr. Trump in 2017 that Beijing and China would come to dominate all technologies, including AI and that America would export little more than soybeans and corn. Right. And then the conclusion is last week's decision helps make that unlikely dream a reality. Right. Again, like really, I mean the H200 are going to mean that in five years the US is exporting corns and.
- [70:20]
B
Soybeans, frozen, concentrated orange juice, pork belly.
- [70:25]
A
Right, right. I mean, because the H200's export does not mean that the US loses anything, any capability. There's no loss of capability. It's just giving China a little bit a step incremental ability to train better models. But it doesn't give away the huge US advantage in compute. And then, you know, innovative companies Right. So. So it's. I kind of, I kind of, you know, I had a coffee spiel moment when I read that because I was trying to picture. I was trying to picture the U.S. you know, just exporting soybeans.
- [70:57]
B
Yeah, I'll make sure to link to that. Op ed. Today is the 18th, and it's interesting. It just came out this morning, the very day that we were recording. So December 18th.
- [71:05]
A
I think it's worth a read. For sure. For sure. And again, the arguments that Ben was on was on the Edward Klein show. And these are very thoughtful people. Very thoughtful people. I respect their views on these things because their views are well considered. But as I say, there's some fundamental issues that I think are unexamined and have not been sort of really hammered out in terms of what the implications of them are. For example, that issue of what the governments do when they get to some high level of AI and capability. You know, there's just not enough discussion of what that would actually mean. There's an assumption, and that assumption, I think is not really based on sort of a really hard look at, at understanding China. Right, right. And how Chinese government would. Would react to that. I think there's sort of a mirror imaging. Maybe the US Government, if it got to AGI, would take down China's critical infrastructure. That's the only thing I can infer from.
- [72:06]
B
There's always a lot of projecting going on with these people.
- [72:08]
A
Right, right. Is that, is that. That's what we would do if we got there first? Wow. Let's take down China's critical infrastructure and threaten to regime change. Right. I mean, really. So anyway, so that, that the pro. I just have a big problem with that. But again, these, these are thoughtful people, and in the Biden administration, there were very thoughtful people about AI safety and AI security and the misuse of AI and cybersecurity. And those are very valid concerns. But again, my view is that we need to collaborate with China to avoid the malicious actors, for example, getting a hold of advanced AI. Right. Because that seems like the much more likely possibility of misuse than a government like China misusing AI in some sort of random way, like taking down US Political.
- [72:50]
B
Paul, that's too sensible. All right, we're going to leave it there. And thank you so much. I mean, this has been just a really wild ride and nothing, nothing that I didn't expect from you. So fantastic. Thanks for taking so much time to talk to me. Let's move on to paying it forward. Who's somebody you've been working with recently or some younger person in, you know, either at DJI Albright Stone Bridge or somewhere else Whose work you think we should be paying attention to.
- [73:20]
A
That is a good, a good, a good question. I would say my choice for today would be Po Zhao who is a. Yeah, no, absolutely.
- [73:33]
B
I'm a big fan of.
- [73:35]
A
He has a good substack.
- [73:36]
B
Oh it's fantastic. Fantastic substack.
- [73:38]
A
I think it's a really good, it's sort of again, sort of facts based. He's not, you know, he's not grinding any axes and I think he's. His, his analysis is really good and good.
- [73:50]
B
Insider written from China by. Right.
- [73:53]
A
Somebody who's exactly.
- [73:54]
B
Deeply China immersed in technology there.
- [73:56]
A
Yeah, exactly. And I tend to, I talk to a lot of people across the semiconductor AI stack who are in China or who have, you know, who are on the ground and on the day to day. Day to day, you know, sort of following these issues as they become more and more complex. And so having, I think we're, we're, we're lucky that we have, you know, people like that, that, that are writing more on this that can help balance because I think a lot of the discussion here is driven by people in the US who, who never go to China or have never been to China and their view is very different. Right. Than somebody writing on these topics from within China who is much closer to the sort of ground truth, if you will. And I find that really refreshing in this.
- [74:37]
B
So pose substack is called hello China Tech. It's at hello Chinatech. Hello China dot com. So check it out. He's written on this very topic just for disclosure. I reached out to see if he would be willing to join you on this and he wasn't able to join this time. But I look forward to yeah. Getting him on the show again sometime. I'm really.
- [74:57]
A
I'm going to meet him in Beijing.
- [74:59]
B
Oh, good, good, good. Yeah, yeah. Well I'll be back before too long, so let's, let's hang.
- [75:03]
A
Yeah, yeah, let's, let's definitely for sure.
- [75:06]
B
All right. What about recommendations? What you got for us in terms of recommendations this week, Paul?
- [75:11]
A
Well, I have a very, not a sort of a Off the. Off the beaten path recommendation.
- [75:17]
B
I see you holding a book there.
- [75:18]
A
Yeah, Big. Yeah, it's as big the Life of Brzezinski, America's Great Profit. And I think it's just a really interesting look. It comes from a different era, but I think it shows how. I think what we're missing these Days are sort of strategic thinkers like Brzezinski who can look beyond the day to day and think about great powers in a sophisticated and nuanced way. And again, I think it illustrates how much now we sort of miss that at senior levels of government. And I think it would be great. And particularly with respect. Respect to China. And so I think it's a good sort of. You know, there are. There was an era when we did have. We did have strategic thinkers on national security that, that were really. That were not just serving in government, but they were. They expanded beyond that and were really seminal thinkers on what am. How America should approach the world, and particularly how America should approach, you know, complicated problems like Russia and China and. And these. And other issues which are still very much salient. And I think on China, you know, we've never needed sort of more strategic thinking than we do.
- [76:28]
B
So let me say an additional word on behalf of the author of that book, Ed Luce, who is a. He's a columnist for the FT and probably my favorite columnist writing today. I just think he. He just never gets it wrong. I mean, everything he writes, I am in emphatic agreement.
- [76:45]
A
I meant to add that. Yeah, yeah, I meant to add that Ed is one of my favorite. Yeah, he's really good. And so there's the China connection.
- [76:51]
B
Yeah.
- [76:51]
A
Yeah.
- [76:52]
B
I mean, he writes on China quite a bit. He's very smart on China, I mean. And, you know, he's Washington correspondent or Washington columnist, so he writes on global affairs. I've had the pleasure of having lunch with him one day in D.C. and I got to say, just a lovely man in person, too. Just a terrific guy. So. Yeah.
- [77:12]
A
Yeah, maybe we can have. Maybe we can have him on. I'd like. I'd love to be on a panel on a podcast with. With them. And you need one.
- [77:18]
B
Yeah. If you want one. Yeah, go. If you got another one.
- [77:20]
A
Yeah, one more. There's. There's a really good substack that I think people should look at. It's called Hyperdimensional, and it's by Dean Ball. Dean wrote the. The US AI Action Plan, and he's a really seminal thinker on AI issues. We're going to have. Probably have him on a. On a. On a call at. At my. My organization. He's really good.
- [77:40]
B
I'm looking at it now, and he's.
- [77:42]
A
Yeah. And he. It's very. He is very thoughtful on these issues. And I think, you know, it's. It's. I. I tend to. I've been reading. I'm just always floored at how good some of the substacks are that, that again like with, with Pojiao and others on China. And I think Dean is, is a, is a sort of up and coming young thinker on this issue and he's, you know, he, he wrote the AI action plan. So he's very, he's very, been very involved in policymaking circles here. But then he stepped back to be able to write a substack. And so it's a good read and he brings a lot of experience and I think creative thinking to the top.
- [78:18]
B
I will add that to my list. I'm looking at it right now. I see his latest is called where do you stand on Ghosts.
- [78:25]
A
Yeah. And he's, you know, the bigger debate in AI, you know, sort of between the doomers and the acceleration over on safety people and then the sort of acceleration. Right, right. So that, that debate is, is very much alive in Washington and affects this discussion around.
- [78:42]
B
Absolutely.
- [78:43]
A
Because it's along those lines. So I think that's, that's a good recommendation as to. And he's also, you know, the Washington, I think in 26, the sort of Washington view on AI is going to be really important both sort of domestically and then globally in the Middle east and other places, how the US government is going to push the USAI stack and also deal with these complicated issues like diffusion and what to do about China and advanced compute. So anyway, Dean is a good person to watch. Thanks man.
- [79:13]
B
Zvig absolutely goes on my reading list. I've got a lot. I'm in the middle of a huge, huge, huge thick book right now, which I don't foresee finishing before Christmas, but. And I've got another couple of assignments to read, you know, things that I've forgotten, upcoming interviews. But then I will absolutely get to Ed Luce's biography of Brzezinski. I've got a book to recommend too, actually a couple of books by the same author by John Green. You might know who John Green is. He wrote a very well regarded YA book years ago called the Fault in our Stars. It was made into a movie even. But John and his brother had been sort of early vloggers. They were the Vlog brothers, Hank and John Green. Fantastic stuff. Just. It sounds good.
- [80:00]
A
Yeah.
- [80:01]
B
But so, you know, let me just quickly tell the story of how I arrived at it and I can throw another recommendation. In the meantime, I was corresponding with the son, one of the sons of Joseph Levinson, who is. Anyone who knows me knows as somebody I'm sort of obsessed with Tom. Tom Levinson. And in his email, I'd read one of his books before, History of Money, which was just fantastically good. He's a professor at mit. He teaches science writing at mit. So obviously he's a really good writer and he knows the science up and down. So when I saw in his email signature, one of the books that he had written recently, it's called so Very Small How Humans Discovered the Microcosmos, Defeated Germs and May Still Lose the War Against Infrastructure. Infectious Disease. Thomas Levinson. Just fantastic book about history of our discovery of the Microbes, microbial world, microbial diseases, bacteria and eventually viruses. But fantastic book. But after I finished that, a new recommendation popped up for a book about tuberculosis by none other than John Green. It's called Everything Is Tuberculosis. Tuberculosis is something I'm kind of interested. Had an interest in. I had written a piece on my substack some. A few months back called the Wasting Disease, talking about consumption, which of course is another old word for tuberculosis. Talking about consumption in the context of the Chinese economy and just sort of punning on that. But I had realized because of my own reading habits that just so much of what I read, the authors or characters in books or, you know, the subtext of it is all about consumption. It's all about tuberculosis. I mean, there's. Everyone seems to be like every writer in a character.
- [81:52]
A
The Magic Mountain. Yeah.
- [81:53]
B
Yeah.
- [81:54]
A
Young Hans Kastor.
- [81:56]
B
Yeah, yeah, he sure was. Yeah. I mean, and I actually, you know, even though I go there every year, I had never actually read the Magic Mountain by Thomas Mann until just a.
- [82:07]
A
Couple of years ago.
- [82:08]
B
Yeah, it's just absolutely great. So anyway, he wrote this book called Everything Is Tuberculosis. It's fabulous. I highly recommend it. And then after reading that, I said, I wonder if there's anything else by John Green that's out there in a nonfiction book. And I came across this fantastic essay collection called the Anthropocene Reviewed. It's a collection of essays. There's several dozen essays, and they're all great. And they all end with sort of a star rating of whatever it is that he's talking about. It just goes wildly all over our Earth and talks about. I mean, it's deeply personal. And the guy is just. He exudes such warmth and compassion and humanity and earnestness. I mean, he's just like the most decent guy, guy you've ever come across, guy you want your daughter to marry. I mean, just this kind of a guy. He's just. Just a fantastic human being in all evidence points to it. I've heard him on many interviews. I've watched him, you know, and read his, his novels as well as his, now his essays. And so John Green. Anything that John Green touches, I endorse. So the Anthropocene reviewed a collection of splendid essays. He's up there with a few other people in my canon of greats. I mean, I would put like the writer Michael Chabon on that list. The historian Will Durant. Those are, those are people who I just admire kind of immoderately. Okay, that'll do it.
- [83:43]
A
Wow. Yeah.
- [83:44]
B
Paul, as always, great to talk to you.
- [83:47]
A
My pleasure. I'm happy to try to add a little bit of clarity to this very complicated situation here, but as always, you're asking the right questions and appreciate your hosting style.
- [84:02]
B
Well, thanks. Thanks.
- [84:04]
A
You always bring a lot to the table and it makes it easier. And you obviously know yourself a lot about this.
- [84:11]
B
Yeah, I try to do my homework. Hey, thanks, Paul. We'll talk to you real soon. You've been listening to the Seneca Podcast. The show is produced, recorded, engineered, edited and mastered by me, Kaiser Gua. Support the show through substack@synecapodcast.com where you will find a growing offering of terrific original China related writing and audio. Email me@cinecopodmail.com if you've got ideas on how you can help out with the show. Don't forget to leave me a review on Apple Podcasts. Enormous gratitude to the University of Wisconsin Madison center for East Asian Studies for supporting the show this year. Huge thanks to my guest, Paul Triological. Thanks for listening and we'll see you next week. Take care.