← The AI Policy Podcast
The AI Policy Podcast
Why is China's AI Sector Booming?
00:46:39
In this episode, we discuss China's focus on AI adoption, the underlying factors driving investor enthusiasm, and the national security implications of China's booming AI industry.
Loading summary
Transcript29 lines
- [00:00]
A
Foreign.
- [00:10]
B
Welcome back to the AI Policy Podcast. Today we're going to be doing a deep dive into China's booming AI sector and what it means for US China competition. I'm Sadie McCullough, the program coordinator here at the Wadhwani AI center, and I'm joined, as always, by Greg Allen. Welcome, Greg.
- [00:24]
A
I should be welcoming you. This is your first time on the podcast. I'm really excited to do this with you.
- [00:29]
B
Me too. I'm usually on, you know, the behind the scenes part of the podcast. So I'm really excited to be here.
- [00:34]
A
For those who are wondering where Brielle is, Brielle was the victim of a tragic promotion. So she has been promoted at CSIS and is now the chief of staff to the entire Economic Security and Technology department. So congrats to her and we'll miss her dearly. But Sadie, so glad to have you on the podcast.
- [00:55]
B
I'm so glad to be here. Let's jump right into this big topic. So in January of this year, the release of The Deep Sea's R1 reasoning model shocked the world, signaling that China was still a fierce competitor in the AI race. Now, nine months after the so called Deep Seq moment, what's going on in the Chinese AI sector?
- [01:12]
A
Yeah, I mean, that Deep Sea moment was like a crazy media circus and political firestorm. You know, President Trump had to talk about it. The chairman of the Chinese Communist Party, Xi Jinping, convened this whole, you know, meeting of the great tech muckety mucks in China, and the CEO of Deep Seq was prominently featured there. So that was a huge moment in the global AI world. I think folks will remember that Nvidia's stock price briefly plummeted after the so called Deep Seq moment. Well, all the tech stocks have recovered in the United States, Right. Nvidia's stock price has exploded, you know, yet again following incredible revenue and earnings growth. So the hypothesis that Deep Seq signaled something terrible for America, I think did not get borne out the way that some of the doom sayers were predicting at the time. But Deep Seek also heralded a moment of optimism for China. And I think that actually has largely continued. And just in the recent few months, we've seen really a crescendo of optimism about what China can achieve in the AI sector. And I think one way to think about it is the following. So, you know, there's this opportunity that China could have had in a parallel universe where export controls never happened. Right. It's possible that if export controls had never happened, the largest AI supercluster on earth could be in China today. Not right now. You know, OpenAI has 700 million weekly average users. It's possible that, you know, the largest platforms for AI could be in China today in the absence of the export controls. And I think what Deepseek forced Chinese investors to realize, forced investors outside of China who are interested in Chinese markets to realize, is that maybe the Chinese AI economic opportunity is not as big as it would have been in this parallel universe where export controls had never happened. But it's still a really big opportunity. There's still a ton of money to be made in China by companies that succeed in adopting AI. And as Deep Seek showed, there's still world class talent in China. You know, Deep Seek, what I've heard from the folks at OpenAI, Anthropic, Google, elsewhere, the smartest people at these companies, they look at Deep Seq and they're like, yeah, these guys are our technological peers. They're really, really, really smart and they're clearly working very hard. I saw an analysis at one point recently that said something like more than 47% of all the AI engineers on planet Earth are currently in China. So this is still a big market, there's still a big opportunity. And even if the boom is not as big as it was, it's still a really big boom. And that's something that folks are seizing on. And I think the other aspect of what they're seizing on is the opportunity for AI applications. Right. How are we actually going to weave AI into our overall economy? Maybe we're not going to have the absolute positive best model on planet Earth, but we can still have great models. We can still have great models that are tied to meaningful economic applications.
- [04:22]
B
So when you say that China's more focused on these applications, what exactly does that mean?
- [04:25]
A
Well, I would start, I guess, by contrasting it with what we see in the west, because I don't mean to suggest that folks in America or Silicon Valley are not focused on applications. Certainly they are. However, there's also this overriding interest and some would say obsession with the idea of superintelligence. And one way that I think this is helpful to understand this is you can think about a AI as a general purpose technology in the way that electricity is a general purpose technology, in the way that the steam engine was a general purpose technology, in the way that the Internet is a general purpose technology. And that is a historic, important moment. But then there's the framing of AI superintelligence, which is to say this is going to be, you know, a thousand times smarter than the smartest human being or 1000 times smarter than all human beings combined sometime in the next 10 years. And that looks a little bit less like electricity and more like all of human history is a before moment and then there's an after moment. So it's not like we're not, we're not talking about the same league as like inventing electricity. We're more like talking about the evolution from like the common ancestor of chimpanzees and humans to humans. Like, that's, that's kind of what we think of when we talk about super intelligence sometimes. And to put it in terms, you know, Demis Hassabas, the CEO of Google DeepMind, he said earlier this year that the AI revolution will be 10 times bigger and maybe 10 times faster than the Industrial Revolution. So that sort of suggests this really crazy kind of transformative moment. So that's what I would say is the frame in a lot of the West. And you know, Mark Zuckerberg, he named the meta new project the Super Intelligence Lab, right? Keeping with that super intelligence time framing. By contrast, if you look at what's coming out of China, and here I'm focused on what Chairman Xi Jinping said in the April Politburo study session on AI. He directed the country to be, quote, strongly oriented towards applications. And so he also emphasized that, like, when you think about what is strategic about the competition between the United States and China, you know, maybe you could say that the United States is great in that we have like these free markets which are great at identifying opportunities to make money and et cetera, et cetera. But where the Communist Party has always identified its advantage is that they can effectively mobilize the whole country in pursuit of strategic objectives, right? When they say, hey, we're going to be great at high speed rail, boom. They build an awful lot of high speed rail. When they say that they're going to be great at solar power and renewables, like, they go from barely present in this industry to they've completely taken over this industry worldwide in relatively short fashion. And so he's basically saying those government muscle movements that China knows how to do very well, which is to orient itself around strategic goals. That is our advantage in the AI race. And what we're going to do is we're going to be better at adopting key applications. And so how could that translate into victory? I think maybe a helpful analogy here is to just like think about the first Industrial revolution, the Steam engine era, right? One way of Winning the industrial revolution, you know, could be to have the best overall steam engine, right? So like, our British steam engines better, or our French or German steam engines better. Right? And that's what's going to determine GDP growth. But actually, I think if you could imagine a scenario in which, you know, German steam engines are only 80% as good as, as British steam engines, But Germany is 10 times better about incorporating those into the economy. And keep in mind, right, like when you're going from horses fed with grass to steam engines that require you to mine coal and, like, dig up more iron ore and, you know, build railroads for you, really is a lot of restructuring of the processes of the economy to take into account the existence of this new industrial revolution. And so basically, China is saying, just because we may not have the best AI in the world from an absolute peak performance perspective, doesn't mean that we might not have the overall best economic benefit from AI if we are better at driving adoption. And then, of course, there's a virtuous cycle whereby more adoption of a technology means more, you know, consumers spending money or businesses spending money that you can then plow back into R and D to ultimately try to achieve some kind of technological leadership.
- [09:15]
B
Sure. So you already kind of started talking about this, but it seems as if the April directive to focus on AI application has been made more concrete in the new AI policy document. On August 26, China's State Council released a document titled Opinions on Deeply Implementing the Artificial Intelligence plus Action. What is AI plus and why is this document important?
- [09:39]
A
Yeah, so the April Politburo study session, I mean, when. When Xi Jinping says something does have the force of policy. But what this is is coming from the Chinese State Council. So that's the same group that released the Chinese AI strategy in 2017. And while this document is titled Opinions, I think all the local governments are like, yes, and those are the correct opinions that we need to hold, you know, immediately. And I will say that I'm borrowing here from some really good analysis of this AI plus document that was done by Matt Shen, who we've done, who we've had on the show before. He's a senior fellow at Carnegie and is doing a lot of great analysis of what's going on in China's AI ecosystem. And then also Patrick Zhang, who runs the Geopolitex substack, and Matt just launched a substack. So I do encourage our listeners to check both of those out. So I think to. First, I want to say that Matt sort of agrees with my overarching statement here, he wrote in his most recent newsletter quote. Notably, the document doesn't cite AGI as the driver of these changes and it pays relatively little attention to frontier AI development in general. Instead, it's focused on how the AI of today ish can be leveraged to achieve the Chinese Communist Party's economic, social and political goals. So very much a focus like, hey, LLMs are already really impressive right now. We don't need some kind of crazy 100x in performance improvement to recognize that we have an exciting opportunity now. So the document is called AI and this has some heritage in Chinese policy documents. Ten years ago they released a document called Internet plus. And the framing there is basically we're going to combine that technology of the Internet with specific applications and industries where it matters. So if you think about combining the Internet with taxis, you get Uber or in the Chinese equivalent Didi. Is that sort of their Uber esque company or Internet plus, like shopping. That's the e commerce revolution. And so there, there's an analogy here between the Internet plus plan and the AI plus plan. What is going to happen when you apply AI on an industry by industry basis and generate the kind of explosive growth that we've seen? And I think, you know, for, for somebody like me who the iPhone came out when I think I was graduating from high school is when the first iPhone came out. So the mobile Internet revolution, you know, I lived through it. It was pretty shocking and startling. You know, the pace of technological change in China, it's been so much more. I mean, they went from like, everything is paid for with paper money to like, we don't even have credit cards, everything is paid for with mobile payments. And it's like actually hard to live in China if you try and use paper money at this point.
- [12:43]
B
And how long was that transition between paper money and now?
- [12:46]
A
Like a handful of years. Right? So we're not talking decades, we're talking like three to five years tops. So like a jarring pace of technological change is a very familiar experience in China to a degree that I would argue is even more extreme than what we experienced here in the United States. And you know, if you're the Chinese government, you're saying like, yeah, you're welcome, we did that. And so when they say they're going to drive, you know, a similar category of transformation across the economy with this AI plus plan, you can bet that every local government is paying attention and every business is paying attention. So they've identified six areas where they sort of see this AI plus opportunity. And that's the AI plus science and technology, AI plus industrial development, AI plus consumption upgrading, which, you know, for AI or for Chinese economic policy nerds, they'll notice that that's an important one. AI plus people's livelihood and well being and AI plus governance capacity. So those are all areas where basically every local government, right, who's responsible for a sector there, every ministry of the government is supposed to like ask themselves, how could we drive. Oh, and I missed one there, AI plus global cooperation. But you know, how could we drive more accelerated AI adoption? And what's also interesting is that they set quantitative targets for all of this. So there's supposed to be 70% adoption of AI by 2027, 90% adoption by 2030, and then by 2035, they're supposed to be what's called this intelligent economy and society, which is this new era, similar to how like the Internet is a new era. And I think, you know, when you Hear like these 90% numbers, maybe they sound really, really high. But like, I don't necessarily think so. If you think about like the Internet experience in China, these are ambitious targets, but they're not completely unreasonable ones.
- [14:51]
B
Okay, so with all of this adoption, how is the government's emphasis on these, of these AI applications and the AI sector's renewed optimism and excitement impacting Chinese markets?
- [15:02]
A
Right? Yeah. So if, if there's going to be a huge acceleration in the adoption of AI applications, then somebody should be making money off of that. Right. Who is actually going to be providing this AI, both in terms of like the software, cloud environment, type of AI part of the story, but also the hardware that's powering all of that. And here we've got some pretty good analysis coming out of a lot of investment research houses, places like Morgan Stanley. So in a recent analysis they did over the summer, they predict that China's AI industry is going to grow from $3.2 billion last year to $140 billion by 2030. So that is a lot of growth. And that wasn't even including like the AI infrastructure part of the story and component supplier. So when you include that, we're talking $1.4 trillion is the size of the quote unquote, AI economy in China. So that is a huge share of overall Chinese GDP that they're expecting is going to be captured by this so called AI industry. When you see like, hey, this is going to be a huge part of our economy. Obviously there's a lot of people who want to invest to be the company who Fulfills that. One other thing that I think is interesting here is there's a recent Gartner survey that suggests that corporate use of generative AI is already, you know, growing incredibly rapidly. So in their survey they found that only 8% of firms last year were saying that they were using generative AI. This year reran the same survey, it's up to 43%. So like when the Chinese government, you know, says every government agency, every state owned enterprise, you know, it's your job to adopt AI as fast as possible. I mean, that message is being carried. But even if you look at the Chinese private sector, when that deep SEQ moment happened in January, there was an explosion of optimism in saying, hey, we should adopt this. And keep in mind, right, Deep seq, if you use the open source version of it and you run it on your own computing hardware, it's free. So that excitement around adoption has really gone through now. That's the kind of the economic opportunity. But what about the inputs to that system, the actual investment? And here it is now clear that the biggest Chinese tech companies are preparing to spend jaw dropping amounts of money. So on this podcast we've talked about how companies like Amazon, Google, Microsoft, OpenAI, et cetera are like spending hundreds of billions of dollars per year to build out the AI infrastructure of the future that's going on right now. I mean, I've heard people who basically said that the AI infrastructure boom is like the equivalent of a government stimulus package, like the type of thing we did after the Great Recession or like during the COVID economic crisis. It's just coming from the private sector because there's so much investment in AI. So the Chinese AI boom is not quite that like enormous, but it's still enormous. So According to a July 2025 report from Morgan Stanley, AI capital expenditure investments by China's top six cloud providers. This is companies like Alibaba, which is like the Amazon of China, or Tencent, which is the top social media company of China, or Baidu, which is kind of like the Google search dominant firm of China, although they also, all these companies also do other things. But it's now set to reach 380 billion RMB this year. That's the equivalent, you know, if you translate that currency, that's $53 billion this year and that's up 60% from 2024. So when you hear like a company like Nvidia, which during the H20 debates was talking about like a $50 billion Chinese market opportunity that's growing massively, this is what they're talking about like these firms, they are traded on stock exchanges, which means they have to put out their investor relations reports and those things have to be true or they're subject to regulatory sanction. So they're saying, we expect we plan to spend this crazy amount of money. Now, if you like translate that into gigawatts, which everybody loves to talk about gigawatts in terms of data centers, that's like adding 3 to 4 gigawatts worth of data center capacity every year. And like a gigawatt is like a Hoover Dam worth of electricity.
- [19:38]
B
Huge amount.
- [19:39]
A
So like a lot. Okay, one final thing, right, what, what we were just talking about, that's only the big private sector companies who are going to be, you know, spending to invest in AI. There's also the government and state owned enterprises, which in China are a much, much larger part of the story. So if you think about, for example, the three largest state owned telecom companies, right, like this is like the AT&T and Verizon sort of equivalence of China. They're spending tens of billions of dollars on AI infrastructure as well. So there's just a ton of this new investment and a lot of excitement.
- [20:20]
B
If China's largest tech and cloud computing providers are getting ready to invest hundreds of billions of dollars in AI infrastructure, that should mean massive growth for AI chip designers and makers as well. So which companies stand to make the most from China's AI investment boom?
- [20:36]
A
Yeah, so I think, you know, you start with Nvidia. Nvidia is the company that's making the most money off of America's AI investment boom. And that's because they make the most important chips and data center, you know, insides for this story. And Nvidia is still making a ton of money in China. That's been true both before and after the export controls. In fact, they're China revenue has exploded since the export controls were adopted in 2022. So a little caveat here. The H20, which is Nvidia's, you know, chip that they designed for the Chinese market, they had sold, you know, tens of billions of dollars worth of these chips in terms of, you know, I think it was $16 billion was the number I saw in terms of prior orders of H20 chips. Well now the Trump first, the Trump administration banned the sale of those chips in April and then allowed them again a couple months ago, which we've been.
- [21:39]
B
Talking a lot about on this podcast.
- [21:41]
A
Yes, of course. Now what's interesting is that recent reporting coming out of China has basically said and Nvidia acknowledged this in their most recent investor relations filings. There have been no new H20 sales in China since the Trump administration allowed this. And I think there's, like, multiple reasons for that, the first of which is, and I think this is the most important part of the story the Chinese government is actively telling companies, do not buy H20s. But according to all the journalism on this story, when you're interviewing executives at these companies, they still want the chips. They still want Nvidia chips. So why aren't they buying them? Well, one reason is the Chinese government is telling them, don't buy them. But the other one is that they're holding out for the possibility of buying a next generation Nvidia chip. Recall that Donald Trump said that Jensen was coming to see me about allowing exports of a more advanced chip than an H20. So these Chinese companies are preparing to build, you know, massive, massive data centers. They're already working on doing precisely that. If those data centers are filled with, you know, Blackwell generation chips versus Hopper generation chips, the return on investment goes way, way up. So I think they're delaying their purchase to see, you know, what they're going to be able to get. And coming back to that, you know, July 2025, Morgan Stanley report, they did some estimates of, like, what foreign market share was of GPU sales in China versus what Chinese market share was of GPU sales in China. And they found that Chinese companies only had 11% of the market in 2021, 20% in 2022, 24% in 2023, and 34% in 2024. And this is in July. So after the H20 sales were allowed, but before Nvidia disclosed that none of them were happening, they said that allowing H20 sales was going to shrink Chinese companies, companies like Huawei or Cambercon, their market share from that 34% back down to 23%. So Nvidia sales, you know, are expected to be a significant part of the market in this scenario. And I think that's most likely going to be the case unless the Trump administration blocks the sales of the H20 again, which some Republicans in Congress are calling for. The Trump administration hasn't really given any indications that that's the direction they're they're likely to go. Okay, so Nvidia probably going to make a lot of money. The other question is, like, the Chinese competitors of Nvidia, and here I want to focus specifically on Cambercon. So Cambercon is a really, really strong group of Chinese chip designers focused on GPUs. Normally we focus on this show about Huawei, and that's because Cambercon has had almost no chip sales in China for the past few years. Because, you know, to the like, Huawei was hurt by the export controls, but Cambercon was, like, destroyed by the export controls. And that's because they were completely cut off from tsmc, and Huawei basically blocked them out of smic. So the best Chinese local chip manufacturer, Cambercon, couldn't even get access to them. Well, guess what just happened? Cambercon's stock price is up more than 5,500% since a year ago today. And we're recording this on September 11th. In fact, the stock has tripled just over the past two months. So, like, why is this happening? Well, the first is, like, now SMIC is making chips for them again, so they're actually able to sell chips. And I think that reflects the growth of capacity at SMIC and also the fact that Huawei is in the process of building its own fabs, which is kind of like a betrayal of smic. And so I think SMIC is, you know, more willing to say, like, hey, we need to be allocating some of our production capacity to Cambercon more than we were in the past. So Camber Khan's revenue, according to their, like, most recent investor relations filings, and this is a publicly traded stock again, so we learn things about it that, you know, we don't learn about Huawei because Huawei is theoretically owned entirely by its employees. You can't go buy shares of Huawei. But according to this public disclosure, Cameron's revenue rose more than 4,000% in the first half of 2025, which basically goes from, like, not selling chips to selling an awful lot of chips. And so you ask yourself the question, like, why do people want to buy Camber Con chips? Why wouldn't they want to buy Huawei chips? And here I think the best answer that we have comes from some reporting in the August 2025 Financial Times. And this is a quote from. I'm probably going to get this name wrong, but Lynn King Yen Yuan, he said Cambercon struggled to gain traction until the end of 2024, when it collaborated with ByteDance to make its chip more compatible with algorithms trained on Nvidia's ecosystem. And he's a semiconductor analyst at Bernstein. So recall when I was talking about all the challenges that Huawei has had with their chips in the past. Number one, it was a bad design. Like, they did not design their chip adequately for the LLM revolution in anticipation of demand. The chips were also, like, unreliable. It was very difficult to network the chips together to build very, very large clusters. But overwhelmingly because the surrounding software ecosystem was awful when it comes to Huawei. So if Cambercon has made a breakthrough in making chips that are compatible with all the software in the world that is written for Nvidia, like the Cuda ecosystem, that is a big, big advantage because there's so many folks in China who are currently using a lot of Nvidia chips, still kind of plan to be using a lot of Nvidia chips, but would like to buy more domestic. And if Cambercon now has the capacity to make their stuff compatible with Nvidia, then switching to Cambercon is a lot less intimidating than switching to Huawei has been, where it's like, rewrite your entire software stack so that you can adopt these lousy Huawei chips. And I think that explains a lot of the Cambercon boom. I'm sure there's additional complexity to that story that we're glossing over here, but I thought that was a really powerful point. Now, by the same token, Huawei does still stand to benefit from this. So I said before that, like, Huawei sales dwarfed those of Cambercons. I mean, if you look at 2024 data, which is available from the consultancy IDC estimates, they said that there were something like 600,000 AI accelerators sold by Huawei in China in 2024. And Cambercon was like less than 10,000 or something like that. So the prop. So Huawei actually had a lot of sales in 2024. The problem is that the Huawei chips were bad. So the customers were like government agencies who were ordered to buy Huawei chips. And I would encourage folks to go read the analysis done by Chinatalk back in February of this year, which talked about China's chip overcapacity. And it was basically just like people had bought a lot of Huawei chips to put them in these, like, 10,000 chip GPU clusters sort of distributed all around China, where it's actually really impossible to do anything interesting or useful. And so you have a lot of Huawei chips where the utilization rates are really, really low. That's what Huawei chip sales looked like for most of 2024. And I said that, you know, the 910B was a lousy design that's not super useful for LLMs. Well, you know, the 910C has been delayed again by Huawei. That's their next chip. And Then there's still the sort of question about like, what's the complete redesign? I do expect Huawei to come back in a big way at some point, but they're still struggling. And into that moment, you know, Huawei's massive investments in SMIC have suddenly led to Cambercon, you know, being a big beneficiary of that. So that then brings us to the next company that's probably going to make a lot of money off that and that is SMIC itself. Right. A Chinese chip manufacturer now in a position to make more AI chips for more types of companies. And that also probably has downstream benefits for Chinese semiconductor manufacturing equipment companies, companies like nara. So nobody's stock is up as much as Cambercon stock in the Chinese stock market. Like I said, their valuation compared to their earnings. This is like an eye watering number. So this is from the August 26th South China Morning Post quote, the frenzy over camera Khan, whose stock is trading at an eye watering, trailing twelve month price to earnings ratio above 4000 compared to 60 for Nvidia, reflects a growing belief that China, that the country is on a path to develop an AI ecosystem independent from US hardware. But I just want to dwell on that number, the price to earnings ratio. So like how much is your stock worth versus how much profit does your company make? For camera Khan, it's a ratio of 4,000 to 1. And Nvidia, which some people say is crazy overpriced, it's 60 to 1. So that's a lot of optimism in Cambercon's future growth that people are saying here. So other stocks, smic, nara, they're not like crazy Cambercon up, but they're up by a lot. And that reflects investor optimism in the Chinese AI future.
- [31:41]
B
Wow, those numbers are crazy. But let's shift gears now and discuss the national security implications of all of this. So the US has tried to slow China's progress with export controls time and time again. What does China's AI boom suggest about the effectiveness of this strategy?
- [31:57]
A
Yeah, so I would say that the export controls had an effect and are having an effect, but they're not having the scale of the effect that they could have had if they were implemented in their sort of most aggressive version, which was like what the US government originally thought they were doing. There's been a spate of good reporting coming out lately. You know, a lot of Biden administration officials are now out of government and a little bit more loose lipped. And there was a very, very good story in the wire. China by Elliot Chen and Noah, I think his last name is Berman. That came out earlier this week, sort of telling the story of that export control policy and naming names in some cases for the first time, which is pretty interesting. And one of the quotes that I thought was so interesting in that story that I hadn't seen anywhere else, is that it? There was a senior official in the Biden administration who claims that President Biden directed them in the October 7th export controls to, quote, go for the jugular. Right. So really, like, try and, you know, whack or more than whack, you know, China's AI industry in a big way. That's what they thought they were doing with the October 7th export controls. The reality is they didn't. And they didn't because they enabled the sale of the H800 chip. That's the chip that, you know, Deepseek used to train its models. They also enable the sale of an awful lot of advanced semiconductor manufacturing equipment that is now, you know, inside the SMIC facility. It's now inside the CXMT facilities, now in the YMTC facilities. So while, you know, we're not in the world where we could have been, like where China is buying EUV machines from ASML and Huawei is cranking out, you know, three nanometer chips that are competitive with TSMC's latest and greatest, that's a plausible alternative universe that we could have been in. It's plausible that the largest data center on earth could be in China right now. And Chinese executives, I've mentioned this before, but you know, the head of cloud computing at Tencent Deep Seq CEO, you know, it's not that these export controls don't hurt, but fundamentally, you know, we did not have a maximalist export control strategy. We tried to show restraint, especially after that first tranche and once all the lobbyists descended and tried to water down the policy. And so when you do stuff like tell, you know, forecast that you're going to, you're going to ban sales of high bandwidth memory in July, but you don't actually do it until December. Well, it turns out Chinese companies buy a lot of high bandwidth memory. So I think we're suffering through the consequences of the muddled execution of that export control strategy, incurring an awful lot of cost and incurring, you know, only a fraction of the potential benefit that we could have had. What else to say? I do think, and I said this on the, you know, some of the recent podcasts, I do think that China's commitment to that independence of U.S. technology. You know, that was a strategy that was fully, fully baked by the time President Biden walked into office in January 2021. So, you know, I've mentioned before, gosh, what's it called? I think it was the three, five, two strategy. But anyway, that, you know, the Chinese policy document that basically said companies have three years to get off of Western chips, you know, that was stopped by cutting off Huawei from tsmc, by cutting off Huawei from EUV machines. So that's the future that we, you know, could have had in the absence of some of the export control moves taken by the first Trump administration and the Biden administration. But, you know, where are we now? Well, the, the Chinese government is absolutely, positively all in on domestic substitution of foreign AI technology. So you know, on the AI software part of the story, I mean, that's been China's policy for, you know, decades, right? Like Amazon does not operate in China with its E commerce service. You know, Twitter does not operate in China with its social media service. That's because Chinese industrial policy, which they framed as security policy, has like completely blocked up those industries. Well, on the hardware side of the story, China's been trying to do that for a long time. But now we have some pretty interesting reporting about the specifics of what that's looking like in the post H20 allowance area. So here I'm just going to read from a Aug. 17 report in the South China Morning Post. Quote, publicly owned computing hubs across the country have been asked to source more than 50% of their chips from domestic producers to support the indigenous semiconductor sector. So think about that, what that means, that that basically means that all government AI data centers and potentially all state owned enterprises are facing a huge buy local mandate, right? This comes on the heels of a July 24 report from Reuters which found that the Chinese government's, quote, Ministry of Industry and Information Technology is collaborating with China's three state telecoms companies on ways to connect data centers in a network to create a platform that can sell the computing power. So basically what you're seeing there is like China Mobile, China Unicom, etc. All these big state owned enterprises and all the Chinese government ministries which are big, big data center buyers, they're being directed that at least 50% of the chips that they buy has to be local. And you know, recall earlier we were basically saying that these government agencies in the state owned enterprises, they can be like between a third and a half of all AI infrastructure spending. So if a half of a half of all AI Infrastructure spending has a buy local chips mandate. That matters a lot, right? In terms of, like, what the market prospects look like for companies like Cambercon and Huawei over the long term. And by the way, I think that that 50% mandate that is in effect right now, I think that's very likely to go up. That 50% mandate that now applies nationwide that was copying an earlier policy by the Shanghai municipality. And if you look at what Shanghai is doing, they just increased the quota to 70%. In Beijing, the mandate is that it has to be 100% by 2027, and in Guillaume, it has to be 90% by 2027. So these are big, big buy local mandates. And the point here is that, like, even though the Trump administration is willing to sell H20 chips, that is not enough to persuade the Chinese government that they should be, you know, trusting of the Trump administration or of Nvidia that they're going to be there in the future and that they should build, you know, all this infrastructure based on that. Another thing I guess I should say is that, like, there's a lot of build your data center in China mandates that apply to various AI applications. So if you think about, for example, the autonomous driving industry, well, it takes a lot of computing power to train the autonomous driving AI model that runs, for example, a Tesla car or a Waymo car. And Waymo doesn't operate in China, but Tesla does. And because of local Chinese law which says that your autonomous driving models have to be trained in local data centers, the AI driving model for Tesla in America and the AI driving model for Tesla in China is not trained at the same data center, not trained using the same data set and different in its overall performance because the one in China doesn't have the, you know, latest and greatest fanciest Nvidia chips. And so the point here being that like a lot of the big consumption of AI computing capacity, there is an increasing willingness to mandate that stuff be built in China. So if you think about, for example, a company like TikTok, you know, which is ByteDance, which was the company that we mentioned previously, helped Cambercon figure out how to be Nvidia compatible in its chips. Well, ByteDance, you know, leases a lot of capacity from companies like Oracle, et cetera, et cetera. You know, you could imagine in the future, the Chinese government says, hey, hey, hey, you know, more and more of that computing capacity has to be built and run on Chinese soil using Chinese chips. So really trying to build that all Chinese ecosystem, which, you know, the Chinese government has been trying to do for a very, very long time. Just as, you know, many of their made in China 2025 self sufficiency and local reliance goals from that 2015 era policy came true in other sectors. You know, they're now making progress on it in AI and semiconductors in a big way.
- [41:20]
B
Great. So now that we're a few weeks post the Trump administration's decision to allow H20 export controls to China, how has that decision changed all of this?
- [41:30]
A
Yeah, I don't think it has changed the story really that much. I mean, as I mentioned before, Nvidia is saying they're not having any H20 sales, any additional H20 sales in China. I think the real big thing is that China has decided that it cannot possibly trust the United States. And I don't think there's any apology that we're going to make or any deal that we're going to cut that's going to persuade them that they should not pursue self reliance as quickly as they possibly can. And so here I think the question is like, if we're in a competition for applications, right, do we want to help build a bridge to the Chinese future where they can supply more and more of this capacity locally? Because, you know, when you think back to that, you did the J.P. morgan estimates of what share of Chinese chips are going to be fulfilled locally, I think one question is like, how big is the market opportunity? If hypothetically, you know, there's no Nvidia chips, no AMD chips, no American chips whatsoever sold to China, they will build fewer data centers. And so maybe China's, you know, chip companies have a higher market share, but the Chinese market will be much smaller because the chip construction is the bottleneck. And in terms of fulfilling that demand for AI applications, a lot more of that will be fulfilled by American companies, you know, fulfilling that with servers that were built outside of China. You know, recall that there still is an overall big market shortage of Nvidia chips. They're sold out well into the future. So that's kind of where I think we are, is China is absolutely positively committed to going independent no matter what we do. And the choice that we really face is how easy do we want that to be for them and how much time do we want to give ourselves to build an edge and advantage in global adoption of AI applications?
- [43:36]
B
Sure. So what are American AI tech executives thinking about this? How are they feeling about all of this?
- [43:42]
A
Yeah, I think it's, it's pretty interesting. And one quote that comes to mind Here is from OpenAI CEO Sam Altman, who said in an Aug. 18 interview with CNBC, quote, I'm worried about China. There's inference capacity where China can probably build faster. There's research, there's product, a lot of layers to the whole thing. I don't think it'll be as simple as is the US or China ahead. And I think that's a pretty astute observation on his part. Right. In terms of inference capacity, you know, the JP Morgan analysis says, and I agree, that the bottleneck is chip availability. So they can build power plants faster than they can be built in the United States. All the sort of surrounding infrastructure of data centers, they can probably build that quite a bit faster than the United States. So if they have the chips, they can build inference capacity faster than the United States. In terms of research, you know, I think China has been number one in AI patents and in AI research papers for several years now. They're absolutely, positively world class. And it's true that most of the biggest breakthroughs are still coming out of American companies and Western labs. But nevertheless, you know, as Deep Seek showed, there's definitely a repository of world class talent in China. And then there's product. And here it's where the fact that OpenAI has, you know, 700 million weekly average users is a big deal. Here's where the fact that, you know, Microsoft, and if you think about like Microsoft Office is already installed on so many different computers all throughout the world. So like that is a pipeline to deliver AI enabled services where the United States kind of already has an advantage, but there's no guarantee that that'll be preserved over the long haul. So what I, what I will say is what I've said many times on this podcast, victory is not assured here. We could absolutely lose. And that's why it's so important that we do what it takes to race ahead and be competitive like China. They're not chumps.
- [45:55]
B
Well, we'll be keeping an eye on US China tech competition, obviously, and more on future podcasts, but I think that's a great place to stop for this week. Thank you all for tuning in and for keeping up with our work at the Wadwani Center. Greg, thanks so much for breaking it down again for us.
- [46:11]
A
Yeah, thanks. And great job on your first podcast. Thanks. Thanks for listening to this week's 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 spread the word. This podcast was produced by Sarah Baker, Isaac Goldston, and Sadie McCollum. See you next time.