← The AI Policy Podcast
The AI Policy Podcast
China's AI Industrial Policy with Kyle Chan
01:25:23
Kyle Chan joins to break down China’s AI industrial policy, its tools, priorities, and impact on tech and US-China rivalry.
Loading summary
Transcript103 lines
- [00:00]
A
Foreign welcome back to the AI Policy Podcast. I'm Gregory Allen and today we've got an episode that I've been excited about for literally weeks now. I'm so excited that this is finally coming out and I can share it with you. We've got an interview with Kyle Chan, who we've talked about before on this podcast after his really, really cool op ed that was published in the New York Kyle is a super interesting guy. He's a postdoctoral research at Princeton University's Sociology Department and an adjunct researcher at the RAND Corporation, where he focuses on industrial policy and the geopolitics of advanced technologies. He's also a 2025 fellow with the Penn Project on the Future of US China Relationships and writes the excellent and very popular newsletter High Capacity, which is on substack. Kyle's work has been featured in Wired, MIT Technology Review, the Financial Times, and we're going to talk about that Times op ed, among other things. The main reason why we wanted to have Kyle on is that he just was the first author on the recently published RAND report, Full Stack China's Evolving industrial policy for AI, which came out on June 26. So, Kyle Chan, thank you so much for coming on the AI Policy podcast.
- [01:24]
B
Great to be here.
- [01:25]
A
Cool. So I want to get into this paper and your understanding of AI industrial policy for China, but I do to get a little bit of a sense of your own life story. You heard me just talk a bit about your bio. But sociology is kind of an interesting angle to attack industrial policy from and to tackle US China relations from. So how did you get interested in China? How did you get interested in industrial policy? How did you find yourself getting a PhD in sociology?
- [01:59]
B
Yeah, so I mean, so China, I have a personal connection. My parents are from Hong Kong and I'm part of that generation of sort of children of Hong Kong immigrants in the US who learn Mandarin rather than Cantonese because the parents thought it was more useful. And they were right in my case. But. But yeah, I actually began originally studying economics at UChicago. Very useful set of tools for for looking at sort of rigorous statistical problems, but I felt like not such a great set of tools for understanding Chinese industrial policy, Chinese infrastructure development, and the role of the Chinese state. Sociology offered me a way to have a broader sort of methodological toolkit. You know, it could use everything from statistics, corporate financial data, government reports, and then of course field work sort of on the ground interviews with government officials and other industry experts in China. And it also let me ask some of these big questions about, you know, how does this work? How does industrial policy work in China? What are the government organizations involved? How does the structure of the bureaucracy matter? How does this all piece together? So yeah, I felt like sociology gave me that sort of systems level perspective that, you know, no shade on my economist friends, but was sort of unique among the social sciences.
- [03:23]
A
Yeah. And I think just so I went to Washington University in St. Louis and one of their, you know, leading lights of their economics department when I was there was a guy named Doug north who won the Nobel Prize in economics. But he did it in a paper that honestly you could have published in a sociology department because it was about the institutions of the Soviet Union and how those institutional incentive structures really plagued the Soviet economy and eventually the post Soviet economy. And so I do think you're 100% right that this sort of systems level thinking and also I think sociology has a really great reputation for doing awesome fieldwork type research is a lovely thing to bring to the table on top of the traditional economics toolkit. And I've read your quantitative, you know, minded stuff too, so you really do have those, those chops. All right, let's start. And you know, folks who are in this audience have a diverse set of backgrounds. Some of them are, you know, AI juggernauts, some of them are more policy type folks. But let's just like start with the, the real basics here. Like what is industrial policy like? Explain it to me like I'm five.
- [04:33]
B
Yeah, yeah. So industrial policy is trying to use policy, state investment, perhaps state agencies, state programs to shape certain industries and to support their development, broadly speaking.
- [04:53]
A
Yeah, got it. So in the case of China, which obviously has had one hell of a run over the past three, four decades in terms of its changing economy, what is the history of, of industrial policy in China? And I, you know, I gave you a date to start at, but feel free to start wherever you want in the story.
- [05:14]
B
Yeah, I mean, I mean you can go back to the beginning of the PRC in terms of just the state involvement. But yeah, a lot of people would start to focus around the reform era in the 70s and 80s where you're shifting away from a more sort of command economy, Soviet style model to one where, you know, the government is still playing a major role, but trying to do so through market forces in a broader way. So you, you know, kind of like a shaping of incentives rather than sort of, you know, top down directives. And you know, through this whole process, actually a lot of the ministries themselves, they get turned into state owned enterprises or sort of seed control to parts of the private sector. So there's always sort of this mix between private and public in China where ultimately, right, the party's in charge, the party state is sort of guiding the direction of the overall economy, trying to support some of the inputs into the system, but is not doing so perhaps in the top down way that maybe some people would have thought historically got it.
- [06:18]
A
And so I think we're, most people are familiar that something called the reform era exists and this transition from a command economy to something that at least takes more advantage of market style, incentives and market structures. So what's the interplay here? What's an easy way to understand the relationship between like what is the Chinese economic policy versus what is Chinese industrial policy? And, and how does that, you know, come to the fore in industries like electronics or digital technology over time?
- [06:55]
B
Yeah, so, so yeah, there's you know, probably broader sort of macroeconomic policy, things guided by a monetary policy, things like that. But then there is really sort of targeted industry level policy. And there you have actors like the National Development and Reform Commission, the ndrc, formerly the State Planning Commission actually, or the Ministry for Industry and Information Technology. And these agencies, their goal is to kind of support industrial development for specific sectors. So you can take an industry like the auto sector for example, where China has been trying to build up a, you know, world class auto industry for, for actually decades now. So sort of long before EVs, long before like BYD started to dominate the headlines, you know, Chinese industrial policy, Chinese state agencies were trying to experiment with different ways of bringing in technology, perhaps through joint ventures with foreign firms. So Volkswagen, Toyota, gm, they all set up joint ventures. They had to if they wanted to enter the Chinese market. And this is sort of a crucial fact about all this. And you know, there is sort of mixed success along the way with some of these approaches. So the JV structure, it helped to bring in capacity, but it was widely seen as a failure in terms of building really world leading companies, world leading Chinese auto brands. And then you have sort of a whole parallel effort with electric vehicles and a lot of investment, both at the sort of basic research level in battery chemistry in electric motors as well as sort of industry support. And that includes everything from protectionist measures to keep out the then dominant Japanese and South Korean battery makers, or at least to make them sort of partner with Chinese firms and ultimately give space and support for Chinese sort of homegrown firms to take over. And so that's what we're seeing today. Companies like As I mentioned before, BYD or XPENG or NIO or LI Auto, some of these new EV companies, they're kind of the latest generation of this broader effort to promote auto industrial policy for many years now.
- [09:09]
A
Cool. And so when I think about, you know, communist governments and economic policy, I really think about, you know, five year plans, right, Which Stalin made very famous, Mao made very famous, and China to this day, right. Still has a very serious five year planning effort. How does it work for industrial policy? Is it like the same thing where like every industry has a five year plan, there's some official blessing of like this is what we're trying to accomplish or is it something else?
- [09:41]
B
To some extent, yes, there are industry level plans, but, but overall these plans are sort of like guiding documents that they don't necessarily, again, they can't kind of like dictate directly to especially private sector players. You know, you need to invest here, you need to do this. Instead they kind of set a bunch of targets, a bunch of goals, maybe suggest areas to focus on. And that I think perhaps more than anything else sends a huge signal to, to the broader sort of market system that this is where you should bet on. This is where the Chinese state is going to support you if you do move in that direction. And, and it'll be, you know, a smart idea to kind of row on the same team.
- [10:25]
A
So yeah, if you see like in the five year economic plan, right, like we're going to quadruple steel output, you know, from such time frame to such other timeframe, and you're like a private entrepreneur and you're like, okay, well if the state owned enterprises are going to try and quadruple steel output, well then maybe, you know, that will offer a great opportunity for me to, you know, make cars because steel will be so cheap and so widely available. Or maybe that will be a great opportunity for me to sell like blast furnace equipment because there's going to be so many Chinese enterprises are going to be doing this. And so, and I think one other part that you didn't mention is around aligning government executive incentives, right? Like every Chinese bureaucrat knows like they get a report card at the end of the year and that report card is like, did you advance the five year plan? And so that kind of clarifying of incentive structures around those goals kind of. Even without like nailing out the specifics of how we're always going to achieve those targets, people really do focus on them and they try a lot of things to try and make it happen.
- [11:30]
B
Absolutely. Yeah. And Especially at the local government level too, where this sends a signal to them. They might actually have their own plans that kind of align with the broader national plans and those might actually get more granular. They might get down to the level of actually developing investment funds or things like that. But yeah, overall this sort of aligns the broader system, government, private sector, all the above, towards. Towards sort of overarching national goals.
- [11:56]
A
Cool. So I want to start not with AI, but with digital technology, because now we have companies like Alibaba, which is sort of the. This is a very loose analogy, but it's like the Chinese equivalent of Amazon. There's other Chinese companies like Tencent, which is extremely strong in social media. There's other companies like Baidu, which does search in, in China, like Google. And a lot of these companies sort of grew up behind the great firewall and the blocking of western alternatives to fill those kinds of digital technology market niches. So can you sort of walk me through the history of how China goes from, you know, completely backward in digital technology to a point where by 2010, maybe they're not world leading, but they do have real TEC giants that are important players with meaningful capabilities. I mean, just as like one data point, right? Andrew Eng, who was a very famous Stanford professor and a really important person in the history of artificial intelligence and neural network research specifically, you know, he went to go be Baidu's head of AI for a while living in China. And the fact that like Baidu was in a position to credibly recruit and Andrew Eng sort of says something about the state of digital technology when China pivots and really starts focusing on AI. So walk me through, like what industrial policy was for digital technology in China. And you know, you decide what the starting point is in, in the story, but why it was at least in, in my view, right, in terms of if, if the metric that they're optimizing for is building strong companies with real capabilities, that was a success. So how did that happen?
- [13:55]
B
Yeah, so, I mean, you can even bring back further in time to the rise of the telecom infrastructure sector. You know, so now we all know about the rise of Huawei, but there was a time when literally just building out phone lines and then later on sort of basic Internet infrastructure was a very new thing and China was really catching up and felt very, very far behind. And a lot of that was actually originally built with foreign telecom equipment makers. And over time some of that got, you know, like domesticated or indigenized, as they say, where some of the infrastructure Got brought in house and through joint ventures again. So, you know, one of the most famous examples was Shanghai Bell, which was I think one of the first JVs in the telecom equipment world, where that was actually a JV with sort of a Bell Labs descendant in, in Europe. And the idea was to bring in that technology and then to eventually produce that equipment in China. And then you had the emergence of the telecom companies themselves, like China Mobile, China Telecom. And again, these companies in the early days were catching up, trying to roll out infrastructure for cellular technology and eventually they become the backbone for digital infrastructure later on. And the funny thing is for the rise of the Internet companies. So while the infrastructure itself might have been very sort of state directed, Right, I mentioned these Chinese telcos, they are state owned, the Internet companies themselves are private. And some of them sort of emerged against the wishes or sort of like outside of the usual official framework, like the rise of Alibaba actually with backing later on from Yahoo. So there's sort of interesting ties back to Silicon Valley, a time when there was a lot more sort of free flow in terms of people, in terms of capital, in terms of investment and then definitely a whole wave of what many people remember as sort of like the copycat industries or copycat companies in China. So, you know, you mentioned Alibaba is maybe. Yeah, kind of like the Amazon Baidu was sort of seen as the, as the Google Sohu or Sina. Like these were, you know, seen as the equivalent of Yahoo back in the day in some ways as web portals. So, you know, these, a lot of them sprung up kind of outside the, the rule of sort of official Internet policy. I mean, a lot of that was, was sort of not quite formed in the early days and there was a lot of skepticism and actually it reminds me a lot now of some of the shifting attitudes towards AI in China.
- [16:41]
A
So we' the government embrace of these digital entrepreneurs came later.
- [16:46]
B
Yeah, yeah. And once they really started to take off, then they were seen as these equivalents, these sort of like Chinese digital champions. And crucially, and this is. I'm really glad you brought this up. Crucially, over time, all the American tech companies, basically you can list sort of like all the ones that we use that we know well from Google to Amazon. Well, Amazon actually was allowed to go in there, but a lot of the other sort of Internet companies were gradually sort of blocked out. Google's the most probably most famous case. And this did two things, right. This kind of like sealed off the Chinese netizens from the rest of the Internet. Or at least created a filter. But it also, from a market perspective, created this sort of shelter domestic market where these Chinese tech companies could become major players in their own right. There's always like an interesting question of what would have happened had that market protective barrier. Not, not been there today. Yeah. So, you know, this kind of brings us up maybe to 2010. I mean, there's, there's a sort of a rise and fall pattern with Chinese tech companies, and we can get into that as well. But, but that, that's sort of like the, the early days.
- [18:01]
A
Okay, cool. So, you know, you said in 2010, 2012 is often cited as like the critical turning point date for modern AI. So 2012 is the ImageNet competition. And that is when a team, I want to say, Andrew Ang, was involved to some greater, lesser extent, although my memory is getting a little bit foggy at this point. But the basic point is this is a team that is using GPUs plus neural networks in the ImageNet competition, which is for image recognition. And they just absolutely blow everybody else out of the water. I mean, the amount of progress that they made that one year going from not deep neural networks to deep neural networks was like as much performance progress and accuracy of image recognition as they'd made over the previous, like five or 10 years using a bunch of different approaches. And so every tech company kind of sees the right writing on the wall that neural networks, machine learning, this is an incredible source of untapped potential. And suddenly you start seeing, you know, massive, massive money going into the AI sector by all the big tech companies. And so I think maybe we can start the story there. You know, when does the AI craze reach China and what the Chinese private sector entities think about it? What does the Chinese government think about it? And when does it start becoming a part of the industrial policy story?
- [19:44]
B
Yeah, so in particular, machine vision and language became really big focuses for Chinese industrial policy as it pertains to AI. And so you have the rise of companies like iFlytech, SenseTime, Megvi that were developing, focused on developing these models for processing images, sort of capitalizing on some of the breakthroughs with sort of American machine vision. And you also have sort of a surge in research at the basic level at places like Tsinghua and Peking University. And it's during this time when the Chinese state sees this as not only a useful tool, perhaps more broadly for economic growth, but especially also for surveillance and for data collection and data processing. And so Huawei is obviously also very much involved in this whole charge into the sort of machine visioning and language space. You have a combination of sort of state backed or semi state backed firms jumping into this, becoming sort of the original wave of. I forget the terms now, but I think they're called like, now they're called like the old dragons. There used to be a time when they were the new hot up and coming AI startups and now sort of, you know, they're the old guard says something about how quickly the space is moving.
- [21:15]
A
Yeah, I'm glad you connected the story to surveillance because 2012 also the year that Xi Jinping assumes power as the uncontested leader of China for his first term. And in that regard he's really prioritizing surveillance, right? He's really prioritizing centralizing domestic political control to an extent that was, you know, beyond even what China as a hard authoritarian country had already done. And AI offers an awful, really attractive option for him because it basically says, well, I, I often tell the story, right? Like, like in the book 1984, George Orwell has this thing called a telescreen, which is the TV in your house, but it's also a camera that's pointed at you. And the thing is you never know if somebody's watching your feed. You're never told. And that's because the totalitarian government has a labor shortage, right? They need somebody who's doing something other than watching, you know, one person per screen. But they could be watching you at any moment. You just never know at what moment they're actually watching you. And AI kind of solves that problem because there's no labor shortage. AI is watching every camera all of the time. Obviously very attractive to totalitarian countries or authoritarian countries. And what I've heard it described from some folks who have done business in China is they said there was that wave of AI startups that came after the ImageNet wave and, and American companies and Chinese companies both really were good at raising venture capital. China at this point already had a surprisingly sophisticated venture capital possibilities, not least of which drawing in American investments. But somebody told me that the difference between Chinese startups and American startups at this point is that both were able to raise investment, but the Chinese companies had revenue and a lot of it. And a lot of that revenue was government customers building out these massive surveillance networks, not just in places like, you know, Xinjiang, but literally all over China. And so since time I actually, you know, got to go Tour, you know, SenseTime's facility in Shenzhen and they were like talking about how many GPUs they had how many cameras they had. And it was a massive operation and it was, it was, they were able to ride that growth wave. And at the same time, you know, by like 2017, they're publishing world leading papers. So when you look at like the computer vision, facial recognition benchmark or like the speech recognition, you know, benchmark, it's totally routine at this point to see SenseTime or iFlytech, you know, at the top of those benchmarks.
- [24:06]
B
Yeah, definitely, yeah. And Huawei with its smart cities slash safecities program. And this is not only being sort of tested and rolled out within China, but this is also being exported, started. So companies like Huawei are trying to sell this to other countries. This sort of techno political model, if you will, of kind of surveillance and control, but also sort of like making urban services more efficient. Right, that's sort of the framing of it.
- [24:36]
A
So the massive surveillance build out is it, is it, is that like a case of unintentional industrial policy where like what they really wanted was the surveillance capability, but what they ended up doing was building this AI juggernaut. Like when does, when would you say AI becomes like an industrial policy focus of China?
- [24:57]
B
Yeah, that's interesting. Yeah. Because so I would probably peg it as as many people do, to the 2017 plan, which was sort of a big turning point. But the funny thing is, I think.
- [25:10]
A
The translation is like the new generation AI, you know, strategic plan or something like that. But this is like the official Chinese government AI strategy which came out in 2017.
- [25:20]
B
That's right, exactly. And the thing is, before that, AI had been part of many different source of plans, including made in China 2025, but it was sort of embedded in other areas like next gen IT or telecom technology or especially industrial automation, things like that. I would say that 2017, that, that AI plan in particular marked a turning point where it became a top national priority for Chinese industrial policy.
- [25:53]
A
Cool. So in that strategy at that moment, what would you say Chinese leaders believed about AI and then from an industrial policy perspective, what were they trying to accomplish?
- [26:07]
B
So I think this is not long after AlphaGo. And so AlphaGo famously is an AI program that can beat computers, can beat humans at go, and this was seen as sort of a leap over say Deep Blue beating humans at chess because of the difficulty and complexity of the game. And I think this shocked a lot of people in the tech industry in China and a lot of the policy folks.
- [26:40]
A
And there had been a alphago Lee Sedol the match when he defeated the world champion was 2016, so about a year before. So almost very plausible, right, that it directly led to some Chinese bureaucrat being like, you know what? We need a strategy for this stuff.
- [26:54]
B
Yeah, yeah. I mean, it was definitely talked about a lot in the tech community in China at the time.
- [26:59]
A
Well, I think. I think there's one other part of the story that's worth mentioning, which is in the same way that, you know, chess has a really important history in American culture and European culture, the game of Go has a really important place in Chinese culture. So, you know, the. The Garry Kasparov IBM Deep Blue match back in, like, 1997 was, like, watched live by, like, a huge, huge number of people. Go was like, mostly reported on, like, by the computer science community. Right, because, oh, this is really exciting what's going on with AI. But I think I. I think I remember reading somewhere that, like, in China, literally tens of millions of people watch that match live. So, like, the. The degree of, like, cultural significance that that Go match had in China was, like, dramatically superior to. To the impact that it had in America, where it was mostly the computer science community that was obsessed.
- [27:56]
B
Yeah, yeah, exactly, yeah. Go was just a whole different level of. Of significance culturally within China.
- [28:04]
A
So. Okay, so sorry, I took us off track here, but in the 2017 plan, what do Chinese leaders believe about AI and what are they trying to accomplish?
- [28:13]
B
So there they set out this idea where AI will become an industry in its own right. They had wanted to turn AI, the AI industry, into something like a 1 trillion R&B industry in and of itself. But perhaps even more importantly, they wanted AI to kind of turbocharge the entire Chinese, to kind of diffuse into a whole range of sectors. Healthcare, education, of course, manufacturing, some of the classic hard tech sectors, as well as even government services. And the idea was that AI would upgrade sort of all of these different parts of the economy and sort of the broader social structure writ large. So there is a sense that, you know, maybe perhaps different from conceptions of the race to AGI today, that AI would be, I don't know, maybe something more like the computer industry, maybe an industry in its own right, but also something that would just ramp up productivity across the board.
- [29:15]
A
And is it fair to say that that document reveals that Chinese policymakers, and I think it was published by the State Council, so this is a pretty important institution that's putting out. Did they think AI was something special at that time?
- [29:30]
B
Yeah, yeah. So it was seen as an area that sort of was its own category altogether, and a category that would Cut across sort of China's other efforts in industrial policy across sectors. So in a way it was sort of a meta layer on top of, you know, China's broader sort of techno industrial policy.
- [29:53]
A
And then what are like the big goals in that strategy? Like what are they trying to accomplish?
- [29:58]
B
Yeah. So the goals are to, number one, make China a world leader in AI, especially when it comes to innovation at the cutting edge. So to make, I think the language was something like to make it sort of the primary center for AI innovation in the world. That's on the one hand. On the other hand, it really emphasized sort of diffusion applications, the proliferation of AI across the economy. And so when you read through the document, a lot of it is sort of like breaking down specific areas, specific sectors where AI might improve productivity or might enhance some kind of government service or something like that. And you can see later on, again, this is sort of going back to the sort of national plan versus local plan. Later on, other local governments, they would create their own sort of AI plans in response and a lot of those would sort of mirror some of the language, but then maybe get a little more granular in terms of which sectors they want to focus on, which parts of AI they want to support.
- [31:06]
A
Cool. So now I want to talk about the industrial policy toolbox. Right. They've said, hey, AI is special. It's, it's going to be, you know, as important as computers maybe, which it turns out are pretty dang important. Yes. And we want to be innovative, we want to be world leaders. What is like the policy toolbox that they reach for to try and achieve these goals? And let's, let's start, you know, like in the time frame of shortly after that strategy and then we'll, we'll update to present day pretty soon.
- [31:41]
B
Yeah. So there's a number of things that they try to do. So one is trying to build out compute infrastructure. So the idea here is eventually this becomes sort of a national integrated computing network. But the idea is to have the state involved with the build out of data centers, fill out connections between data centers in a way that Chinese industrial policymakers have done in other sectors. Right. This idea that the state can play a role not just in providing capital, but in actually building some of the sort of basic underlying tools. And sort of within this project is the eastern data Western computing project, where they wanted to use renewable resources, renewable energy resources in the sort of interior provinces. So think about like Guizhou or Inner Mongolia or sort of far from the more developed coastal regions and build data Centers out there to leverage the cheaper energy, renewable energy, perhaps colder climates out there to run data centers more efficiently. And then models could be trained or even deployed from there to service customers back on the more wealthy east coast where you have sort of the demand for compute, essentially. And yeah, this was part of this broader vision of almost sort of like data centers as a utility or compute as a utility where you can.
- [33:25]
A
When you say like building data centers, do we mean like the government is owning and operating these data centers and providing them to both state owned enterprises and private industry? Or is it state owned enterprises are doing this or is it private enterprises are doing this? Are they like subsidizing it or are they actually like just doing it? What are they doing?
- [33:43]
B
Yeah, so a lot of this ends up getting actually rolled out at the local level. Provincial and municipal governments in partnership with actually some of the Chinese telecom companies like China Mobile, also in partnership sometimes with Huawei. And the idea is, yeah, to, you know, these, these are ultimately meant to be sort of public, public data centers that can be then leveraged by, you know, local governments to either offer compute to local startups or sort of feed into this broader sort of national grid of compute.
- [34:22]
A
Okay, so one thing they're doing is they're building a hell of a lot of data centers. What else are they doing? It?
- [34:29]
B
So on top of that, they are setting up what are called, what we try to call sort of state backed AI labs. And these are often again really initiated at the local level. So there's a few in Beijing like Bai, which is also known as Chiyuan. And there's, there's one in Shenzhen, Pengcheng, there's Zhejiang AI Lab, Shanghai AI Lab. So a lot of these are started by some kind of partnership between local governments like the Zhejiang Provincial government, maybe Hangzhou Municipal government, and maybe even private sector players like Alibaba. And also very close collaboration with universities like Tsinghua or Zhejiang University. And the idea here is these state backed AI labs would do sort of a whole range of more basic research standards, making sort of feed into the R and D efforts of the private sector. And they would do everything from develop their own sort of open source models directly. They might come up with evaluation systems, they might also be involved in AI safety discussions. So they kind of COVID a wide range. But one crucial thing that they do is they also serve as sort of a training and sort of network node in this sort of broader AI community within China. So they help to develop talent and sometimes you get startups spun right out of some of these state backed AI labs like Chu is from the Beijing Academy of AI.
- [36:25]
A
Yeah. And this, this is right around the time when China being a basic R and D powerhouse is like no longer a wish. It's, it's pretty much reality. I mean, like I remember when, when I first went to China, which was in 2010, the phrase that you would always hear is China cannot innovate, they can only copy. China did a ton of work to upskill its universities to actually start producing papers that mattered, to actually start doing research that was worthwhile. And by 2017 a lot of that stuff is bearing fruit in a big way. And so what you're saying here is that AI is going to be a huge part of the focus of the basic R and D effort and then there's a mechanism, an expectation that that's going to translate into private sector, you know, benefits and leadership. I guess one other thing here is just like the actual education part of the story and trying to produce huge numbers of people who have AI relevant skill sets. How is that going on?
- [37:36]
B
Yeah, yeah, so there are whole sort of centers set up at universities at the Chinese Academy of Sciences that are aimed at, you know, cutting edge research in AI. And these also again serve as sort of training centers for a whole new generation of AI researchers. And you can see, you know, so you know, these are backed with everything from government grants to again, a lot of actually partnerships with the private sector. And ultimately, yeah, the proof is in the pudding. When you look at say the co authors of papers for Neurips, you know, one of the top AI conferences or sort of any one of them, you see, see a lot of Chinese researchers and Chinese institutions, both public and private tech firms as well.
- [38:24]
A
So I think what the final thing that we haven't talked about, and feel free to tell me if there's other things I'm missing, but I think the other big tool in the toolbox that we haven't talked about yet is government guidance funds and being involved in the venture capital part of the story. So how does that work? What was going on around the time of this strategic plan?
- [38:42]
B
Plan, yeah, so, so in, in general, government guidance funds are used in a whole range of industries in China. And the idea here is that they're kind of like a VC fund where the general partner with a GP is say some government agency or set of bureaucrats and they sort of hold the initial seed capital, but they invest in a certain area, say AI or robotics or something like that, and they bring in private sector investment and local Government investment as sort of like the LPs. And the idea here is to have like a crowding in effect to sort of leverage up the capital that the central government has access to and try to direct broader, you know, sort of social resources towards target industries. And AI is one of those areas where you have a number, number of different government guidance funds. So most recently There is a AI specific fund that was announced actually earlier this year. $8 billion was sort of the scale and that was actually, you know, partly pulling some, some capital from the semiconductor fund. So there's actually a number of other funds. Some of the biggest actually are in semiconductors before this trying to do similar things. And then another huge fund that was announced just a few months ago was this sort of national VC guidance fund that was on the scale of a trillion RMV and again sort of $140 billion. And it was really an event, this announcement because it was announced by the head of the National Development and Reform Commission and called, I think it was the trans, like to translate it. It's something like carrier class, like aircraft carrier class level fund as in we have a bunch of guidance funds out there. But this one is, is sort of like the mothership and it's going to be the biggest one and it's going to be targeted at high tech areas including AI, including especially embodied AI, but also other, you know, areas like quantum fusion, things like that.
- [40:58]
A
Cool. So, so, so you're, you're, you're taking us to, to present day and I, and I want to get, ask one final question, you know, in the, in the 2017 era before we transition there, which is like if you are a startup in AI and you're located in one of these, you know, tech zones of Beijing or Shenzhen or Shanghai or what have you, like what are all the ways that you're benefiting from industrial policy? Like what are all the ways that the government is helping you, whether you know it or not? You know, like what is, what are all the ways that like industrial policy is, you know, putting, putting some accelerant in your gas tank or whatever.
- [41:42]
B
Yeah, yeah, yeah, that's a good way to put it. So some of them include sort of direct investment, direct financial assistance, whether they are. Yeah, sort of capital injections, loans, things like that. Also you benefit from sort of investment in, in the training, in development of talent, in the university systems, both in your local area and sort of nationally. You're also benefiting potentially from the compute infrastructure that I mentioned earlier. So in many cases local governments will offer these Compute vouchers where you can get sort of discounted access to compute drawing from especially these public data centers. And the idea here is that, you know, some of the big cloud service providers in China like Alibaba and Tencent, they kind of want to keep their compute for themselves to train their large models and stay on the cutting edge. So it's harder for startups to get access to that. So this is a way for them to kind of even the playing field or at least get, get a piece of the pie that otherwise might have been difficult. And then yeah, and then there's sort of broader sort of regulatory support assistance with sort of cutting through bureaucracy and just sort of setting up shop normal sort of business incorporation processes, things like that.
- [42:55]
A
That cool. Okay, so we talked a lot about this 2017 AI strategy document. Now we're in 2025. Did it work? Did they accomplish what they were trying to accomplish? Like where are we today in terms of China as an, an AI superpower, as Kai Fu Lee put it in his book, very memorably.
- [43:18]
B
Yeah, so, so here I, I would have, I would step back and think about industrial policy as it pertains to AI versus other sectors. So one thing that was very interesting doing this report is other people might be familiar with Chinese industrial policy for steel or shipbuilding and a lot of that might involve scaling up, pouring a ton of capital or directing a ton of capital into certain areas. But AI is different, we argue, because of the frequent paradigm shifts and the emphasis on cutting edge innovation. Like you can't just sort of like focus on one direction and sort of double down for year after year and hope that that will bear fruit and the ground sort of move from underneath policymakers and the tech industry in China with the emergence of ChatGPT. And there again previously China had been focused especially I mentioned earlier, on sort of computer vision and things like that. Generative AI was, there were some efforts to build language models. And so I mentioned actually some of the state backed local AI labs were involved with this. But ultimately a lot of people in China were caught by surprise. And some of the policies that were aimed at other sort of the older AI paradigm, some of that could be reused in part, some of the data center build out was somewhat useful, although that happened has sort of mixed implementation or execution, but some of the focus needed to be reoriented. And so yeah, I think what we see right now in some ways is in the past few years has been sort of a scrambling to catch up. So there were some factors that helped China catch up pre chatgpt. So I mentioned some, especially some of these things related to basic research and talent development.
- [45:14]
A
Yeah. Matt Sheehan, who is at Carnegie and we had on the PODC not long ago, talking about the Chinese regulatory part of the story. He said, and I thought he made this interesting point, which is until ChatGPT, China thought they were at least in a tie and maybe ahead of the United States in AI. And then that forced him to sort of recalibrate their assumptions.
- [45:40]
B
That's right, yeah. Yeah. And there was even sort of an ambivalence about generative AI in the early days. And I think Matt Chien might have actually talked about this before, where it was sort of seen a bit as a risk, a liability perhaps. You know, what is the content that's being generated in a way similar, not, maybe not totally dissimilar to some of the concerns in the us, except that the content we're talking about is especially content that could, you know, sort of the usual content that was censored in China on the Internet previously. So political content.
- [46:16]
A
Right.
- [46:17]
B
So, yeah, so there was a sort of catching up, a sense that like, actually China's now behind and while you do have this sort of like nice tailwind of all that investment in sort of broader AI scientific knowledge, there was a big pivot and a lot of the tech companies too really had to shift and sort of double down on and basically build their own LLMs in order to catch up. And now we see sort of the fruits of that bearing out. And, you know, the question is whether, you know, these efforts will continue going forward and what might actually be sort of critical for China to even have a hope of, you know, staying in the race. And, and one of those is actually chips and the export controls and sort of the, the, the, the potentially growing gap in compute capacity between the US and China.
- [47:04]
A
Yep. So in this paper you've got this lovely chart which for those of you who are watching on video, I'm holding up right now. But it's a doozy of a chart because it maps the AI stack. Right. Whether that's the physical applications of AI, the software applications of AI, the data, the training of the models, the R and D, the data centers, the underlying chips, you know, all these different parts of the AI stack. And then thinking about what are the industrial policy tools that are being used in order to support those various parts of the AI stack. So we talked a bit about, like, what AI industrial policy looked like in the, you know, ballpark 2017 timeframe. There's a lot of detail in this paper. And so I'm going to kind of throw it over to you as to like, what's the right way to organize the next phase of the conversation? Maybe by the, maybe by the segments of the stack, maybe by the tools in the toolbox. I don't know. But how would you characterize, like, what is Chinese industrial AI policy today?
- [48:09]
B
Today, yeah. So the top line summary is that China's applying industrial policy tools across every layer of the AI tech stack, from chips to compute to foundation models to applications, and especially into physical applications, including robotics. You can find some kind of policy program or fund or lever that they're trying to pull to kind of fuel progress in that area.
- [48:44]
A
Great. So just defer to. You talk about some of these elements of the stack that you think are particularly illustrative or illuminating and what are some of the tools that are being used using that.
- [48:57]
B
Yeah. So one area is actually chips that you could argue is sort of like an overlap between semiconductor industrial policy and AI industrial policy. And some of this is not, not originally aimed at AI per se, but has become very useful, if not critical for China's AI industry progress.
- [49:15]
A
I think it's probably worth, worth saying here. Right. Conductors featured very prominently in made in China 2025. Yes. And then in 2020 there was document number eight, which was a big industrial policy document for semiconductors. So the Chinese government has been, you know, pedal to the floor, cut the brakes on semiconductors for a very long time. Um, but unsurprisingly, right, the, the Biden administration's AI chip export controls really made crystal clear the relationship between AI and AI chips, the kind of chips that Nvidia makes. So what does the, the Chinese government, you know, see there? And what are they trying to accomplish and what tools are they using to accomplish it?
- [49:59]
B
Yeah, so in many ways they're trying to create sort of Chinese domestic alternatives to the industry standards and hardware, especially sort of Nvidia gpu'. I think the story, you know, you talked about actually a lot on this podcast and you've done some great work on this, of Huawei in particular and its Ascend series of AI chips, trying to offer an alternative that would be safe from or get around sort of US sanctions on export controls on Nvidia chips to China. And these are changing. So, you know, you can think about a chart over time about the performance of Nvidia chips that Chinese tech companies can legally access has been going down over time with multiple now generations of sort of performance degradation. Huawei chips are sort of gradually improving in their hardware capabilities. But what's interesting is even when you're starting to get a convergence between what's available from say, foreign chips and domestic chips in China, there's still reluctance of a lot of Chinese AI companies to make that switch over to say, Huawei. And so there's a whole bunch of reasons for that. Some of it involves CUDA and the difficulty of American tech companies in terms of moving away from the Nvidia platform. And there are other sort of factors, factors such as Huawei's own limited ability to produce chips to a certain extent or uncertainty about supply, chip supply for Huawei chips. But it actually extends even beyond just hardware itself. So I mentioned Cuda. Huawei has its own alternative can. Huawei actually has developed a number of sort of open source layers that are alternatives. So Mindspore is, is Huawei's sort of answer to Pytorch. Then there's even sort of Chinese alternatives to GitHub, like Gitea. OpenAI is sort of like an AI code sharing platform and benchmarking platform. There's sort of all these different Chinese alternatives. A lot of them are open source. And the crucial thing to understand here is that they're really trying to build up a comparable ecosystem system. And it's not just about sort of raw performance or raw capabilities. It's about building up the software libraries, the tools that are mature enough that you can sort of deploy pretty quickly. I mean, if you think about Chinese AI developers are similar to AI developers anywhere. They just want to make the coolest stuff as fast as possible. And they'd rather spend less time on tweaking some sort of less used Huawei architecture and just go with what they see as best in class and move on. And so there's this, this problem right now, sort of almost like a game theory problem where they need to switch as us expert controls become more and more stringent, but they really, really want to wait till the last minute to do that. And in fact, sort of, you know, Nvidia's potentially new, further degraded chip that they might introduce into China. It seems like there's still actually quite a bit of appetite, market demand for that among Chinese tech companies. So that tells you something about, about, you know, sort of the unwillingness to move away from Nvidia or conversely the lack of desire of shifting to say, Huawei platforms.
- [53:19]
A
Cool. So we talked about the data centers part of the story, which I assume is mostly still the same right now. Like they're just building a hell of a Lot of data centers. We talked about the semiconductor part of the story. What about the applications of AI, both software and physical applications?
- [53:36]
B
AI, yeah, yeah. So this is an area where there's a lot of activity both by Chinese tech startups and also by sort of more established players like Tencent and Alibaba as well as their spin offs like Ant Financial. So there are some companies like Bituan that are trying to focus on medical AI and diagnostics and they almost sort of like got pushed out by Deepseek and decided that they'd just rather double down on, on building out medical diagnostic data sets with doctors for example. Or there's efforts say with and Financial, which is a offshoot of Alibaba, trying to develop AI tools for personal finance, things like that. There are efforts to develop sort of AI tools for education, language learning. At this point there's, there's, it's almost sort of like every company is trying to put Deepseek in their model in their application as well. So you see this broader rollout, the industrial policy part of this is that you have a lot of support from local governments for these startups. Again I mentioned Access to compute, sort of these COMPUTE vouchers are useful, these AI pilot zones where startups can set up more easily and sometimes not even.
- [55:01]
A
Have to pay rent. Right. If you put it in one of the government at office buildings.
- [55:06]
B
Yeah, yeah. So it's funny because it's like AI seems to be so software digital, but it still requires some of these physical inputs like office space and physical proximity to other researchers and tech companies doing similar things. So that's where some of these local government initiatives play a role in supporting the application side of this whole process.
- [55:30]
A
Yeah. Cool. So one thing that we haven't talked about, and I'm not even sure it falls under the realm of industrial policy, but I think it's so fascinating and I've wanted to ask you about this for a really long time, so I'm just going to shoot my shot here. And that is the role of intellectual property protection and what you might call for lack of a better word, competition policy in China. And so I want to start by reading this passage from Kai fu Lee's book AI Superpowers. And he is talking about the post2010 rise of hyper competitive and cutthroat Chinese technology firms. He calls them, quote, copycat gladiators. But I think there's something really interesting here that has definitely been a through line of Chinese technology culture. And obviously the government has a role here. So here's the quote. It's a long one, so bear with us everyone. Silicon Valley may have found the copying of Chinese tech companies undignified and the tactics unsavory. In many cases it was. But it was precisely this widespread cloning, the onslaught of thousands of mimicking competitors. And this gets back to what you were saying Kyle, about Baidu being a Google clone and so on and so forth that forced companies to innovate. Survival in the Internet coliseum required relentlessly iterating products, controlling costs, executing flawlessly, generating positive pr, raising money at exaggerated valuations, and seeking ways to build a robust business moat to keep the copycats out. Pure copycats never made for great companies and they couldn't survive inside this coliseum. But the trial by fire competitive landscape created when one is surrounded by ruthless copycat cats had the result of forging a generation of the most tenacious entrepreneurs on earth. As we enter the age of AI implementation, this cutthroat entrepreneurial environment will be one of China's core assets in building a machine learning driven economy. So number one, do you agree with that story and is it still true today? And two, did this just happen or did the government try to make this happen, happen?
- [57:54]
B
Yeah, so I actually largely agree and I think that this sort of, yeah, I guess one word is copycat, another word is sort of diffusion of business models, ideas, features is very, very fast. And not just for AI, but this is where sort of like my other work on industrial policy helps me understand what's happening with AI. Because you look at a bunch of other sectors like electric vehicles today, where you have, you know, over 80, maybe even 100 EV brands in China all sort of duking it out, maybe barely making a profit of any, but sort of fiercely, you know, engaged in this domestic competition. And whoever emerges from that, you know, maybe it's byd, maybe it's some other sort of upstart that we don't even, we haven't even heard about today. They would be globally competitive and quite formidable in international market.
- [58:43]
A
This I think is a, this, this I think, think is, is really an astute point connecting it to EVs. Because if you think about industrial policy, it, it had a bad reputation in economics for a very, very long time because it was connected with protectionism and import substitution. And if you look at like, you know, random Eastern European countries, like what, what was at the time called Yugoslavia, right? Like they had a car industry and it was protected from foreign competition because if it wasn't protected from foreign competition, it was going to be obliterated. Because the cars were terrible. And so when you. And the point here being that like most of these economies, the domestic market is not large enough to breed a market that is competitive enough with multiple players achieving relevant economies of scale. Scale so that those companies can go out and be globally competitive. When you have small countries building industries that are globally competitive, it's usually a story of like a South Korea or a Taiwan where they're overwhelmingly focused on the export market and foreign technology transfers and investing in the country to upskill that industry is like a huge part of the story. What's different about China is that their domestic market is large enough and competitive enough that you can be born behind protectionist walls as Baidu was, as Alibaba was, et cetera, and yet grow up to be globally competitive, even though you started out for a very long time not having to compete on the global stage. And I think, like, what Kai Fu Lee is pointing out here is that like, you might think that that protection is a made, you know, them, them have to deal with fewer threats and therefore be less competitive. But actually, like, it was a, it was a coliseum and everybody's like dying like flies out there. And I think what, what you just pointed out by connecting it to the electric vehicle industry is, you know, China's division of power between the central government, which sets these objectives, and the provincial governments, which usually are the ones doling out the subsidies, has resulted in like so many local governments having their local champion. Yes. For electric vehicles. And so now you have those 80 to 100 companies that you mentioned who were all duking it out. So I don't know if this was like some genius level master plan right, by like the Chinese government, like, how do we create an epically competitive market? Let's have each province, you know, come up with two to three companies and then like, whoever survives that tiger fight, you know, will be the ultimate carnivore predator in the history of the earth. But that's kind of what's happening. I mean, that's what BYD is. They're like the tiger that survived the longest in this coliseum. And there's. I don't know if it's fair to say that there's an equivalent of that in AI, but you are seeing like very, very pro startup. They want to create a hell of a lot of companies. They want to create market opportunities for those companies, companies, and they don't, like, they may restrict foreign competition, but they don't restrict domestic competition, at least not that I could see.
- [62:03]
B
Yeah, yeah, yeah. I, I, that's, I agree with that. The one twist I would add to that is that there are other large economies like Brazil and India that also kind of struggled. I mean, maybe even Russia to a certain extent. They also had a degree of protectionism and sort of import substitution, but they weren't able to generate that kind of, of fierce internal domestic competition that yielded global champions at the level that China seems to be doing in, in other industries. And so, yeah, I think, I think your point, though is, you know, how, how deliberate was this? I do think that there is an effort to kind of like almost generate bubbles in the Chinese system akin to almost like private capital.
- [62:48]
A
So, like not, not a, not a, not a bug. This is a feature.
- [62:52]
B
Yeah, that's what I would argue that the idea is to kind of pour a bunch of resources or get, channel a bunch of resources into a priority area and that would, you know, maybe that results in a whole bunch of redundancy. You know, every province, every city is coming up with its own battery plant or solar factory or solar champion or AI company. And at some point, Right. It doesn't, it's not fun to be one of those companies. Right. But ultimately, at the sort of national level, it does create enough of that, enough of the resources and enough of that sense of competitiveness that allows you to then translate into sort of international.
- [63:34]
A
Yeah. Okay. I kind of like what you're saying about, like, intentional bubbles. Maybe another way to frame it is like the. Right. When you're, when you're, when your country is nobody in a, in an industry that is globally competitive, the right amount of investment is over. Investment. Yes. Yeah. You know, because if you, if you're like, okay, we want, I don't know, 15% of global automobile market share. Right. Like the Soviet system would have been like, okay, well, we need to build so many factories that can produce so many cars, and that number of cars will be 15%. Right. But like, you know, when you're trying to, you know, match supply and demand as like a bureaucracy, you know, you're probably just going to create these lousy companies whose, like, primary skill set is like, interfacing with bureaucrats as opposed to building awesome cars.
- [64:25]
B
Right, right.
- [64:25]
A
Whereas if you're China, you say, like, okay, we want 15% of automobile manufacturing capacity, so we should build enough factories to take over 35% of global automobile manufacturing, because 20% of that is going to be wasted on loser companies that go nowhere where the 15, like, the people who survive will definitely be capable of taking 15% of the market. Yeah, I think, I think you're onto something there.
- [64:51]
B
Yeah, yeah. And I think all this, including sort of the AI full stack industrial policy, I think a lot of this is sort of not so scientific and precise as it is sort of almost five spaced, right. Like nobody's quite sure including the policymakers exactly which layer is going to be the most important and they're not exactly sure which amount of funding would be, you know, sort of just the right amount. So in general they kind of do this kitchen sink approach where they're trying to attack sort of every level of the tech stack, or they're trying to attack every sort of segment of the supply chain when it comes to EVs or batteries. And they're trying to just make sure that there's err on the side of having too much investment and too much production rather than air on the side of not having enough. Especially given China always has a sort of catch up mindset that like technologically, economically, it still sees itself as catching up, trying to, you know, reach U.S. levels of, you know, what industrial development or technological development. So in that sense, you know, better to step on the gas harder rather than to sort of carefully modulate how much you're, you're going for it.
- [66:03]
A
Okay. So obviously, you know, China is an unfathomably massive place with a lot of diversity of opinion. Even in the inside the Chinese government. Of course there's diversity of opinion, although a lot of that is opaque intentionally, you know, to the outsiders. But to the extent that we could sort of oversimplify and rack and stack and rank China's goals, you've talked about self sufficiency, right? They want to have a Chinese option for every single element of the stack. We've talked about like global leadership in technology. They want to have the best stuff. We've talked about adoption. They just want to make sure that as many different segments of the economy as possible are using AI and adopting it as fast as possible. How would you rank those three goals? If Xi Jinping was to say, here's the thing I care the most about, here's the thing I care the second most about, here's the thing I care the first most about, and maybe I forgot some other goal that's even more important. So feel free to add to, to my list.
- [67:11]
B
Yeah, I, I, I would probably put at the very top. This is going to sound a little bit vague, but sort of like techno economic power or something like that, or national power broadly. So to the extent that the economy is bigger, that's good to the. But, but that's not enough. And to the extent that China is starting to move up the value chain in certain industries, that's also very good. But it needs to be sort of like across the board and it needs to be in a way sor mentioned before with the self reliance aspect where China feels like its, you know, development trajectory can't be. You know, they always sort of characterize it as blocked or knocked off, maliciously interrupted.
- [67:52]
A
Right, right.
- [67:53]
B
That's always how they put it. So like this sort of idea of just sort of moving up the. Yeah, there's obviously geopolitical implications to all this, but sort of moving up across all these different levels at once. And I think AI then is sort of another way of getting there. Another, you know, both a metric in and of itself of success and also a tool for, for reaching sort of a greater level of like techno economic development.
- [68:22]
A
Yeah, I think that's very well said. So I certainly agree self sufficiency is a massive goal. I certainly think that they, they still care about leadership. Although parts of the self sufficiency thing kind of make me think that they're willing to trade weaker technology in exchange for more self sufficiency. Certainly in the short term. One that we haven't talked quite as much about is the adoption part of the story. You know, I remember after the sort of Deep Seek hype cycle that happened in, you know, January and February of this year, the CEO of DEEPSEA got a meeting with Xi Jinping Ping and then I'm blanking on the name of the Chinese government agency, but some Chinese government agency basically directs all the state owned enterprises, like starting using Deep Seek in everything that you do. Can you talk a little bit about like, and sorry, please repeat the name of that organization. But can you talk a little bit about like what the Chinese government is doing to drive adoption of AI?
- [69:29]
B
Yeah, yeah. So that organization specifically is sesac, the State Assets Supervision Authority, something like that. I always pray the full acronym. But basically it's sort of almost like this state holding company that actually holds the most powerful central state owned enterprises. So a lot of the oil, energy, telecom, you know, these are all sort of under that CSAP umbrella.
- [69:56]
A
It owns huge equity shares of all of these state owned enterprises or this is the organization that owns those.
- [70:03]
B
It's like the parent organization for all, all of them. And then what they can do as the sort of parent organization is they can tell their companies to do certain things. And actually before AI there was a push by them to increase Carbon efficiency, for example. So AI is now a new area where they're saying you must all come up with a plan for rolling out AI in some form or fashion. And that's especially helpful. Helped with Deep seq's open model, open weights model. So that is something that is a lever that I think is different than what you would see certainly in the American system. But more broadly on this adoption question, I would tie it actually a lot to the open source push because I think they're very connected where now we actually see a shift among some of the larger tech companies, private tech companies in China. But actually a lot of tech companies in China like this open source approach because yeah, a lot of their, their main problem is building up a user base, building up an ecosystem, building enough up, up enough sort of dependencies that other companies would want to make that leap and try a Chinese platform or try their specific platform and similar also with a lot of the, the models and tools that are being developed by these state backed AI labs I mentioned earlier, almost all of those are open source. So there's this idea that, you know, you, you create sort of like limited barriers to, to adoption and that will help diffusion. At the same time you have sort of like, you know, if that's sort of like removing barriers, then you have sort of more proactive things like sesac, but also at the local government level, you know, Beijing has a plan, Shanghai has a plan. You see different municipalities roll out AI adoption plans. You know, they actually try to spell out, you know, what aspects of the municipal government services might be automated or might use AI in some form. So there's definitely an effort both sort of trying to break down the barriers, but also trying to push for adoption, especially at the local level.
- [72:20]
A
Very cool. So now I want to ask you to pull out your crystal ball and let's talk about where we might be headed, where China might be headed in this story. So, so China has this industrial policy towards AI that you've characterized. To me it seems at least directionally consistent with the 2017 strategy, which is full steam ahead.
- [72:49]
B
Right, right.
- [72:50]
A
It doesn't seem likely that they're going to turn that off anytime soon. There is of course US efforts to try and slow China's Russian rise coming from multiple angles. So where do you think China is headed? And I guess here I'm kind of asking you to bring in your, your New York Times piece where I don't want to spoil it, but you can talk about it. So where do you think China might be in terms of AI and competition with the United States or whatever in a year, in five years, in 10 years. And let's assume know both the US and both China mostly stick with the approach they've had now and then maybe at the end we can talk about some policy interventions that you would recommend.
- [73:36]
B
Yeah, yeah. So I see a number of different trajectories, but one big one could be that China would end up doing very well in their race with the US and AI maybe keeping pace or you know, in some areas perhaps in terms of sort of model development despite sort of reduced access to compute, you know, might have some innovations there that would be, would be new to the American system.
- [74:11]
A
Now wait a second here, Kyle. You know the title of your New York Times piece is in the future, China will be dominant. The US will be irrelevant. And what you're saying here sounds a little bit more of a mixed bag. So you know, talk us through the difference between your, your title and what you just said.
- [74:28]
B
Yeah, yeah. So I mean overall actually aside from AI, because there's, there's a big wild card there with the compute. I do see that in a, in a whole range of industries. We're talking about EVs, auto energy, renewable clean energy, we're talking about traditional industries going back to shipbuilding, we're talking about also even aircraft production. There's, there's a huge push in China, industrial policy push in many of these areas and there's a lot of progress being made. And in many of these areas China is either catching up or even possibly, you know, arguably starting to edge ahead and rolling the picture forward. I argue that because China is sort of doubling down on its sort of model of state supported but still, you know, private sector, you know, driven, you know, sort of this combination of both innovation and sort of scaling up that China will like either be a leader or a close competitor with the, with the US and with Europe and Japan and other countries in a lot of these, in a lot of these fields. That is one where I have to hedge a bit because yeah, there, there's a lot of, there's some big wild cards here.
- [75:47]
A
Yeah, because I mean in the piece you, you said, and these are juicy quotes, so I do, I definitely encourage everybody to go read the, the piece. But you said America, by contrast, may end up as a profoundly diminished nation sheltered behind tariff walls. Its companies will sell almost exclusively to domestic consumers. The loss of international sales will degrade corporate earnings, leaving companies with less money to invest in their businesses. And then it goes on traditional high value industries, such as car manufacturing and pharmaceuticals are already being lost to China, China, the important industries of the future will follow. Imagine Detroit or Cleveland on a national scale. I mean, that's a profoundly bleak picture. And you're saying that actually AI might be one of the bright spots where the United States is more competitive than it is in say, car manufacturing or pharmaceuticals. Is that, do I understand you correctly?
- [76:43]
B
Yeah, yeah. So, yeah, so part of the story for that New York Times piece was also looking at the US research trajectory. And so I was trying to draw this contrast where China was doubling down on funding for basic research, R and D, on tech innovation, and forging ahead on sort of all these industries, while at the same time in the past few months we have in the US a cutting back on the very things, the very factors that made the US a superpower. So that's again, a lot of these same factors. Investment in basic research, creating an open environment for foreign researchers, including actually ironically, Chinese researchers who are very active in the US AI ecosystem. And so I just saw that if we roll that picture forward, if we roll forward the trends of the past few months in the US with sort of these cuts and these crackdowns on the sort of core factors of production is one way to put it for American leadership. While at the same time China's doubling down and making actually quite rapid progress in a whole bunch of areas that then, yeah, in 10, 20 years, this could have not just economic consequences in terms of lost corporate revenue, but this could have consequences in terms of jobs, especially the higher paying jobs that might be attached to high tech industries. And even the sense that U.S. scientific and technological leadership up until this point was taken for granted. A line that didn't make it into the. The piece was we take for granted that the next Nvidia or Google will be started in the US and now that's not as, as certain. So, yeah, part of this is sort of rolling forward the trajectory of these two countries based on especially what's been happening lately in the US and in.
- [78:37]
A
AI specifically do things like, you know, OpenAI announcing somewhere between 100 and 500 billion dollars for Stargate, other, you know, big tech companies announcing these massive pools. Is that what gives you confidence that the situation might be better in AI or is it something else?
- [78:57]
B
Yeah, so I think for AI, the compute is definitely a huge factor, both in terms of you just have incredible spending on compute capacity and expansion of compute capacity in the U.S. i think there was some estimate that this year alone alone, the largest four US tech companies will spend like over $300 billion or something like that on AI infrastructure, including data centers. And that seems to be.
- [79:29]
A
A large.
- [79:30]
B
Amount of investment going forward. At the same time, China's compute situation is dependent on US export controls and China's ability to get a around them. So here, while the US is sort of like charging ahead with more compute, and also the chips are getting better, right? So they're able to take advantage of increasingly sophisticated AI chips, especially from Nvidia. China's getting less access to those chips, getting more degraded versions of those chips, is still trying to build up its own alternatives which are not anywhere near the level of what's available available internationally outside of China. And that gap is growing at the per chip level and that gap is growing at the sort of national compute level potentially. And so, yeah, you know, one, you know, related wild cards are whether the US will sort of ramp up export controls even further, make them more stringent, and on the Chinese side, to what extent Huawei or other GPUs makers in, in China might be able to develop comparable offerings or at least enough to, to sort of stay in the race.
- [80:43]
A
And what I have started saying about Huawei specifically as a competitor to Nvidia is, you know, Nvidia is going to go from phenomenal awesome chip to even more phenomenal awesome chip chip. But Huawei, the Ascend 910B is like a bad chip. And it's not just bad because of export controls. Like, Huawei had a lot of design optimization decisions for like what types of workloads they might have envisioned their customers doing it and like it was not especially well designed, you know, to run LLMs of the architectures that we run right now. And the Huawei competitive to Cuda is like a train wreck, at least for right now now. And so while Nvidia is going to go from, you know, phenomenal to even more phenomenal, Huawei is going to go from non viable to viable. And that actually might end up being a more important strategic transition because like Huawei made a lot of dumb decisions with designing the 910B, but they have smart enough people to fix those decisions. And the Cuda, you know, alternatives are terrible. But that is the kind of problem where if you just throw 10,000 software engineers at it for a couple of years, they'll make a boatload of progress. And so yeah, Huawei is in a very, very, very tough spot right now. But I do think that they're going to make this non viable to viable jump and that's like something really important. To watch over the next, you know, two to four years.
- [82:16]
B
Yeah, yeah, yeah, exactly. And I, and I think also kind of related to all this, you know, questions about pre training versus inference or test time, compute. You know, so some of the Huawei essentially chips have been used more for inference. And to the extent that that could be an area where Chinese tech companies can make progress on more versus sort of scaling up just at the pre training level, that could be another factor. There are also questions about sort of the future of the AI industry in general, about whether, for example, agents will mean that cutting edge models are more important than ever, that they actually might not be commoditized, that it might really matter that every step of the way you're doing it 99.999% accurately versus 99% accurately. So, yeah, I think, you know, in a world where, where that doesn't matter as much, then China might have sort of good enough chips. You know, if, if Huawei can get their act together, if sort of the ecosystem can be slowly built up on, you know, Chinese domestic platforms, then they might be okay. The question is whether there could be even future paradigm shifts. Compute could be even more important. So it can go in many different directions here. So that's why I sort of hedge on the AI issue. It's hard to extrapolate in the AI world from even just the past few months.
- [83:49]
A
Well, this is why you're a rigorous scholar at a fancy university and not a, you know, classless DC pundit. You know, just throwing out hot takes and expressing absolute confidence in all of them for no, no real reason. But thank you for laying out the possible scenarios of the future. Thank you for, you know, walking us through the, the past, present and future of Chinese industrial policy towards AI. Kyle, I thought it was an awesome conversation and I just want to give two, you know, shout outs here, first of which is the, the full paper is available. Kyle co authored it with Gregory Smith, Jimmy Goodrich, Gerard DePipo, Konstantin Plitz, and all good people and highly recommend reading that. And I also highly recommend reading Kyle's substack where there's lots of good stuff on so many different aspects of China's economy, industrial policy and US China competition, not just in AI. So, Kyle Chan, thanks so much for coming on.
- [84:53]
B
Thank you. This is really, really fun.
- [84:55]
A
All right, take care.
- [84:56]
B
Thanks.
- [84:59]
A
Thanks for listening to this episode of the AI Policy podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform. And subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next time.