Deepseek and China’s AI Ecosystem
24:11 Share

This is CNA Talks, the flagship podcast of CNA, a Washington, D.C.-based research and analysis organization. January 27th, 2025, New York, London, and Singapore fears that a low-cost Chinese artificial intelligence model would threaten the dominance of AI leaders, causes global investors to dump tech stocks, erasing nearly $1 trillion in value.

The model in question was launched by Chinese tech startup DeepSeek. Today, we examine what this model tells us about China's AI ecosystem, military applications for this technology, and other developments in China's high-tech sector.

Welcome to CNA Talks. I'm John Stimson. Today, we are going to discuss the latest issue of Intersections, CNA's news digest describing the interplay between the People's Republic of China, PRC, technology acquisition and defense industrial-based development efforts, U.S. response, and emerging technology rifts with military implications. And I'm excited to welcome Chris Cairns and April Herlevy back to the show to talk about this newsletter. Thank you guys for coming back on. It's great to have you here. Great to be here.

Thank you, John. Great to be here. All right, let's get right into it. Earlier this year, the PRC AI startup DeepSeek announced the launch of a new AI chatbot. The news sent shockwaves through the US tech sector, wiping out nearly a trillion dollars in value and causing Nvidia to lose over $500 billion in market cap, the largest single-day loss in history.

The DeepSeq chatbot is notable for its cost effectiveness. It supposedly can achieve similar performance metrics as other LLMs that use much higher and therefore more expensive computational resources. April and Chris, what is your reporting reveal about China's AI ecosystem and how that helped create DeepSeq?

Well, I think what's really interesting about this particular topic is just the speed at which with it's come to fruition. And I bring this up because for several years, CNA had an AI and autonomy newsletter. We published that from 2021 to 2022. And at that time, DeepSeek wouldn't even made the headlines. I went back through some old issues just to see

This company is so new and the technology has just come to fruition so quickly. That's one of the things that sort of makes this remarkable. Obviously, there's a lot of debates about what is this model really doing? What are the background characteristics? But what's important to note about China's ecosystem writ large, since you raised that question, is

So there's two things going on. There's decentralized activity. So companies like the one that has created DeepSeek are doing what tech startups do. They're doing research and innovation and trying to figure out tools that they can commercialize. But there's also this push from the top down in China to do all this work with AI. So there's centralized directives telling these companies, here's where we want the focus to be. So those two things have come together in the case of DeepSeek.

From the centralized perspective, what the government in part is doing is providing the raw resources that are necessary for China's AI ecosystem to grow. Above all, the supply of semiconductors, which as we know, there have been export controls placed on exports of U.S. semiconductor technology to China. And one of the stated goals of those controls has been to provide

slow down or even in some areas halt china's ai development by denying china access to both the quality so the state of the art as well as the quantity of chips it needs so something the government is trying to do to counter that is throwing large amounts of state investment money at

both China's domestic semiconductor development, but also in acquiring chips from abroad through a variety of different pathways. So without that supply, without those sort of hardware resources, the ecosystem wouldn't be able to grow. And I think one of the things we try to highlight in the case of DeepSeq is how has DeepSeq been able to grow so quickly? And there's, as April pointed out, there's still a lot of debate about how really innovative is their model or not.

how resource intensive is it or not compared to others, but they wouldn't be able to do what they can without at least some advanced chip technology.

April, I want to go back to a point you brought up, and that's this idea that DeepSeek was able to do like, you know, what startups do, which is like innovate the, you know, bottom up kind of approach. And I think that like looking at it from the outside, a lot of people who aren't, you know, China watchers would probably think that that's not necessarily how their economy functions. The impression that a lot of people have is of like a top down, you know, like a lot of state enterprises and things like that. So just this idea that like China...

does have that kind of a bottom-up innovative ecosystem. I think that might surprise people, so maybe you could elaborate on that a little bit. - Sure, I'd be happy to. And I think it's particularly emphasized that that's in the AI, the artificial intelligence and other areas. There are some other areas that are really bottom-up, but we saw it the most because over a decade ago, you had what they call the bad.

That's an acronym for Baidu, Alibaba, and Tencent. These names people probably have heard of. These are some of the biggest tech giants in China. And those companies were really at the forefront of thinking about how they can operate in the Chinese ecosystem. And this was before a lot of the national level, top level guidance came.

it relates to AI came down. So you had these companies, quite frankly, moving out more quickly than the government. So there's those big tech companies. But what else is really important to note is that there are particular locations where this ecosystem has built up. So for instance, Baidu is based in Shenzhen, China, which a lot of people might know because of the Shenzhen Special Economic Zone, which is very famous.

But another ecosystem that doesn't, I think, get as much play until recently is the city of Hangzhou. And that's actually where DeepSeek is located. And that's where Alibaba set up operations many years ago after they decided to move their headquarters from Shanghai. So there are these hubs where tech companies are coming together. I think it's an understudied portion of the Chinese economy. And I think it's something we, you know, there should be a lot more research on, quite frankly. Yeah.

Yeah, it kind of makes me think of like China's Silicon Valley. Although ironically, Chinese media calls both Shenzhen and Hangzhou China's Silicon Valley. So I think there's some debate about who's got the Silicon Valley in China. It's interesting. I'd be curious what people in each respective city say about that.

Beijing to that list. Oh, Beijing too. Special economic zone, which was one of China's first high technology development zones. And it's located right next to, um, Tsinghua university and Peking university, as well as several other universities in a district of Beijing called the high end district, which is, um,

It's known for higher education and has a large number of universities all clustered within a few miles of each other in that portion of the city. But it's also known for high tech development. So there's sort of an immediate presence of talent in a very small geographic area, not unlike what we see clustering around Stanford and Silicon Valley or MIT in the Boston area.

So, you know, your report's called Intersection. So we're going to talk about the military element of this as well. And how does how can deep seek be used in a military application? It's not just deep seek that the PLA wants to use. As the PLA enters what they call intelligent warfare, they're viewing a wide range of models because there's so many opportunities militarily to take advantage of AI, whether it's

Target, you know, target acquisition, command and control, logistics, intelligence, militaries across the world, and certainly the PLA is included in that, understand the potential value of AI. So they are using all the models. I think what's particularly attractive about DeepSeek is the idea that the computing requirements are less. And so that can help you do more in terms of running your models and getting the results you want.

I would just add one area that I have seen is in human UAV interfaces or human drone interfaces more specifically. There's been some research into whether, for example, a human could input commands which would be processed by an LLM. And then, for example, a swarm of multiple UAVs could coordinate to then interpret the

that human command and carry out some behavior such as attacking a target or conducting reconnaissance at a certain point. So that sort of human machine interaction is one area where LLMs are of interest.

That's really interesting, the idea that you could command a swarm of drones with an LLM-style prompt, basically. That sounds very science fiction. Just the idea of being like, you know, attack... I don't know, just as an example. Swarm that tank, and then the drones are going to do that. I think when you see things like those drone performances now, it's like a bunch of pilots working together to make that happen. It's not any one program. So that...

I don't know, it just sounds very science fiction that we could be using verbal commands to operate drones like that. At least I haven't seen any evidence that this is actually operational, that swarms of drones are actually able to take human commands, process through an LLM, and engage in these behaviors. So it's still a very nascent area of research, but definitely of interest.

And I just wanted to bring up the point that thinking about how to operate swarms is a huge area of PLA research. We have other colleagues in our China and Indo-Pacific Security Affairs Division that are looking at exactly that issue and trying to understand how they would be employed. And so what I was going to say is just there, I think there are questions about, it sounds like science fiction today, but if you look at how they're doing teaming in Russia, Ukraine,

that has moved much more quickly than we would have anticipated. So it may be a little while before we actually figure out how to use an LLM to direct a swarm, but those are the areas that are being researched. Right.

Okay, well, moving on from DeepSeek and into other policy developments, in your latest newsletter, you mentioned that the PRC announced new restrictions on the release of military-related content online. You know, what are these restrictions and why are they significant? Chris, we'll go to you first here. Sure. So, I mean, the restrictions can be viewed as part of a long string of both laws and regulations in the PRC regarding online content.

And particularly in the military context, there's a lot of sensitive information that the Communist Party worries about getting out or being portrayed in a way that is not in Beijing's perceived interests or is viewed as prejudicial to the Communist Party's image, especially with high profile scandals such as the recent

removals of leadership in the PLA rocket force, one part of the PLA that's been especially prone to political scandals recently, as well as just general turnover among a number of other PLA senior leaders. I think there's a concern that the Communist Party's preferred narrative needs

to be synchronized across a whole range of military and party outlets, and thus this focus on regulation. So I think there's two audiences for that. One is the domestic audience, so portraying a certain positive image of the PLA to the Chinese people.

that is consistent with the party's objectives and to the troops themselves to maintain morale. But then the other angle of that is open source analysts outside of China who are trying to glean information about the PLA from information that's online. And as we've seen more and more attention given by, you know, intelligence and foreign governments to open source analysis, it's understandable that you'd have this attempt at tightening control by the PLA to prevent some of that information from getting out.

Yeah, the idea of open source analysis. I mean, it's common. We do open source analysis at CNA. It's a common part of what we do. And it's not I know your team does a lot of it. The Russia's team as well looks at a lot of like telegram channels and things like that. And I'm sure that that is a mutual thing that's also being done at Chinese analysts looking at U.S. U.S. output.

So that area of wanting to control things certainly makes sense. Yeah. So going back for a second to the open source analysis, it's really amazing how much, even in a closed information system like that of China's, you can glean from open source analysis. A site that we monitor, part of our production of the Intersections newsletter is called the War Zone, which

contains a wealth of open source imagery about PLA development of ships and platforms and weaponry. Recently, the Warzone did sort of a feature story on the PLA's development of amphibious transport ships that some have speculated might be used as part of a potential action or even invasion against

Taiwan, and there's a wealth of imagery there. And some of it comes from Chinese social media, is often picked up by foreign social media. I've seen Japanese social media pick it up, other sources, and then ends up sort of digested on the war zone. And it's a very powerful information source. So you can understand why PLA and the Communist Party would be keen to reduce that at least. Mm-hmm.

Okay, so another policy reform mentioned in the newsletter was reforms meant to enhance innovation in military equipment related to scientific research, which sounds like the textbook definition of intersections. April, maybe you can tell us a bit about that.

Yeah, basically, this is building on what Chris has already talked about. I mean, the way the PLA, like many governments, puts together their policy is putting out regulations. In this particular instance, it's about newly revised regulations on military equipment and scientific research. The China's Ministry of National Defense put these regulations out. There's basically five key aspects to them that we list here in the newsletter. But what I think is really important to note about this is some of this is

actually a reflection of all the reforms that the PLA has gone through in the last decade or so. They've made a bunch of changes and sort of the nuts and bolts of how you do business, how you conduct research over that time has changed and it hasn't always caught up. Every time you make a big organizational change, then you have to match the bureaucracy to those changes and that takes time. So this is an indication of

that there's these larger reforms that have occurred, but they still hadn't got around to figuring out how do you actually manage this scientific research? Chris, anything you want to add on that point? Yeah, so to April's point about the management and organization of military scientific research,

being a part of the PLA's overall reform, they're looking, and by they I mean PLA analysts and PLA affiliated academics, are looking to the US in certain ways at how we organize military equipment research. If you look at DARPA, for example, or if you look at NRL, the Naval Research Lab, or you look at

O&R, Office of Naval Research, or the Air Force Research Lab, and how the respective military services sort of structure and guide what they want to get out of military scientific research. There's a lot of best practices that in some research that we've recently done on military

PLA views of US microelectronics development. So, you know, I think chips for semiconductors, you know, radar chips, all sorts of other chips. They look to DARPA in particular as an example of how to do that research right and how to structure those grants and projects to get the maximum benefit out of them. So, you know, we have to be very careful when we're trying to, if we attempt to mirror image

what the US is doing in terms of drawing lessons for PLA reforms. But I think in this area, it's definitely worth paying attention to what they're learning from what the US is doing.

I think that that reminds me of something that I, you know, Russia, Russian reforms to their military in the run up to the Ukraine war in many ways, like mirrored, like an attempt to go closer to the United States model of like smaller units, more efficient and perhaps like less of a, you know, less of a Cold War era mentality, you

You know, I think that like the the standardization of militaries, we see that all the time where, you know, there's a lot of effort towards copying and trying to, you know, duplicate what's been successful. So it's interesting, though, that you mentioned that this is an area we've seen that with China, but it's not a common, you know, it's not necessarily a common practice.

Can you maybe elaborate on that point a little bit, Chris? So one of the things we looked at is with respect to certain technologies, I mentioned microelectronics, is areas where PLA analysts perceive that the U.S. is effectively ahead versus not. And, you know, when it comes to the organization of research, that's an area where they're still, I think, seeking to emulate U.S. best practices for technology.

how scientific research is organized and funded. Whereas in another area, maybe AI development, that might not be the case because it's too new and it's still emerging. And there, the emphasis is more on how the PLA can out-innovate the U.S. or even leapfrog what the U.S. is doing rather than playing catch-up to U.S. practices. So I think it really depends on what domain you're talking about. Gotcha.

Well, just to finish up here, let's get back to semiconductors. We talked a little bit about them last time, but what are the latest updates in that area? April, you want to start us off? Sure. We talk about, in this issue of intersections, a pretty niche issue in the sense of one of the concerns about how China is getting some of its equipment to make semiconductors is

draws on a very like minor argument. So basically I'm actually going to read this to get the terminology right, because there's an issue with how

basically lithography equipment is characterized. So the European Union considers extreme ultraviolet lithography as dual use and sensitive. And so it should not be exported to China. The main company that exports these items is ASML, a Dutch company, but they also export what's called deep

ultraviolet lithography, a technology that is one generation behind Xtreme. And there was a Nikkei and Reuters report that there was concerns about those pieces of technology, the deep ultraviolet lithography being exported to China. So we actually cover that issue. Like I said, it all hinges on one word in terms of how you define the technology.

But it's issues like that that are being debated within the European Union about what kind of sensitive technology should or should not be exported to China. Gretchen, anything you want to add? Yeah, I've had a few points on that. So the first point is there's an ongoing debate about how

how broad we want restrictions on exports of semiconductor technology to be. And there are differences in positions between different countries on that. So as April was mentioning, the Netherlands government in particular adopts

the position that extreme ultraviolet lithography, so the absolute most cutting edge semiconductor lithography production technology, that's off limits, too sensitive, but deep ultraviolet, no, no, no, that's, that's older generation, slightly older generation. So, you know, maybe we shouldn't restrict that. Whereas I think

with the U S government, certainly under the previous administration, and we'll have to see where it goes under the current administration was adopting an increasingly expansive view where they started with certain technological specs of what was considered too advanced to export. And we've seen over time, those specs being broadened out as there's been a realization that China can still do a lot, even with, you know, one or two older generations of technology. Um,

So it's still very much an ongoing debate about, you know, how much is the right amount of restrictions. And I would note that from the company's perspective, they want to sell as much as possible. And so they have an incentive to export right up to the line of the technical specs that are permitted while still, you know, not being a violation of these policies.

Right. I mean, the companies at the end of the day, they're profit driven. Their goal is to make money. So more customers, more profits. But I'm curious, like, are these chips like upgradable if they're sold like the lower end version? Can they be can they be can you take that and build on it into a more, you know, into the more upgraded version or are they like really completely different specs? Well, that's not where that's not possible.

The key thing to know with ASML, the Dutch company, is they create the tools that make the semiconductors. And so we're talking about generations of technology that are the lithography tools to make those chips.

So the lithography machines are enormously complex and ASML is the only company in the world currently that manufactures extreme EUV, so extreme ultraviolet lithography machines. But they're also a major manufacturer of deep ultraviolet lithography machines. And so a large fraction of their business and revenue comes from that. So to them, this is a really important policy distinction.

Yeah, I think that's about all we have time for today. April and Chris, thanks so much for coming back on. It's been great talking to you. Thanks for having us. Thank you, John. Pleasure to join. For our listeners, there'll be a link to the latest issue of Intersections in the show notes. You can get into a lot more detail on everything we covered today there. And if you'd like to subscribe to Intersections, you can email intersections at cna.org. And that email will be on the show notes as well.

But I want to thank you all so much for listening. And we'll see you next time on CNA Talks. The views expressed are those of the commentators and do not necessarily reflect those of CNA or any of its sponsors. CNA Talks is produced, edited, and mixed by John Stimson. Our theme music is by Edward Granga. If you enjoy our show, we'd love it if you could give us a five-star review on Apple Podcasts and tell your friends about us. Thanks again for listening, and we'll see you in two weeks. ♪

♪