Who is Winning the US-China AI Race?

Image

Artificial intelligence is fast becoming a new front in global competition. Is this the defining U.S.-China rivalry of the 21st century, and what really matters in determining who comes out on top?

The release of China’s DeepSeek R1 model last year seemed to kick off a new realm of competition between the U.S. and China.

In this episode, Luke Bartholomew is joined by Robert Gilhooly, Senior Emerging Markets Economist, to discuss the nature of the AI race between the U.S. and China.

They cover who is winning now, whether government strategies around export controls and energy policy can determine who the winner might ultimately be, and whether framing the contest as a race with a single winner is the right way to think about how the technology and policy is likely to evolve.

Transcript

Luke Bartholomew: Hello, and welcome to Macro Bytes, the economics and politics podcast from Aberdeen. My name is Luke Bartholomew, and today I'm joined by Robert Gilhooly, Senior Emerging Markets Economist here at Aberdeen, to discuss the state of the so-called ‘AI race’ between the US and China. What's at stake in that race? Who looks to be currently winning? And what might determine who the ultimate winner is? All extremely important questions, both from a market and geopolitical perspective. So, Bob, Robert, thanks so much for joining us today.

Bob Gilhooly: Thanks, Luke. Good to be here.

Luke: So, I think a good place to start this conversation, and indeed the story more generally, is with the release of China's DeepSeek R1 model about a year ago, which rivalled US large language models but was trained, put together, at a fraction of the price. And this was a hugely significant development for at least two reasons. First, from a market's perspective, it seemed to imply that the competitive moats of many large US AI firms was much smaller than previously thought. You know, if these models could be trained much cheaper then the barriers to entry provided by that sunk cost, did much less to entrench the competitive advantage of the US firm. So, this caused a really quite significant repricing of various AI-related assets, including quite a large sell-off in US markets. And then the second thing, and perhaps more related to our conversation today, is that it seemed to open up a new front for US-Chinese strategic competition, because it implied that China was much closer to the AI frontier than previously thought, implying, you know, there was a genuine contest to be had here, a race worthy of that name, two competitors that really were in the game. And this is a contest of potentially fundamental significance, because if the development of AI does ultimately lead to the emergence of artificial general intelligence, you know, the Holy Grail, at least for some AI research programs, then this could have radical implications across a whole bunch of domains, including US-Chinese strategic competition. So that race between the US and China around AI has sometimes been compared to the 20th century superpower competitions between the US and USSR, you know, in nuclear, space, and arms domains. So, in the way that those races defined in many ways, the second half of the 20th century, perhaps the AI race between the US and China will define the 21st century. So, Bob, I guess my first question to you is, do you think that that comparison is appropriate? Is that the right kind of analogy for thinking about this contest?

Bob: Yeah, look, I mean, first of all, some of the political kind of rhetoric and framing does at least rhyme. Last year, President Trump declared that this was a kind of race in which the US ‘had to win’, as he put it. And that did have some echoes to, say, Kennedy's speech around the space race. And I guess the surge in just the amount of money being spent also does give it a sort of feeling of urgency as well. But otherwise, you know, I think many of the parallels here are actually fairly loose. The space race had a clear finish line, whereas the AI race doesn't, for the most part. Going to the moon, technically very difficult in the 1960s, and at least in some ways, you know, challenges were high, but surmountable, and not clear necessarily, I guess, kind of whether the creation of true artificial general intelligence is on the horizon or not. And definitely a key difference this time around is that private firms, I think, you know, they wouldn't have stepped in to finance and drive the space race. The R&D costs and uncertain payouts would have been very prohibitive. Whereas this time round, the commercial payoff has been very clear. And this has very much meant the US firms, in particular, but also more Chinese firms more recently have very much leant in hard without needing that kind of encouragement from governments. So, the biggest difference, though, you know, maybe that's an economic one. We could perhaps think of the 20th century races as, you know, partly being this kind of discrete outcomes, which had some commercial spinoffs. Whereas maybe AI is more we should think of this as a general-purpose technology, diffuses through economies over time. In that sense, maybe a more appropriate comparison might be the roll out of the adoption of electricity, rather than like the Apollo program, or the Manhattan Project, that were so key within the 20th century.

Luke: So even granted that this metaphor of a race is a little bit loose, in some respects, I think perhaps we could sort of maybe shrink the domain of the question and think about just which model is better at the moment, or at least which country has the firms with the best model? I mean, are we able to answer that question, in the sense of who's winning that much more narrow race?

Bob: Yeah, we can at least look at it across a few different dimensions there. So, the US large language models are seemingly slightly ahead as judged by reasoning, knowledge, maths and coding tests. Both China and the US in these rankings have, kind of say, five models in the top 10, US taking the top two spots, and five of its LLMs are kind of in the top eight. So marginally ahead, but you know, that is very close. And as you were kind of framing at the start there, you know, Chinese models have caught up very quickly to make it into those, kind of, top 10. Moreover, China's top large language models are almost all open-source models, unlike the US focused on closed models. Hence, we've actually really seen a very rapid adaption and adoption of Chinese models worldwide. And you just saw today reports that the open router data puts the three most used models by token consumption this month all in China's hands. So, we have Xiaomi's MIMO model, Alibaba's Qwen, and then also still up at the top there, DeepSeek 2. And in terms, you know, if we want to broaden it out a couple more dimensions there, US does seem [to] still be quite a long way ahead in terms of total compute. Estimates there put that around about 70% of global capacity. So that's effectively the kind of capacity, if you will, to run the models. And I guess the US does also seem to be winning the kind of spending race for what it's worth. And maybe that tells us a bit about going forward, US tech companies have spent more than 865 billion dollars since 2022 on AI-related capex already, and they plan to spend around about 600 billion dollars this year alone. It's a little bit difficult to compare this with what's going on in China because it's not just the tech behemoths in China doing all the work there and spending is a bit more opaque and, kind of, spread out across firms. But as best we can tell, it's likely that China is spending quite a lot less. So, at the moment, you know, I guess that kind of impetus and energy is a bit more maybe in the US core.

Luke: And as you say, I mean, one of the differences between this and the 20th century races is that it's much more a private sector competition, or at least the private sector is leading here. But I think that doesn't mean that states don't still have a pretty big role in potentially determining the environment in which those private sectors operate. And therefore, perhaps who's more likely to win even in those, sort of, more narrow domains as we've set out there. So of course, the US has its AI Action Plan. And then from the Chinese perspective, AI played a really quite prominent role in its 15th Five-Year Plan. So perhaps we can talk about both of those in turn, but maybe we'll start with the US. I mean, as you see it, what is the US's strategy, the government strategy with regards to AI?

Bob: Yeah, the Trump administration's AI Action Plan, I think does talk a fairly good game, highlights the need to cut regulatory hurdles, build out energy capacity, build out computing infrastructure, attempt to kind of export American AI globally. And then, you know, just integrate it across defense and federal agencies too. So in essence, you know, kind of boosting demand and opening the public purse where possible. But to a large extent, I think we could summarize it as kind of, get the government to stand aside and let the private sector get on with it without the burden of regulations. I mean, not [that] this was heavily regulated to date, but just kind of stepping back and saying, we're not going to heavily regulate you. It's maybe, kind of, one way to get firms to invest. That said, the US strategy is definitely not all kind of free markets – export controls on the most advanced chips, particularly Nvidia ones, and other semiconductor manufacturing technology are key levers the government is pulling to hinder China's progress. Export controls, I think, are likely to continue to give US firms a fairly strong advantage in training those frontier AI models. And that will make it a bit harder for, kind of, China to catch up or maybe even leap ahead. It certainly doesn't hinder other types of use though. So, your inference capabilities, i.e. running models, that's much less computationally intensive. So that can still go ahead. And then just finally, kind of alongside, I guess, kind of expanding the market demand across, say, the US government. One relatively new tactic has been to use AI as a geopolitical lever, effectively kind of helping to embed the US tech stack across more countries. In particular, compute seems to be becoming a key pillar of the US relationship with the Gulf. Indeed, countries in that region have had to almost, kind of, actively dump some Chinese technology in order to get access to frontier AI capabilities and US technology. But that seems to have been a decision to some extent that they've been willing and keen to take to get access to this cutting-edge tech.

Luke: So, as you say, the US perhaps talks a very good game, but maybe not all US policy is helping the AI roll out equally. I mean, one potential issue is that data centers are becoming enormous consumers of electricity, which is a constraint in and of itself for the grids to provide that electricity, but it's also putting significant upward pressure on electricity prices, and that's becoming a source of political contestation. In fact, public opinion does seem to be turning against them. So, do you think it's possible that politics per se and then the power-supply issue are going to be able to hinder the rollout of US AI?

Bob: Yeah, it's possible it might slow it down at the margin. Cost of living concerns ahead of the US midterms in November, certainly have been making the US AI rollout a little bit more difficult. We've seen some local legislation and proposals to regulate or even cancel data centers emerge. And even while President Trump has kind of been talking up AI as part of that national action plan, he's also said on Truth Social that big tech effectively needs to pay their own way to stop higher electricity bills feeding into households. I mean, for now, at least the hyperscalers broadly kind of seem happy to oblige and go along with this. Microsoft, for example, has committed to compensate for higher power prices and also, kind of, try to contribute in other ways to local communities that might be affected by their data center build-out. And other big tech firms are turning effectively to, kind of, behind the meter generation to effectively avoid the grid altogether. So, you know, effectively having their own little power source. But I think it's going to be hard to completely shield US consumers while also like really rapidly building out capacity. The US grid is effectively going from a long period of very minimal growth over like almost the last two decades to what we think of as probably being a sort of 2% if not more annual growth rate for the grid out to 2030. So, there's some risk here, I think, that you know, it's not just kind of the cost and local opposition. Maybe there's also a risk here that kind of people skills, industry capacity to expand the grid can effectively lag behind this sudden rise in demand. And that contributes to being somewhat of a bottleneck in terms of power.

Luke: And then by contrast, China has been rapidly expanding its grid since 2021. I mean, I think back then that might have been sort of counter-cyclical policy to deal with the real estate bubble deflating back then, but it has had structural payoff as well. It's boosted China's self-reliance and reduced vulnerabilities to imported energy, no doubt, something that, right now as the Strait of Hormuz remains closed at the time of recording this podcast, that feels like it is paying dividend. So how big an advantage is China's grid policy do you think in this AI race?

Bob: Yeah, I think to a large extent, I think this just means it removes the chance that the power grid becomes a bottleneck on the AI rollout in China. And I think it probably means it's going to further open up its lead in terms of lower electricity prices compared to the US. Now, of course, partly that will depend on how stress and strain plays out on the US grid. But it's also worth saying, you know, China was already working hard to expand its generation capacity even before the power needs of AI became apparent. You know, indeed, as you just noted, this was definitely a key lever that China was using to counter the drag from the kind of deflating of the real estate bubble. And actually, China's added something like a staggering 1,100 gigawatts of new solar and wind capacity since 2022, with about 400 or so gigawatts added in 2025 alone. At face value, that's getting close to as much as the new capacity as the entire US grid. And while I think we were already of the view that the authorities would be continuing to add a lot of renewable power capacity, the combined needs from AI and probably a bit more impetus to reduce dependence on foreign energy imports, given everything that's been happening in the Middle East, suggest the power rollout will probably go faster in China than we previously expected. So, I think that will probably kind of reinforce these, kind of, divergent trends in electricity prices more in terms of China's favor.

Luke: And you mentioned earlier, Bob, that US export controls of leading-edge chip technology has been a key part of the US strategy. And in fact, I think that goes back at least to the Biden administration. The broad consensus at this point is that China is indeed several years behind in advanced semiconductor manufacturing. So how much of a limitation do you think that is in practice on China's AI rollout?

Bob: Yeah, it's certainly been interesting that despite long-running very substantial subsidies and government encouragement, China has really struggled to close the gap on semiconductors, even before these, kind of, latest restrictions came in and the AI race more heated up. It feels like we've been saying that China's five-to-seven years behind the US for well more than five to seven years now, which maybe tells you something a bit about the industry, could maybe tell us the US has a kind of advantage here that's actually just kind of more fundamentally difficult to overturn. But thus far, US export controls certainly haven't been a firm limiting factor. If they were, we just wouldn't have seen this surge of Chinese model performance. That said, it wasn't necessarily because there were good domestic options for Chinese firms to substitute into. It was reported in the Wall Street Journal that when DeepSeek was being developed, they tried to use less advanced chips from Huawei and other domestic vendors, but the results just weren't acceptable. So, they actually ended up turning back to Nvidia chips for their training. Huawei has also reportedly been trying to develop a workaround by joining millions of lower-capacity chips together, which some have dubbed ‘swarms beat the Titan’. But it's hard to know whether that really will prove to be a durable solution, especially if you really do need the most advanced chips, if you really are wanting to push towards artificial general intelligence. I suppose it's not impossible that some US export restrictions could get watered down as part of President Trump's upcoming visit to Beijing. There certainly has been a bit of an active debate about whether US policy should be to try to get China hooked on the US technology stack. But access to the most cutting-edge chips still seems unlikely, and the US could tighten other aspects at the same time. For example, maybe limiting Chinese access to data centers located elsewhere in the world, which do remain open for use by China at the moment. Ultimately, I guess, though, all limitations will depend on what you want to do. Training the most advanced models is still going to need advanced chips or a feasible workaround. But if you want to do inference, so running the models, you just don't really need the most advanced chips to do that. So that will be less of a constraint.

Luke: And then finally, China's economy is still very different, structurally speaking, to the US. It's much more manufacturing-focused rather than services and consumption in the US. So, to what extent does that structural difference play out in China's AI strategy? And in what ways might that change how we should think about the nature of this, again to use a metaphor that I know you've challenged but it's still very helpful, this ‘AI race’?

Bob: Yes. I think China's ‘AI plus’ strategy, as they term it, it could maybe imply they're just running a different race. Maybe it's still a race, but just a slightly different one. They seem to be focusing a bit more emphasis now on the applications of AI and diffusing the technology through its very large manufacturing sector, rather than trying to achieve the ‘machine god’ of artificial general intelligence. Now, I suppose aside from US tech restrictions, maybe should we put it, encouraging it to go in that direction of applications, I think the pull from advanced manufacturing might have moved China's strategy in that way anyway. AI should be able to streamline processes, accelerate product development timelines within manufacturing, but China could also be better placed to implement what are called ‘world models’. And these use large reams of data to try to allow AI to understand the physical world. Very useful for autonomous driving and robotics more generally. The International Federation of Robotics actually thinks there are probably more than two million industrial robots in China. And the Chinese Communist Party is well aware of its upcoming demographic challenge, which gives it another reason to, kind of, maybe lean more into kind of robotics to augment a dwindling workforce. I suppose the other way to think about this kind of race is it's not impossible the runners diverge and then join back up further down the road if it's a particularly long race. But if this is what we're maybe looking towards, if strategies and emphasis really are diverging in the near term, maybe that will just make it feel a bit less like a race in the near term. Maybe that would also have the advantage of just, kind of, toning down some of the related geopolitical tensions compared to an outright sprint. So, there's maybe some benefits in that sense for the race diverging.

Luke: All right, well, that is all we have time for this week. We have run our race, so to speak. So as ever, please do allow me to remind you to like, subscribe, wherever it is that you get your podcasts if you have not already done so. And then I must thank Bob for joining us, for all his insights, and to thank you all for listening. So thanks very much and speak again soon.

READ MORE FROM ABERDEEN INVESTMENTS

Important information

This podcast is provided for general information only and assumes a certain level of knowledge of financial markets. It is provided for information purposes only and should not be considered as an offer, investment recommendation, or solicitation to deal in any of the investments or products mentioned herein and does not constitute investment research.

The views in this podcast are those of the contributors at the time of publication and do not necessarily reflect those of abrdn.

Aberdeen Investments Global is the trade name of Aberdeen's investments business, herein referred to as "Aberdeen Investments" or "Aberdeen". In the United States, Aberdeen Investments refers to the following affiliated, registered investment advisers: abrdn Inc., abrdn Investments Limited, and abrdn Asia Limited.

UNITED STATES RESIDENTS

The purpose of this website is to provide general information about the US-registered investment advisers which are part of abrdn, and the strategies they manage. The information provided is not intended as an offer or solicitation for the purchase or sale of any financial instrument.

Past performance is not indicative of future results, and there can be no guarantee as to the accuracy of market forecasts. Opinions, estimates, and forecasts may be changed without notice. This site does not provide financial or investment advice and does not take into account the particular financial circumstances of individual investors. Before investing, investors should seek their own professional advice. The views and opinions expressed are provided for general information only, and do not constitute specific tax, legal, or investment advice to, or recommendations for, any person. We suggest that you consult your financial or tax advisor, accountant, or attorney with regard to your specific situation.

AA-070526-207692-1

The InsuranceAUM.com podcast is available on