Who’s winning the AI race – and does it matter?

AI is often compared to the nuclear or space race — but is that the right analogy? We explore how the US and China are competing in AI, from models and semiconductors to power grids and government strategy and ask whether this is really a race at all — or something far more complex.

Transcript

Nick Robinson

Hello, this is Nick Robinson from Aberdeen and you're listening to the Emerging Markets Equity Podcast. Together with our expert guests, we'll dive into the driving forces behind emerging markets and uncover the opportunities that are shaping the future of this exciting region. So just over a year ago, markets were shocked by the launch of China's Deepseek R1 model, a credible rival to US large language models, yet developed at a fraction of the cost. This event knocked a trillion dollars off stocks and opened up a new chapter in US-China strategic competition. Since then, China has made significant further inroads into the top 10 language models and has emerged as a global powerhouse in the development of AI applications. So where do the US and China stand on key elements of the AI race? And what do differing government strategies imply for both nations? To help answer these questions and more today, I'm delighted to be joined by my colleague Robert Gilhooly. Robert Gilhooly is a senior emerging market economist and has recently written a paper on the subject of the AI race between China and the US. Welcome back to the podcast, Robert Gilhooly. It's great to have you back on.

Robert Gilhooly

Thanks, Nick. Great to be back.

Nick Robinson

Brilliant. Well, there's a lot going on in AI. Why don't we start by just thinking a bit about what makes this different from the nuclear and space races?

Robert Gilhooly

Look, I think the nuclear and space race analogies capture some of the kind of perceived or at least kind of potential stakes here. But maybe it kind of misses some of the mechanisms and also the implications too. The space race, you know, culminated largely in a symbolic victory for first the moon. Nuclear race kind of stabilized into something more akin to deterrence. There were, of course, commercial spin-offs from both, but to some extent anyway, they were a bit more akin to kind of binary or discrete outcomes. And in contrast, AI, we hope, should be more of a general-purpose technology that kind of slowly diffuses across multiple sectors of the economies over time. So maybe, you know, the analogy actually is a little bit more appropriate if we think of kind of electricity, how those revolutionised economies or the Industrial Revolution rather than a kind of Manhattan Project type event. And definitely a key difference this time, which I do think is important to emphasise, is that private firms wouldn't have stepped in to finance and, say, drive the space race, for example. You know, the R&D costs, uncertain payouts would have been pretty prohibitive. Whereas this time, the commercial payoff has really been very clear, which has meant that US firms in particular, but also Chinese firms, have lent in very hard in this case with some very substantial investments.

Nick Robinson

So why don't we move on to the US and China AI landscape? So, I suppose if listeners were to ask who's winning right now in terms of who has the best models, what do you think the honest answer is? And what I suppose explains China's sudden catch up that they made.

Robert Gilhooly

Yeah, I mean, US large language models are probably slightly ahead still, as judged by kind of reasoning, knowledge, maths, and coding tests. And there's an interesting leaderboard here produced by Open Compass. And in that leaderboard, you know, both US and China actually have five models in the top 10, but the US is occupying the number one and number two spots, and its five large language models are in kind of the top eight positions. So, that does give at least perception that the US is in the lead, even if it's very close. But, as you kind of led off with on this podcast, there have been very rapid progress in China's models. A key difference between China and the US is most of theirs are actually open-source models rather than the US closed models. So potentially there, they could be spurring more rapid adoption of their models worldwide. So, it's not just about, I guess, kind of who's got kind of the technically best models, but it's kind of also kind of how these are kind of diffusing around the world and kind of kind of what sort of pace. I mean, in terms of the catch-up look, it was it was a big surprise. Perhaps, maybe it's always easy to say this in retrospect, I suppose, but perhaps, we shouldn't be surprised if we look at kind of China's human capital within this. It's well known, you know, China has kind of the largest number and percentage of STEM graduates coming out of their universities. And China's also been pretty close behind the US in terms of the numbers of top rank AI researchers indeed just looking at some recent research suggesting that this is a kind of crown that they took in sort of early 2025.

Nick Robinson

Suppose one thing we've kind of seen is just the extraordinary amount of money that's been pouring into AI infrastructure. I mean, the US now controls something like 70% of global compute. And I think in your paper, you argue that China's true spending on AI is somewhat understated because of investment being spread across various state-backed firms. You know, do you think, if we were to look at those headline numbers kind of adjusted to what China's really spending, it would be a bit more similar? Or is the US still way more advanced in terms of spending?

Robert Gilhooly

I think there probably still is a gap, but the gap probably isn't as big as some of these headline numbers that are being thrown around. So, US tech companies, we think, have spent around about £865 billion since 2022 on CapEx, and they plan to spend more than £600 billion in 2026 alone, which is, that's around about 2% of US GDP. In contrast, if we add up some of the kind of major Chinese tech firms, these are still pretty big firms.  This is probably what I mean - they've kind of only spent around about £70 billion and their kind of plans to expand, while still rapid are falling well short of those kind of US tech companies. But I think this does understate China's broader spend. The build out of data center infrastructures is very clearly driven by the hyperscalers in the US. But in China's case, more of the capital-intensive investment sits with a kind of broader, more diffuse array of firms. This includes some state-backed telecom companies rather than its kind of high-profile tech behemoths. And China's spend and state support, I think, overall is a bit more hidden than that in the US. On top of kind of direct investments, we've got grants, subsidies from central and local governments. The People's Bank of China is a re-lending programme dedicated to boosting the flow of credit towards sectors deemed of strategic importance, of which, I think AI is certainly part of that. And consequently, we think Chinese firms are likely to receive these kinds of larger indirect financial subsidies than those in the US. So, I guess I'm confident that this is narrowing this gap from the headline numbers, but it's impossible to know exactly by how much. So, I think it's kind of fair to say that even kind of despite these kinds of Chinese characteristics of the system, the US probably is still winning the spending race. It's just not winning it as much as those headline numbers would suggest.

Nick Robinson

Right, okay. And I suppose thinking about the broad importance of government strategy in terms of development of AI, I mean, how would you characterise the main differences between the US government strategy and what China's been doing and whether or not that's likely to have a big impact in who ultimately develops AI faster and wins this so-called race?

Robert Gilhooly

Yeah, to some extent, I think the US strategy is to kind of get the government to stand aside and let the private sector get on with it by minimizing that kind of burden of regulations. If we look at the Trump administration's AI action plan, which came out last year, at least talks a pretty good game. It highlights the needs to cut regulatory hurdles, of which, to be honest, I don't think there were that many to begin with, but also to build out energy and computing infrastructure, try to help export American AI globally, and then kind of also integrate it, I guess, across the government, so defence, federal agencies. So, in essence, I guess, kind of boosting demand or opening up, you know, the public purse. On that side. But the US strategy is definitely not all free markets and getting out of the way. Export controls on the most advanced Nvidia chips and more generally, semiconductor manufacturing technology are definitely key levers the government is pulling to try to give US firms the edge over Chinese firms. I do think export controls are likely to continue to give US firms a fairly strong advantage in training AI models that will be kind of hard for China effectively to catch up with. So that's definitely kind of a key difference in some of the strategies.

Nick Robinson

Yeah, I suppose you mentioned kind of regulation as an issue within the US. And obviously, power is a massive issue in terms of electricity. And we all know that data centres are enormous consumers of electricity. How close do you think the US is to hitting a power bottleneck?

Robert Gilhooly

Yeah, look, I mean, we've definitely seen some political blowback here already. There's been local legislations, in some cases proposals to either regulate or even kind of cancel data centre projects. And President Trump has taken to Truth Social. And you've got to imagine your own cap locks here, but he doesn't want Americans to pay for higher electricity bills because of data centres. And he said, big tech companies must kind of pay their own way on this front. And at least for now, the hyperscalers seem relatively happy to oblige. Microsoft's committed to compensate for higher power prices. And contribute to local communities, while others are kind of turning to behind the meter generation to try to avoid the impact on the grid and households altogether. But attempts to shield consumers plus the Trump administration's aversion to renewables, I do think this could be problematic. The IEA (International Energy Agency) estimates the US power consumption is going to rise to around about 2% growth per annum by 2030, and this following on from effectively two decades of fairly minimal power growth in the US. And the bulk of that, acceleration is due to data centres.

Nick Robinson

Yeah, and I suppose with data centres, and you kind of mentioned this, there is some, I suppose, public opinion that's turning against data centres and a bit of backlash against AI more broadly in the US. Do you think that's a political risk for AI expansion?

Robert Gilhooly

I think the cost element maybe is, but I think probably ultimately not. There are obviously some concerns of kind of other dimensions as well around kind of labour market impacts, but I think we're just too much in the early stage for this to really become a kind of political constraint in its own right.

Nick Robinson

Yeah, and then I suppose thinking about China, I mean, obviously quite different there in terms of them, well, the Chinese having added grid capacity equivalent to the entire US grid in just a few years. So, with China's more, I suppose, top-down mechanism of government, how much of an advantage is that in terms of just being able to get more power built?

Robert Gilhooly

Yeah, if we're just kind of scoring different elements or aspects of this race, I definitely put this kind of in the camp of here, we've got a risk of US bottleneck developing or at least generating an additional cost via the power sector. Whereas it seems like there's kind of, I'd say, little to no chance that the power grid actually becomes a bottleneck in China. China's already had the edge in terms of electricity costs. If we look at, say, cost in per kilowatt hours, And I think this cost advantage is also going to get compounded on a couple of fronts. One is just that China's pushing us so far ahead in terms of its renewable power drive. And actually, you know, we had a record 420 gigawatts of capacity added in last year alone. And I think the kind of longer lasting impact from the Iran conflict for China is going to be a doubling down of their desire to kind of cut their reliance on foreign energy centres. So I think that kind of large structural move within China is going to help to drive down electricity costs quite a lot further. And you can also say, you know, China's also maybe that cost advantage is also amplified. By China's kind of more cost efficient open source models too. So in that sense, there is some advantage here in this kind of state directed system, I think for China, especially on that kind of power front anyway.

Nick Robinson

Yeah, that does seem very clear. So particularly while these models are huge consumers of electricity and compute, that's going to be something the US is always going to struggle with, I think. But I suppose one area where the US definitely does have an advantage is just the quality or the ability of advanced semiconductors. I mean, how limiting is that for China in practice to be quite far behind on that?

Robert Gilhooly

Yeah, I mean, it's interesting as well because this is a limitation that's still present despite very substantial subsidies and government encouragement to try to catch up with the more advanced semiconductors. China's really struggled to close that gap. I think we've been saying for probably the best part of a decade now that China's sort of five to seven years behind the most advanced semiconductors and that gap hasn't really closed very much. So, I think that would suggest that even if the government redoubles down on its efforts to develop domestic chips, this gap is going to persist for at least a few more years. That said, there's a risk of maybe making too much of the access to advanced semiconductors. It's not been a particularly firm barrier to date. In parts, China, you know, has obtained some advanced chips, Deepseek was reportedly trained on some of the more advanced Nvidia chips. And also, at least, for now, there does seem to be a route to get access to data centers with more advanced semiconductors located elsewhere in the world, too. And, you know, US is still allowing China to buy some Nvidia chips, maybe not the best ones, but the H200 models are still able to be sold. I personally wouldn't rule out some loosening in some of these restrictions after Trump visits Beijing either, but I still do think that's probably a world in which Chinese access to the most cutting edge isn't that likely either. But you know, the limitations here, I guess they also partly depend on what you want to do. If you want to be training the most advanced models, you do need those advanced chips or a solid workaround. But if what you want to do is inference, i.e. Running the models that are kind of already developed, you don't need the most advanced chips for that. So, some chance there, this kind of reinforces the Chinese focus on applications rather than trying to achieve artificial general intelligence like some of the major US firms seem to be doing.

Nick Robinson

Okay. So, as we see, perhaps a switch to more inferencing from training, then that relative advantage of the US perhaps declines a bit.

Robert Gilhooly

Yeah, I think that's right.

Nick Robinson

And I suppose, you know, Huawei does have a slightly different strategy in terms of how they're generating processing capacity in terms of this swarms beat the Titan approach, you know, where they stitch together millions of lower capacity chips to try and achieve the same outcome. Is that working? Is that a strategy that's viable?

Robert Gilhooly

Well, you know what they say, necessity is the mother of invention and that is a kind of example of illustrative or, interesting workaround technique. I don't know. I think it is difficult to know how viable that is over kind of the long run. So, I mean, I thought it was interesting. It was reported in the Wall Street Journal that when Deepseek was actually being developed, they tried to use less advanced chips from Huawei and also kind of other domestic vendors. But the results didn't really work as well as they wanted it to. So, they ended up turning back to Nvidia chips for some training. So, I think, yeah, it's kind of, I think the jury's out a bit on that question.

Nick Robinson

Okay. And I suppose one thing you've talked about in the paper you wrote is just this idea that actually China is running a slightly different race to the US in terms of their focus on applying AI tools to the manufacturing base. Could you talk us through that a little bit and how you see that?

Robert Gilhooly

Yeah, look, I think some of this kind of depends a bit on how you think AI is kind of going to change the economic structure and how you can kind of apply it to your current economic structure. So obviously China's got a much larger manufacturing base, you know, that's roughly 25% of the economy versus a bit under 10%. In the US. So, in that sense, it's perhaps unsurprising that China's AI plus strategy is really kind of pointing first and foremost at trying to boost productivity within manufacturing sector, trying to streamline process, accelerate product development timelines. And China, I think, could benefit more from what's called world models, which use, large reams of data to allow AI to try to understand the physical world better. Typical examples there would be like autonomous driving and robotics. And China's thought to have around about two, 2 million industrial robots in kind of 2024. So that just gives you a kind of flavor, I guess, of kind of the potential for China to kind of turbocharge productivity to some extent within manufacturing. And I think these kind of world models to some extent are also getting a bit of support from a different regulatory approach in China, as we were kind of saying earlier on, you know, the US approach to some extent take a step back. But if we go back to the 2021 reform agenda that came out of the two sessions then, China has to some extent been, I think, at the forefront of efforts to harness the power of data, effectively acknowledging that data is just another factor of production for a modern economy. So, they did change a couple of their kind of data laws, in particular the personal information protection law, might still require user consent for users' data to be shared. I still think it's probably the case that China's a bit less encumbered by data privacy concerns in the US. Indeed, we've actually had a couple of recent efforts that have included creating public data sets and kind of marketplaces for firms to trade data with government bodies. So, I think that kind of potential kind of freeing up of data within the Chinese economy, if it's done well, that could give them an advantage in helping kind of train models, tailor them to tasks, and speed up the adoption across manufacturing as well.

Nick Robinson

Yeah, okay, so that's interesting. So would you say, I suppose, that when we think about capital allocation that China's doing versus the US at the moment, perhaps China's actually kind of oddly doing a better job of capital allocation this time round, whereas perhaps the US more risk there, given that they seem to be throwing an awful lot of capital in the quest to get to AGI.

Robert Gilhooly

Yeah, I think you could think of that in terms of that's a much kind of riskier aim. I suppose the counterargument here is, look, at least there's solid demand for these US closed models. And it's a little bit less clear exactly how you keep pushing out so many open models, given their kind of the ability to monetize those open models is a little bit less clear cut. So, you know exactly how these business models work in China could be kind of like part of a counterargument to that. But yeah, I agree with the general point. You know, you want to try to create the machine god of artificial general intelligence. That's definitely a very risky undertaking.

Nick Robinson

Right. And I suppose I can think broadening out, thinking about the rest of the world rather than just the US and China. With US export controls and I suppose US ability to direct sales of these ships to certain places, we've seen that the US has been able to use AI as a geopolitical lever in some cases, particularly in the Gulf, when you think about what the Saudis are trying to achieve in terms of AI investment and the UAE. How effective is this likely to be, do you think?

Robert Gilhooly

Yeah, I think it could actually be reasonably effective. As you said, it's kind of been, I guess, one part of the US government AI strategy - to use AI as a geopolitical lever and therefore, help to embed the US tech stack across more countries. I do think that compute is becoming a key pillar of the US relationship with the Gulf. There are some interesting examples there where kind of countries in the region actually had to effectively kind of dump Chinese technology in order to get access to frontier AI capabilities from the US. Look, I mean, the Iran conflict, of course, could, you know, maybe lead some countries to reprioritise towards defence spending in the near term. But if we're kind of thinking more medium term, large sovereign wealth funds suggest they've got the capital to deploy. They have abundant energy, which again could be kind of supplanted both by kind of renewables or fossil fuels. And I think that all suggests, you know, Gulf states, when they put their mind to it, can expand those AI capabilities very rapidly. You know, that gives wide swathes of Middle East, North Africa access to US technology potentially. And then, by expanding access to the US tech stack via the Middle East, that potentially also lessens the strain on the US power grid that we were talking about earlier. So, there's a few interesting dimensions, I think, to this kind of gulf strategy, if you will. But I think those kind of all hang together fairly well.

Nick Robinson

Okay, so kind of going back to some of the initial questions, do you think given this very different way that the US and China are approaching AI, do you think it's right to call this a race at all? Or do you think, you know, actually what's going on here is something slightly different?

Robert Gilhooly

Yeah, I think there is a risk here that we're kind of trying to force this into a kind of a race analogy. There are definitely some aspects where you could argue that maybe the race is is diverging. So, as we're kind of saying, maybe US firms are kind of attempting to kind of shoot for the moon for artificial general intelligence versus China. So, AI plus strategy, you know, more focusing on kind of current tools to augment its fast-manufacturing ecosystem. So, you know, that maybe could suggest that, you know, if we're going to stick with the race analogy for a moment, the runners are somewhat diverging. But, you know, equally, This is definitely a marathon, not a sprint, so I wouldn't rule them kind of joining back up to some extent, further, further along the road. In terms of this, I guess, kind of misconception, you know, perhaps it comes back to what we discussed at the start. If we were to look back, say, Nick, on the internet race of the late 90s or early 2000s, would we really say that was a race? Was there really a clear winner? It's not really very clear that there is a finish line when it comes to kind of the broader changes in digitalization that effectively AI is just another phase in. And I guess, AI will change the structure of economic activity in both countries, just as kind of the digitalization dynamics from the internet and computer revolutions did before them. So, it's not clear, I guess, to me how lasting any leads really, really were in that case.

Nick Robinson

Okay, great. Well, thanks, Bob. That feels like a good place to draw the podcast to a close. So, thanks a lot for joining today, Bob.

Robert Gilhooly

My pleasure.

Nick Robinson

Great. And thanks to everyone who took the time today to listen in. If you enjoyed today, then please download our other podcasts from our website or wherever you normally get your podcasts. Watch out for the next episode and tune in.

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 Aberdeen. The value of investments and the income from them can go down as well as up, and investors get back less than the amount invested. Past performance is not a guide to future returns, return projections or estimates and provide no guarantee of future results