AI Political Economy, Geopolitical Role of Middle Powers

I'm very happy to be here. Thanks for having me. I think it's a very exciting time to be thinking about Australia, of all countries, in particular in the middle-powers context, with as much going on in both compute and safety. So I'm particularly excited to get to give this talk today. I'm Anton, I work at Carnegie. I work mostly on the role of AI middle powers, and I'm trying to find out to what extent that is a coherent category of AI policy to make, and a coherent way to think about the countries in the world that aren't the US or China.

For today's talk, I'll just try to give a little bit of context about the shared challenges, the shared potential solution space, when it comes to geopolitics and middle powers, and then what some of that might imply for single middle powers that are looking to chart that course. Maybe we'll start with a very general view of the landscape. The general landscape we have is that, just like everyone else in the world, the 193 countries that aren't building very strong AI systems just have a general problem, which is: the US and China are building AI systems and they're taxing them.

They're regulating them. They have their space of challenges to deal with and their space of opportunities to deal with. They're building very powerful systems. They have a lot of control over what these systems do. They have the ability to extract resources from these systems to solve some of the problems caused by these systems. And I think much of the rest of the world doesn't, in the same way, have that kind of influence. The question then is what do we do with the rest of the world?

And I think very realistically, there just isn't a lot of basis for a lot of these countries to get a lot of say on what happens in this world. Both the US and China have a lot of interest, from national security and economic reasons, to basically chart their course on AI policy alone. The rest of the world, largely, neither has the awareness nor does it have the hard or soft power to really influence what is happening with these systems. So I think, as a consequence of this,

being the realm of international AI policy where most policy is just being set by the US and China. I think one of the most obvious and maybe intuitive first steps is, well: if we're going to get anything with international policy, whether that's some kind of treaty that we might want to happen, whether that's just getting the individuals paid to single countries, it does just have to start with developing a position of strength and national sovereignty in single countries. I think, reasoning from that, we can think a little bit about what is even the category of countries that is most worth thinking about.

By first approximation, I think we can talk about the middle powers here, which are advanced economies that have some stake in AI. I think maybe one way to think about this is the G20, and then you take the US and China away, and then you add some other countries. You add countries that are particularly security-relevant to the US, you add countries that have a sufficiently strong position in the supply chain. You might, for example, add the UAE because of data centres and so on. But you have a narrow band of countries that have some assets that they can potentially use, that have a strong position in AI.

They're not just completely susceptible to whatever is happening, the whim of the great powers, but they're also not in a default position of strength. What they have is quite a lot of sort of policy contingency. If they play their cards right, and if they use their assets well, they can have a lot of influence and power, and maybe they can secure their position and maybe we can get some things right. But if they don't, then the default outcome is that things go pretty badly — which in combination means that there's some action space.

So that's an interesting area to actually make policy. I think if you look at these countries, that's maybe the most coherent category, because it is not a category that is defined by geography or by the specific assets they have right now, but by the ability to make policy choices correct there, then we can actually meaningfully change the trajectory of the world. And I think, in some sense, Australia has maybe the most interesting and exciting example days, just because it turns out a fairly idiosyncratic and — on the broader world stage — pretty moderate specific strength, or just the ability to potentially build out a lot of this compute fairly quickly in this current context, which is: building out compute matters a lot, can be turned into this position of very strong

short-term leverage. And I think a lot of middle powers have puzzles that are exciting and interesting in similar ways. And I think their shared challenge is, well, let's find out what that is, let's enable going for these plays, and let's look at the structural features that these plays might have, so that we can pull them off. Continuing from that, I work in Washington DC. And I think this case, part of this middle-power argument, is sometimes understood to be a very adversarial thing. If you talk about technological sovereignty, and if you look to Europe and the conversations that are happening about weaning

off the US tech stack and building their own competitors and competitive models, and just finding some way to gain more leverage or have more control of their own destiny in the AI and broader tech-policy sense, this is often understood as a broadly adversarial thing, to think about the middle-power strategy thing. I think especially for countries that have strong security cooperation with the US, and therefore a strong starting point for a fairly healthy bilateral engagement, that doesn't need to necessarily be the case. I think there is a way to redefine the question of middle powers and geopolitics

as one of complementary usefulness and complementary leverage of the US. And so I think it's just worth keeping in mind the more constructive and more cooperative version of framing this entire argument in the bilateral sense. For example, in this instance, the Australia-US context. A lot of these middle powers are major US potential treaty allies with important capacities that the US could either get access to or be shut out. And insofar as that's the case, there is a mutual incentive for these allied countries to use these assets in a way that retains and secures

access to all the things that the US is doing while building its models, extracting resources from them, and so on. And on the other hand, allowing the US access to these important assets in countries elsewhere in the world. There's a bunch of different specific aspects of this. For example, there's a bunch of talk about the AI supply chain, if you look to the European AI policy: ASML and Zeiss. And even if you look to East Asia, you have HBM manufacturing, Tokyo Electron. There are a lot of things that feed into the supply chain, where the Americans are building the chips and building the AI systems

that are just trying to be important to the US to keep around. Lots of that is in middle powers. So there's a lot of leverage from other powers, and also a lot of opportunity for the US to get mutually beneficial alliances going so they secure access to these systems. A similar concept that I think on both sides of the aisle is fairly popular in America is this idea of allied scale and broader alliance AI capabilities. The classic canonical version of this is to talk about manufacturing specifically.

The Americans built these AI models, but then it turns out, to turn a country of geniuses in a data centre into hypersonic missiles or vacuum cleaner robots — whatever you want to do with them, any sort of real-world value that you might want to extract — you actually need some ability to build things in the real world. That can be something that middle powers can contribute and get some mutual leverage in. I think another version of the allied scale is there are only so many places in the US to build data centres.

There is not that much political appetite for building a lot of the data centres. But it turns out, to turn good AI models into a widely available consumer product, or a sort of broader infrastructural good that just leads to good economic outcomes, you just need a bunch of inference. You just need gigawatts or gigawatts of compute any way you can get it. And so it turns out one version of allied scale play is that you don't actually need to build the products. You just need the allies to build the data centres that are used for inference the products.

And then that's maybe one way to generate some bilateral leverage. And then I think what you get from that is basically a situation where you can either offer the US a pretty good deal, which is mutual access to the things we do well and access to the things that you do well. Or alternatively, realistically, a lot of these allies are going to panic, they're going to get very nervous, they're going to potentially reorient towards geopolitical rivals, like China, to deal with this very volatile situation of being exposed to all these risks.

And eventually that might obviously be the second best, worst equilibrium for the US. So in that context, I think the deal to the US should be pretty clear and pretty obvious and mutually beneficial in a very simple sense. You can just distribute a lot of capabilities across the alliance. That is good for everyone. And I think making that pitch usually goes over a lot better and makes it much more likely to do good middle-powers policy than to frame this primarily as a sort of adversarial "let's move away from the US" kind of measure.

Having said all that, I think the way to think about this is unfortunately mostly: what could go wrong. And I think what could go wrong fairly specifically is that the middle powers capture all the risks — all the things that could go wrong with AI — and mitigate the benefits — all the things that could go well with AI. I think three broad challenges play into that. The first thing, just getting access to frontier AI is not a given. I think when I wrote this slide for the first time,

which was something like six months ago, people were a little bit more sceptical of that broader case. I think we've now seen a lot of evidence for why access to frontier AI might very clearly be restricted, rightly so. The export controls, not really export controls, but still sort of directly export controls on the table. You've got a US government that's more and more interested in the security implications of making these frontier models widely available. You've got more and more crunches in the inference supply that makes it so there's not enough AI to go around for everyone.

So there is an open question: do you even get access to these systems to reap all the economic rewards? I think, well, that's unclear, but what is clear is that you're going to be exposed to all the risks and all the disruption. No matter where the workforces get automated and build much more efficient products or much better services, they will eventually take the domestic labour market of any exporting country, because there is international competition. And wherever abilities to misuse AI systems — whether that's cyber security stuff or biosecurity stuff — wherever these capabilities emerge, they will eventually cross borders

and reach the middle powers as well, just with the difference that the middle powers won't have access to the frontier tools to build up the defences against that and start dealing with these systems. So by default, you don't necessarily get all the defensive capacity, you don't necessarily get all the potential upside. But what you do get is a quadrant of the risks and the disruption. And then the question is how do you deal with that? And I think the realistic scenario right now is you don't really, because you don't have substantial political leverage over what the models do.

If you get this and you're very concerned that the bio thing is going a little bit too fast, and it looks like there are serious risks associated with these models, you can't start regulating. You can't just stop these models from coming up. It's the United States of America that is in their policy toolkit. If you're a middle power that is incidentally susceptible to that kind of thing, you don't have the same toolkit to deal with that. You just have to accept whatever the Americans choose to regulate or not to regulate.

And then you have to deal with that downstream. The same thing goes for dealing with social repercussions of AI models. And I think that's maybe the more underrated version of this, which is: usually we think of the US, especially in tech policy, as a sort of laboratory for policy for the rest of the world. They get the tech first, and then they get to come up with some thoughts on how they deal with the tech. Maybe they come up with some good regulatory and taxation approaches.

And then once this tech has diffused throughout the world, other countries can also copy some of these approaches and then find their own solutions to the disruptions — for example, the labour market effects that we think about that AI might have. Specifically with AI, that's not really the case, because the US has the privileged special case of just being able to tax the AI models. A bunch of the compute is in the US, the valuable firms that design the chips and that produce the models are in the US and they're listed in the US, and the labs specifically are just incentivised to give stakes and give shares and give taxation access to the US government in ways that they're just not incentivised to cut in anyone else in the world.

So if it turns out that the AI thing just requires us to pass very expansive social safety nets and policy that is, and then everything turns out fine — as long as we can do that, then we might just end up in a world where the US does just fine because it has enough tax revenue to do that, while the rest of the world could adopt these measures if it only had the tax revenue, but because it is all US-centric, there is no way to get around it.

And I think that is the general shape of the challenge that I think most of the middle powers here. And I think that also gives rise to the general set of solutions that middle powers should probably pursue first. Well, if we need access to these models and we can't build them ourselves — and I'll just take this for granted for the purpose of this talk, I think separate conversation to what extent that's true — but I think it's looking like it's pretty hard to build very good AI models if you're not worth $1 trillion.

And not a lot of countries outside the US are worth $1 trillion. So the first obvious thing you can do is you negotiate access guarantees. If you find something that the US really wants, and if you use that as a chip for trading, for guaranteed access to US frontier models, you're probably not going to get all the way there, in the sense that the NSA is always going to have better access than a random consumer in another country in the world. And there are going to continue to be these limited access, Project Glasswing style programs where

very good AI systems aren't given to everyone. But what we can ask for is some extent of parity. You can ask for most of your businesses to have a similar amount of access as American businesses do. You can ask for your consumers to have similar amounts of access as American consumers do, and I think that already gets you plenty of the way there. And I think there are ways to use leverage for that. In the Australian case, the most obvious way to do that is just to build up the compute.

If you are x percent of the compute budget of a major lab, it turns out that's a really relevant stake. That's a really relevant factor that really incentivises both the lab and the US government to continue giving you access. If your threat is that it's baked into the contract, that if they don't give you access, you cut off the compute, you rent out the compute to someone else. So I think the most obvious low-hanging fruit for any country that can build out compute is to just use compute as the anchor for this access thing.

I'll just very briefly go over this, because I know there's a lot of other good talks, and the specific safety interventions that are worth doing in both the middle powers generally and in Australia specifically. The obvious solution to the risk exposure is just to build up resilience. The US has maybe the most efficient solution to dealing with the risks, which is just trying to get ahead of the risks. Everyone else can't do that. So everyone else is even more incentivised to find ways to be more resilient to the risks and to build out defensive capacities, whether that's UVC lighting for getting rid of some of the pathogens, whether that's wastewater surveillance and genome screening, all these things that we might be able to do that are downstream of pathogens being introduced

by misused AI, whether that's just building out cyber security resilience, whether it be building out broader societal resilience to the labour market stuff. I think many middle powers are both necessarily incentivised and also therefore in a position to do some real good for the world to build out these resilient capacities. The US can focus on what can we do about the models themselves and the systems themselves. And I think the rest of the world is more incentivised to figure out, well, given that there are going to be dangerous models, what are ways to make the world — to make our countries, and then also the rest of the world —

safer to the effects that they might have? I think that is the basic logic that will also lead to positive spillovers from getting the middle powers thing right, because generally these resilient solutions, they do generalise very well. If we come up with good technological innovations to deal with the vectors of AI risks, we can export them through our broad alliance. We can just pass them throughout the world, and then we get to a broader level of resilience of the world just downstream of a single middle power starting to realise: "Wait, this is something we should really be thinking about."

And then the last thing about political leverage is, well, we just have to find something that's going to be valuable so that we can still have economic activity in our country that allows us to tax it, allows us to get the fiscal revenue. It's kind of hard to say what that will be. In the labour economics context, a lot of thinkers usually use the framework of, well, what will be scarce if there's a lot of abundant intelligence floating around, what are the things that are still going to be very valuable?

Again, compute is the most obvious solution, because it just is very easy to compare. It's very easy to build out. It makes a lot of sense, but you can go in with a broad range of things that might be valuable. Whether that’s human preference jobs like health care or artisanal goods or that kind of thing? Whether it's tourism, whether it's — a whole range of literature on that question. But I think the common trend is just, well, if you have something that gets more valuable as AI becomes more and more abundant, then you probably do have a decent position, at least to have some sort of revenue and tax base to deal with some problems.

If you don't, and if most of that is quantified, and if you're a country that has just a very high share of services economy – and that's mostly what you do – I think that leads to — that makes things a lot more difficult. Going on from there, what do you actually do about all this? I think I've sort of teased and mentioned this throughout the talk already, but I think the most obvious thing is compute just makes a lot of sense to buy in the short term.

I think there are some wrinkles to that, because compute is very expensive and compute is not going to matter forever. So at some point the compute constraints are going to ease a little bit. We're going to come up with new solutions to get compute online in more parts of the world, we're maybe even going to come up with solutions to get compute into space and build it out there. But in the near term, which is to say in roughly the next 3 to 4 years, there's just not remotely enough compute to go around.

Both the labs and also the US government — because it cares about the valuation of the labs and cares about the political input of most of the labs, maybe not as much Anthropic, but definitely all the others — is incentivised to go along with what the labs want as well. So if you have compute to offer in the current situation, you can extract a lot of concessions from that. You extract tax revenue that results from the models being serviced and maybe even developing your country. You extract direct leverage to the US, because you can, in the limit, just turn off the compute and be very inconvenient to the US.

You gain indirect leverage over the US, because you can inconvenience a developer that you've given the compute, and then therefore extract concessions on stuff like frontier access. And I think what you also get is the sort of OPEC-style mechanism, where it turns out if you just have a decent percentage of the global compute supply, you just get influence over the broader global market. Because if compute is just a somewhat financialised good that is priced somewhat like a commodity on the global market, then being able to restrain some compute production and maybe turning turning off a data centre really sort of does things to the global compute price, in the same way that OPEC countries can reduce or increase oil manufacturing.

So you get this broader set of geopolitical leverage along with that, that doesn't even need any specific bilateral engagement, but just works through pricing power. The downside of that is, of course, you have to be really good at building compute to do this, because the only way that the labs and the US government are going to go for this kind of deal, that hands over so much leverage to a country that builds that, is if that country is just so much better at building out the compute than all the alternatives.

And I think, realistically, in the current scenario, that means doing it very quickly. I think that is the core bottleneck to getting the compute play right. The costs on energy, on construction, on basically all the other things are basically marginal compared to just the effect that it has that you get this huge capital expenditure and chips just online a month or two earlier. Very concretely, I think if you can get something on the order of 1 to 2GW online in a time frame of something like 18 months, you probably have enough leverage to do a lot of the things that we've discussed.

On a similar note — that sort of skip over this a little bit because it's a little bit less relevant to the most obvious case for Australia — you can leverage bottlenecks upstream and downstream. And I think that's going to be a strategy that a lot of other middle powers pursue. If you have a strong position in the semiconductor production manufacturing supply chain you can basically do the same thing as with the compute. You can threaten to withhold, you can offer to expand, you can offer deep mutual integrations that are attractive to the US government and so on.

And I think you can basically run the same playbook, but with more specific bottlenecks for AI that aren't just the compute thing. I think maybe the more interesting part, to perhaps close on, is coordination as an important measure here. I think there are two reasons for why coordination in some way is under-priced and important for middle powers. Of course, there's like big pay offs. Oh, well, wouldn't it be great if all the middle powers work together, they all pool their leverage, they all work together, and then they're much stronger than even the US.

And they get all the concessions they want. I think realistically, that kind of thing just doesn't work. The US starts offering good bilateral deals to some of the countries, and they start defecting and they start going with the bilateral again. And they're the US's best friend again. And then at the end of the conversation you have a very small alliance of very rebellious middle powers, but all their friends have sort of been bought off or leveraged out by the US. And it probably doesn't work out that way just because of some extent of political awareness and interest in that kind of thing we have.

But way short of that, there are still things we can do around coordination. The most obvious to me just seems to be coordinating around division of labour. Right now, if you're a middle power you probably are very nervous about a lot of things. You're nervous about your access to compute, you're nervous about your access to frontier models. You're nervous about whether you have these bottlenecks. You're nervous whether you get to have this taxation capacity. You're nervous whether you have the tech ability to build up resilience and have the safety infrastructure.

And it turns out that you don't need all of this in your own country. You just need all of this in some country that you trust. I think the specific Australian implication is you can just think a lot bigger about the compute stuff specifically. Because a lot of the alliance – broad set of middle powers – is interested in having a place that runs a lot of compute. They'd be very happy if they got to spend their very scarce gigawatts and political energy on doing something else and then building out compute.

And if there are deals, advance market commitments, offtake guarantees, that kind of thing to be had around compute, then I think in turn, Australia can expect a lot of concessions for capabilities it itself does not have from other countries around the world that are seeking a location for a grid and everything that could sustain the level of compute build that they can do at home. Luckily, we're in a paradigm where the latency stuff doesn't seem to matter as much, for instance. So it turns out there is actually a coordination play to be had.

And for a country that has a very spiky, distributed strength of the potential ability to build out compute, that means, well, maybe you don't have to do all the other stuff as well. Maybe you can just use that in sort of bilateral engagements with the middle powers that are also interested in this kind of stuff, to extract the rest from the rest of the middle powers and have the sort of mutual integration. That seems like one obvious way in which this coordination between the middle powers, perhaps.

And then the second part of coordination that seems to really matter is coordinating with the US. Because going back to the frontier access question, when they're in the room and they're thinking about “Well, are we export controlling the next model? Which firms, which governments should get access to this model?” there's basically two factors. The first factor is, well, does it help us in any way if we give them access to this model? And I think that's what most of this talk has been about. It helps you

to give them models for allies, because then you get access to their capacity and they have some leverage against you. So if you don't give it to them, things actively look worse for you, because they stop exporting their lithography machines, or they stop giving you access to that compute. That's part of the reasoning that you attack with most of the things we've developed in this talk. The other part is how risky is it for the US to export a model and to give other countries access to the model?

And I think that basically just comes down to security integration, and that just makes it very, very convenient to be a member of Five Eyes is I think the main upshot here. Because if the US has some baseline level of trust that your government and also your infrastructure isn't easily breached and penetrated by foreign agents and adversaries, it just makes it much, much, much, much lower cost to export these models to you. I think just for that reason alone, it's just a very convenient position to have a strong US bilateral.

Not so much for the day-to-day politics, because those are always going to be volatile in the Trump administration. But for the fact that the intelligence community just tends to trust its longstanding deep security embedded allies a lot more than the rest of the world, which puts these allies in a much better position to get compute, to get that frontier model access export to without security concerns arising. And now I’ll close on the side that is deflating and a pessimistic note, which is, well, all of this is nice, but that's only the start.

That's how you get to cut into participation and potentially having some sort of access to these models, and how you don't get completely disrupted and pushed out by all the risks that might manifest. But downstream of that, we will still need to get the basics right to use these systems well and redistribute their effects well. There is still just this baseline impulse of, well, you just need these good bilateral relationships. You need a decent amount of economic growth. You need to be able to distribute that in helpful ways.

And so I think there are some clever things we can do on AI policy to sort of pull the rug sideways. And most of this talk has been about them, but also, as AI gets bigger and bigger, much more of AI policy is just going to be general security policy, general foreign policy, general economic policy. And so the kind of depressing part of this is, well, if you don't get these fundamental basics right, the rest probably also doesn't matter all that much. But still, getting the AI policy stuff right does matter on the margins.

I think, to be honest, when I look at the middle powers of the world and I travel around and talk to a lot of people about them and talk to a lot of companies, I feel like Australia is maybe in a uniquely good position to make a lot of things happen right now. And I suppose — thanks for having me, and I wish the best of luck with it.