The Race Is On To Control AI | Episode 274: Afonso Salema & Sam Winter-Levy

Sam Winter-Levy
Overnight, the White House can essentially call up one of these companies and say, you can't release this model, you can't serve these models to particular countries, to particular players. One aspect of Anthropic’s release strategy for this model was to give advanced access to a small group of hand-picked companies that formed part of what Anthropic’s called Project Glasswing. All of those companies initially were American, so if you're the Europeans, if you're the Australians, if you're states around the world, this puts you in a very vulnerable position. You don't have frontier models of your own, you're dependent on chips created by US dominated supply chains in US data centres that are primarily in the United States. And your access to these systems, the access to the best frontier systems coming down the pipeline, is dependent on essentially the goodwill of a group of US companies and the government that oversees them.

Michael Liebreich
Hello, I'm Michael Liebreich, and this is Cleaning Up. Even the briefest scan of the news will show that AI and security are two of the biggest topics of our times. But what happens when the two collide? Whether it's physical or cyber security, keeping our rapidly expanding AI infrastructure safe is a huge and hugely important topic. To tackle it, I've asked James Cameron, Chair of Cleaning Up's Editorial Council, to take the hot seat and host this episode. James is a barrister, social entrepreneur and independent adviser on all things climate, finance and law related. If you want to know more about his background, we'll put links in the show notes to his two appearances on Cleaning Up in his own right. And so it's with great pleasure that I welcome James Cameron to host today's episode of Cleaning Up.

James Cameron
Welcome to Cleaning Up, and today we have a three way conversation. And in that conversation, we're going to address three topics which are now completely intertwined: data centres, security and geopolitics. And it's a live, current, right in your face issue today, not least because on the first of March, there was an attack on Amazon Web Services installation, two of them actually, in UAE and Bahrain by Iran. Probably the first time a data centre has been directly targeted by a state in a military conflict, and so we need to think about what that means. Are they becoming, are they forming a list of strategic targets? But equally, there's a race on. It's both a commercial and a sovereign race to be masters of this phenomenally powerful technology, and it's producing huge demands on energy and water and concentrating the minds of pretty much every type of decision maker all around the world. And we are very lucky to have you two to talk about this, we have an entrepreneurial perspective, we have a public policy and academic perspective. And now in the tradition of Cleaning Up, I'm going to ask you two to introduce yourselves in your own way and your best way, and I'll start with Afonso.

Afonso Salema
Yes, of course. So thank you so much for having us. It's a pleasure to be here on Cleaning Up. So very shortly, I work on the physical foundations of the AI economy. I lead Exus Digital, a data centre developer in Europe. I also founded TerraByte Data centres, and as AI is becoming industrial scale, the challenge is not only to build buildings full of servers, it's getting reliable, affordable grid, cooling networks, and that's what I do. In my very little spare time, I'm also a keen AI enthusiast, which means I'm developing, I work with several models, testing harnesses, building my own harness in my own software with AI models to get a sense not only on the software and then that side.

JC
Sam.

SWL
Great, thank you so much for having me. So I'm Sam Winter-Levy, I'm a senior fellow at the Carnegie Endowment for International Peace in DC, a non-partisan think tank, working on the geopolitics of AI, with projects looking at risks to data centres, what determines where data centres are built around the world, working with the US government and other allied governments to try and understand the geopolitics of this race that's unfolding before us.

JC
So I'm going to start with a proposition to you Sam to start with, but obviously Afonso you have views, which is that compute has become a strategic resource, one that countries compete over, one that companies compete over, and that has consequences for all sorts of decision-making across the world. Is that the right framing? How would you phrase it yourself, is that about the right way to look at where we are?

SWL
Yeah, I think that's absolutely the right framing. I think compute is emerging not just as a kind of economic input, but as a core strategic input that will shape how much power different countries have on the world stage. So technological revolutions always elevate new resources to strategic importance. You could think about the internal combustion engine and oil, the railroads and steel, and I think compute is the kind of core physical substrate of this new AI revolution that's unfolding around us today. So I think the countries that are able to secure access to large quantities of compute, whether that's within their own territories or just kind of secure supplies of that compute, I think will have a significant advantage over countries that cannot. And similarly to the companies too, we see right now every major, every frontier AI company is desperate for compute.

If you ask any of the major AI labs what they would love to have more of, they will say compute. And so you can see this playing out in multiple domains. You can see this playing out in the race between the AI labs. You can see this playing out between the US and China, where US export controls on advanced AI chips, on compute have become one of the kind of key flashpoints in the US-China relationship. And you can see this playing out the middle powers and the rest of the world where Europeans, the huge gulf that has opened up between the United States and Europe, the United States and the rest of the world in terms of access to compute could be a kind of major impediment to these other countries' ability to achieve their ambitions and aspirations when it comes to AI. Because I think it's quite easy to underestimate just how big the gap is right now between the US and the rest of the world when it comes to access to compute.

JC
Sam, we might come back to that possible trade between compute and access to compute and partnership over access to compute in a moment. But one of the things I know you are very focused on is the temporal dimension, is the speed to power, the speed to compute, the race, literally the race, to get those resources in place. We'll come back to siting issues in a moment, but perhaps you could comment on that because it's a phenomenally important dimension to the issue, isn't it?

AS
It is and I just wanted to go one step back on what Sam just said, because you're absolutely right, compute is the scarce resource right now. But probably most of the listeners to this podcast are still thinking compute as the laptop or the desktop that they had in front of them. And what changed, it's not that desktop, it's the ability to use different types of compute for different tasks. Something that was limited to special labs and universities with supercomputers. Right now, it's in the hands of companies, of governments, of individuals to use that kind of super compute for additional tasks. And that's thanks to AI that enables on how we use that compute. Having said that, yes, the race is on, the race is on and there's a desperation to gain more and more compute. I was looking for statistics that would allow people to understand how far and how fast this is going. I think there's not a single source of truth to try to understand how many tokens are being consumed right now and how much that actually turns into power. So I'll give you some back of the envelope if you want offline, I can tell you how those were calculated. But if you look at School(AI) for example, when they start talking about how many tokens were consumed and how many were going through the system, and in May 2026 they were telling us that they were processing 9.7 trillion tokens per month. That number has grown to 3.2 quadrillion tokens in May this year, the first year was 2024, not 2026. So in two years, it has grown 330 times. Given that most people don't understand how tokens translate into power, a rough estimate would be in 2024 per month, just the AI portion of Google compute accounted for 16 megawatt hours. Last time that we've got this number, it accounts for 5.3 terawatt hours per month. And Google does not control the entire AI working, they probably have only 20% to 25% of the share of the share. So you can imagine how big the entire consumption of AI tokens are right now.

SWL
Just to add one stat there. So Google, just one US company controls around 25% of global compute. So just to give you a sense of 25% of advanced AI computing power is in the hands of one US company. And that's true across the board, so around 75% of all advanced AI computing clusters are within the US, just 5% within Europe. And every publicly reported major AI cluster within Europe contains less computing power than just one Anthropic Amazon data centre in Indiana. So the scale of the investments flowing into this stack, this is the largest peacetime industrial mobilisation we've seen since World War Two, more than a trillion dollars flowing into this. But it's overwhelmingly concentrated in the US for now. We can come back to whether that may be changing in the coming years, but yeah, compute is so important, just the scale of the investments flowing into the US sector specifically, just dwarf every other country.

JC
So should we just dwell a moment there so we get this laid out clearly for those who are tuned in now? It's not just a question of the US versus China say, or the US versus China and Europe. But it's also what's happening in the Gulf and the other areas that are trying to rapidly catch up through investment and their own technology innovation. And so is it possible to kind of divide the world up into sort of blocks of compute? Is that a good framing also for managing or understanding the geopolitics of this technology?

SWL
So on compute right now, I think it's probably a two player race, it's the US and China. Yeah, more than 80% of global compute is in one of those two places. And realistically, only China has the ability to compete with the US when it comes to developing a kind of full stack, AI supply chain, the chips, the inputs to the chips, data centres and so forth. So I think realistically right now it makes sense to think of this as a two horse race. That's not to say that other countries don't have big ambitions or don't matter. As you said, the Gulf has has very big aspirations and they want to be one of the places that attract some of that compute that attracts the US companies to build our massive data centres in the region. Europe obviously, Afonso I'm sure I can speak to this more, Europe obviously is trying to build out more compute as well.

But right now, I think the most helpful way to frame it right now is that this is a two horse race, the US's position is unbelievably dominant. But the US build out is going to run into constraints now, permanent constraints, grid constraints, a domestic political backlash that we're starting to see now, and that gives opportunities for other countries if they want to capture a slice of this build out. We can come back and talk about whether they really do want to capture a slice of it or whether they should want to capture a slice of it. But I think for now, this is more a two player get raised than a broader set of blocks and competing against each other.

AS
Completely agree. I would just add that you have the US which has shown its ability to deploy new data centres and has the chips. Don't forget who has the chips to actually run these. It had China with more limited chips, but an enormous ability to take on new data centres and new compute onwards. And as Sam says, we started seeing some clouds on the US growth on data centre capacity.

JC
So we have a race for chips too. And that's both a physical supply chain and of course, a political supply chain with risk attached to it, and then we have site selection. So maybe Afonso, we should go to you since you've done it and you've been through this process and you also know that there are significant political choices in site selection, but also physical choices, water, energy, they go together, of course. Maybe talk us through how you would deal with site selection now, given your experience in Europe. If someone comes to you for advice, what would you do?

AS
Well, it's not only in Europe, anywhere in the planet, there's some axes that you need to ensure that they exist. I mean, when I started, when I started looking into sites and entering the sector, I found out there's a graveyard of failed projects which failed on one of these axes. So you need to be very aware of what these are. The first one is, everyone knows, is power. And power is twofold. Power is access to power, that's number one, but also is actually energy. So you cannot just simply put one gigawatt data centre anywhere you see on the grid. It just doesn't work like that, it's too big of a beast to be fed. And you need to make sure that the energy is there to supply it and cheaply. Because when you are doing infrastructure, sometimes you lose track on who is the ultimate user of those facilities, and these are people like myself, I was in this situation trying to find compute for my own uses, benchmarking, I can use the US. The US is offering me cheap GPUs, they're impossibly scarce or not even available, or expensive in Europe. So why should I put my compute here? So power is number one.

The second one is, we are building mission critical facilities. That means that these locations need to withstand any event that can happen. So natural hazards, as broad as wildfires, excess heat, excess cold, come to fruition because if you stop for a couple of minutes, you will have an enormous impact on the GDP of the country that you're running. These are very, very, very large facilities. Second, you come to cooling, where your water comment came. And cooling data centres have that reputation, given the extensive use that they had of adiabatic chillers, but that is changing. Using water or not is a matter of design, of design choices in the location that you need to establish. If you are stashing in the Gulf, highly likely you need to use a lot of water. The temperatures there, they demand that you have smart solutions to get the cooling unless you don't mind of being extremely inefficient on the power side. So there's a trade off, the location weather conditions will determine what type of cooling solution and the type of cooling solution will determine whether you consume water or not, or you can avoid it.

JC
Okay. So that immediately makes you think that apart from the sovereign compute strategy that we've talked about and apart from the access issues that we've talked about, there's now increasingly, because of our first proposition, there is now a physical protection element. How does that factor in? And I just want to add one more challenge to you both to figure out is, okay so we now know that a state may well target your private commercial piece of infrastructure, which is also critical infrastructure to the country. Who pays for the defense? The Russians are now telling their hyperscalers who are under attack from Ukraine that they ought to increase their own defense of their own facilities. Who pays for the defense of a large data centre of the types we've been talking about?

AS
Before we tackle that, one of the elements that I haven't said, which is crucial exactly to answer that question, which is the network. You can have a beautiful building, if you don't have multiple paths and resilient paths or fiber optics that connect globally and connect in the main log hole routes, it's just an empty building.

JC
I'm quite curious to know what Sam thinks about the sovereign defense of very large private facilities like this.

SWL
Yeah, so I think this is a question that is still being worked out. I don't think we have a great answer of who should protect these facilities. I mean, to give one concrete example, these data centre facilities and the Gulf, I think building large scale computing clusters in the Gulf was always going to be a risky proposition given the nature of the region, and given that these data centres are fundamentally quite soft targets. These are big buildings, you can see them on Google Maps, they're hard to hide, they're filled with very fragile chips. If you fly a cheap drone into one of the chillers or the transformers, you can as I said, even just a few minutes of downtime can cost millions of dollars and it's quite easy to take them out. And that is exactly what we saw play out during the conflict, we saw the Iranians target multiple AWS data centres, these are just cloud data centres, not the kind of big AI ones in that building, but they also threatened Stargate UAE, the big OpenAI cluster there.

So when it comes to protecting these data centres, there are things you can do. You can use anti-drone devices, you can use more netting to protect parts of it, you can use more kind of reinforced concrete. All of this of course drives up the price, I think a lot of this should be the company's responsibility. I think historically, companies thought a lot about natural disasters, they thought about cyber attacks. They thought much less about physical risks, drone strikes, missile strikes, I think that's starting to change. Now, as the Afonso said, these data centres are going to underpin critical, they're going to become critical infrastructure. And so then there's going to be some balance of responsibility between the state who has certain unique affordances and information and the ability to protect parts of these, but also the companies who are going to be making decisions about where to site these facilities, you know, how to build in resilience and redundancy.

And I think we don't yet have a great doctrine if you will, for how to think about who should pay the cost of these different things. If you're talking about putting air defense systems around a data centre, that is something only the state can do. Many of these air defense systems are systems in very short supply and so any air defense system you need to protect a data centre in the Gulf is one that's not going to do Ukrainians, or to Taiwan, or wherever else it might be needed. So I think one thing I would urge is that if you're going to build data centres in dangerous regions, you should not expect that, you know, the US government is going to suddenly be on the hook to defend them. And so you should be kind of pricing some of those decisions into your decisions about where to put them. But there's always going to be some balance between companies taking care of certain risks, and then some that are just asking the government to defend them. And I don't think this is just a Middle Eastern problem.

JC
No, no, no.

SWL
There are data centres in Europe you could easily see become targets for Russian sabotage.

AS
So if I may, on the physical security of this, when you build a data centre, you have to take into account a lot of attacks. Missile attack was not on the list, at least the ones that I used to look at, but drone attacks yes, we will look into it. The security on a major data centre is very, very strict, with seven layers of security with several anti-tampering devices, so that you cannot get hostage of the fiber. There are many, many layers of security on the physical block. But I don't think anyone was expecting a direct hit as it happened, like you rightly said. Now having said that, I actually can take stock and hence my network concerned, we can say take stock on what happened on the submarine cables. Submarine cables have become vital piece of today's economy, and is used the UK. For example, two warships have been ordered exactly to patrol the waters to make sure we don't have any submarines trying to sabotage those cables. So Sam is absolutely right, I think it's still undetermined who's going to take care of that but there's already a willingness on governments to to protect these kind of facilities, given how critical they are for their economy.

JC
So can we just explore two pathways to figuring out how best to share the cost of that kind of protection, which clearly does involve the sovereign, certainly when it comes to deploying military hardware, although who knows, some of these corporations are so wealthy they can definitely buy their own. When the Amazon Web Services sites were hit, there were consequences for businesses in the government. Now, they were relatively short lived 48 hours, you lost ride sharing on Careem for 48 hours or so, but you lost mobile banking for a couple of quite significant banks for that period of time, definitely an inconvenience. But 48 hours is survivable for a society, it might be harder if it were longer or went deeper into the economy. There are consequences therefore that have to be paid for, there are some compensation issues about those people who have suffered, where does that get picked up? How does the insurance industry cope with the private sector losses associated with that?

And then just as we've had a story unfold about the sovereign race to compute, so there is a sovereign defense issue to protect these critical resources and they have to be paid for. And I don't think our publics are really very aware and may be quite resistant to the idea that these extremely wealthy countries and companies that are putting up these sites require more of our defense resources in order to protect them. How do we manage that? And is there some scope for some international agreement, especially for the middle powers, to share the cost of this protection? Or frankly, at a time when international agreements are extremely hard to contemplate, maybe that is wishful thinking.

SWL
Yes, I think one point to bear in mind here is, the cost depends on what the data centres are being used for. So as you said, the AWS data centres that were hit in the Gulf, these are primarily being used for food delivery apps, for businesses, some government functions. So for sure an inconvenience, but not as you said, not critical for all of society when they went down. As you said, the outages were relatively brief. So the data centres themselves I think are out of action still, at least at least one of the AWS data centres over here is still out completely, but they were able to shift workloads out to another, to data centres elsewhere. But the cost can be much higher moving forward. For example, if Stargate UAE had been hit in the middle of a big training run, that is a cost to the company that would be an extremely expensive cost to the company. As we start to see these AI systems underpin just a kind of much broader sway of economic life, national security life and so forth, then outages could become much more significant than they have been so far.

So I completely agree that these costs will depend on what the underlying data centre has been used for. In terms of how to think about limiting attacks for them, maybe some sort of international agreement, some sort of norms. I mean, you've seen there are norms in wartime and all sorts of targets, legitimate targets versus not. Those are sometimes observed, sometimes not observed. I think it's hard to imagine, you could imagine a situation in which states exercise mutual restraint. Maybe they'd agree not to target each other's data centres because they do this kind of critical infrastructure for each other and they don't want to get into a kind of race to the bottom. There have been periods during the Russian and Ukraine war where both states agree not to target each other's energy infrastructure. At the same time, there'd be many periods where they haven't agreed to do such a thing. And so I think it's quite hard to imagine a situation where, especially for data centres that are being used by governments that have been used for military uses or for public sector uses, it's quite hard for me to imagine a situation where states totally rule those out as targets during wartime.

I mean, even during the U.S. and Israeli strikes on Iran, they hit data centres, they hit government data centres in one, that was one of the earliest targets they looked to. So if these data centres really do become central to how governments rule, how they project power, how their intelligence agencies work, it's hard for me not to imagine that they wouldn't become one of the first targets on a targeting list in wartime. If they're purely being used for civilian purposes, then maybe you can carve out some norms around mutual forbearance around those targets. But that would be difficult, especially in this moment where international law is not going through its kind of strongest period, let's say.

JC
Although doesn't it tell you very plainly why you have these international agreements or why they matter? So that could be my bias, given my training, but I just think the Geneva Convention was thought to be absurdly idealistic, but has probably saved a lot of lives over the years.

SWL
I think that's right. I think you may need a conflict where the costs are more salient of these things going down. Like right now, we don't really understand what it means for data centres to go down. The idea that the US are trying to agree, you know, the idea that you could have a kind of broad based international treaty focused specifically on data centres to me feels a little bit premature, just because much of society doesn't understand what these data centres are, much society may in fact want data centres to be targeted right now, given the domestic politics of this. So, you know, I admire the optimism on this, but at least for now it feels premature to think that there would be broad political support for this.

AS
And my question is, is that really necessary? Let's pause just a little bit, because what you said is right. I mean, it took 48 hours to get everything back on track, and that's the system actually not working as designed because it should have been faster. So the entire internet, how the data centres systems work is, you are obsessed with single points of failure. So if you have everything on a single data centre,then that's a particular high risk, insurers should demand high premiums, but that's not how the system is designed. The system is designed to have these distributed across several, several locations. So if one gets a hit somewhere, something gets physically damaged, it automatically heals itself and re-establishes. So the question should not be, should we have a protection around data centres, but are the players supplying this capacity have enough resilience in a single country and are not be held hostage by a single supplier? And that should be the question number one.

JC
So that's to be designed in, that kind of resilience needs to be designed in. And do you have to mandate that as a state or is it something that commercially you just do because that's the way you keep trading?

AS
When you sign an agreement with one of the big players, you have those kind of SLAs, which demands that type of protection, so the risk bears on them. That's why they then they are so harsh with us developing infrastructure because they need to comply with these extreme resilience targets.

JC
Right.

AS
So that gets embedded, but that's on a personal customer. If you're looking on a national level, that's a different question, which I don't think people are asking.

SWL
Exactly. I think at a national level from how the US is the only provider of these chips and so many of these things you know, if you're talking about single points of failure, much of the world, they are entirely dependent on our US government, ultimately. And we've seen that US government can change when it comes to some of these export control authorities, the US can turn off access. The US can mandate that the chips can't be sold to one place or it can be sold to another place, there are certain models can be accessed by certain people, but not by others. So yes, those sorts of things. So I completely agree that you need to build in resilience, redundancy across the board, that becomes much harder in a supply chain as concentrated as the AI one is around a certain very small group of states, many of whom have shown no hesitation about weaponising these interdependencies in search of leverage and another another.

So I would say it goes both ways. The United States, the Biden administration for example, was very concerned about the risks of building too many data centres in the Gulf, they were concerned that on a default trajectory, you would see a growing concentration of data centres in the Gulf. And so they introduced this policy called the diffusion rule, which set limits on just what proportion of global computing power companies could deploy to the Gulf states. And that was their attempt to try to avoid these types of single points of failure, this type of excess concentration or excessive points of weakness in one particular region. So I think it works both ways.

JC
Sam, could you just briefly go through some of those interventions that have happened? I know some of them have been quite quixotic and transactional and sporadic, hard to pattern, but there've been quite a few effectively trade barriers put up to the movement of vital parts of this new emerging system. And we have seen some pretty capricious trade policy making on the hoof from this administration in the US. Maybe we should just have a quick review of the sorts of things that have been going on that we need to be mindful of.

SWL
Sure yeah, so I can maybe give three examples of this. And I think one takeaway from these three examples will be that this isn't just a Trump administration phenomenon. I think any US administration is going to increasingly reach for some of these tools. So just to give you three examples of how I see these supply chains be weaponised. The first I think you will be probably familiar with, is US export controls on advanced chips to China. So in the Biden administration in 2022/2023, US introduced controls on the sale of advanced chips, the advanced chips that go into these data centres, to China and a group of other countries as well. Including some restrictions on chips that can be deployed to places like the Gulf States and other places where the US thought there was a risk of chip smuggling or digression to China. So that's one example of just a control on a civilian technology or a dual use technology to try to restrict a competitor's access to it.

I think a second example is the diffusion rule, which I just mentioned a couple minutes ago. This was a policy introduced in the last few weeks of the Biden administration and subsequently pulled back by the Trump administration that essentially introduced a global licensing regime, a global regulatory scheme for where chips could be deployed. It divided the entire world into three tiers, tier one were close US allies, they could continue to get unrestricted access to chips. Tier three were US adversaries and competitors, they had very serious restrictions on how many chips they could get. But tier two was all these countries in the middle, including India, including Israel, including parts of Europe, Poland was in tier two. And even in countries in tier two, the US imposed limits on how much computing power they could get access to. So this was a kind of pretty sweeping global regulatory regime for the deployment of advanced computing power.

And then just to give the kind of third example on this, I think that's your point about the kind of capriciousness of some of these instruments. Just in the last few months, we've had the Trump administration go back and forth on an export control that limited access to Anthropic’s most recent model Mythos, because the Trump administration was worried about the cyber capabilities of these models, the national security risks of broad deployment of these models. And so the Trump administration, and in part because of Trump administration's own relationship with this particular company Anthropic, they have a complicated relationship let's say, the Trump administration tried to block access by non-US citizens to Anthropic’s model Mythos. Anthropic couldn't really distinguish between who could get access to it, so they just pulled the model entirely. And then went through this kind of elaborate back and forth with the government is still essentially playing out in various ways to figure out how they could serve it more broadly. But I think for much of the world, the takeaway from there should be that overnight, the White House can essentially call up one of these companies and say, you can't release this model, you can't serve these models to particular countries, to particular players.

One aspect of Anthropic’s release strategy for this model was to give advanced access to a small group of hand-picked companies that form part of what Anthropic called Project Glasswing. All of those companies initially were American. So if you're the Europeans, if you're the Australians, if you're states around the world, this puts you in a very vulnerable position. You don't have frontier models of your own, you're dependent on chips created by US dominated supply chains in US data centres that are primarily in the United States. And your access to these systems, the access to the best frontier systems coming down the pipeline is dependent on essentially the goodwill of a group of US companies and the government that oversees them. And that obviously is a very vulnerable place to be in. Now, what your options are are much less clear, we can talk about that in a second, but I think these types of interventions by the US, some of them are kind of Trumpian and you know, transactional and capricious.

But others, I think are inevitable when you have a technology like this that has major national security implications that is being developed by private companies and that the government does not yet at least have a kind of coherent, fully worked out framework for thinking through, you are going to see restrictions on who can access these systems, you are going to see, especially with compute constraints, you are going to see rationing about who can access them. And in many of these decisions, when it comes to access, when it comes to rationing, the US government is going to prioritise US interests and US companies interests.

AS
I would just add that that process actually lasted a little bit less than a month. Because the US is also keenly aware they're not competing alone. So what the incident did, at least from my perspective, was to increase the awareness on how vulnerable, like Sam said, we all are. And that push, and you could follow through on all the developer boards and everything, how automatically everyone started developing their own model routing strategies to use other models, including the Chinese models. So that effectively create a mistrust towards US and open up the possibilities for other models to creep in into established companies processes. And that's a balance.

JC
Isn't that an irony that somehow or other we've finished up with an organisation in China that is more open and creates more open access to its technology than their equivalent in the United States of America.

ML
Cleaning Up is proud to be supported by its Leadership Circle. The members are Actis, Alcazar Energy, Arup, Climate Imperative Foundation, Copenhagen Infrastructure Partners, Cygnum Capital, Davidson Kempner, EcoPragma Capital, EDP, Eurelectric, KKR, Mitsubishi Heavy Industries, National Grid, Octopus Energy, Quadrature Climate Foundation, Schneider Electric, SDCL and Wärtsilä. For more information on the Leadership Circle, please visit cleaningup.live. To keep up with all that's going on in the Cleaning Up universe, make sure you subscribe to our newsletter. Written and edited by my longtime New Energy Finance and Bloomberg NEF colleague, Angus McCrone, it comes out every second Monday. Angus provides the latest on the episodes we're recording, the events we're hosting, stories we're watching and what Bryony Worthington and I are up to. To sign up for the Cleaning Up newsletter, visit cleaningup.live.

JC
But can we just circle back a little bit because the scenario you paint there is alarming if you live outside the US and feel the loss of power. But there are more and more instances of some kind of backlash within the US from communities where these data centres are being housed. And it has a little bit of political momentum beyond the local resistance to a physical infrastructure. And it seems to be coupled with a growing mistrust of the enormous power of the private corporations that are involved. And so is there some possibility that there is control from within the US that could manage this disproportionate power that the US has over the rest of the world? Or, and I know Afonso you've got some views about the extent of the reaction against data centres, and how in fact, they can be made very positive for a local community, at least in your experience in Europe. Can we just play this out a little bit about how far should we go with understanding the resistance?

AS
So my field experience has been strange to say the least, because when we announce a project in a new community, there's a lot of willingness to subsidise. Especially when looking at locations where they're economically deprived and that they see as a great boon of investment that will create big ties to the community. And there's a lot of good proofs that the large scale data centres do that. But NIMBY is getting to the data centres as well, mostly driven by people who don't even live in those communities and try to be very vocal about it. We had on a specific location, trucks of people, buses of people, coming to protest who have actually nothing to do with the community. But we were doing good work with the community so they stood up for us. And you see that in Europe, and you're starting to see that in the US where the drivers are not whether data centres are a good member of the community, truth be told that the data centre industry was never very transparent, and that's a lesson that needs to be taken from the energy industry. But yet, most of questions, doubts and blockage are more due to myths than anything else.

JC
But we've got a leader of one of the parties here in the UK, the Green Party Zack Polanski, who has called for a moratorium on the building of any new data centres for very exact this reason, it has become a political issue that can garner support from all parts of the country, not just those who may be resisting it because it's in their location.

SWL
And to be clear, in the US this is now a major political issue. This is not just the Green Party calling for a moratorium, within places like Texas even talking about more data centre moratorium right now in the US. So this is one of the rare kind of bipartisan political issues right now is opposition to data centres. So I think the political backlash is real in the US. That's one of the factors that's going to make it much harder for the US to sustain the pace of the build out within its territories that we've seen. And so a lot of companies now are looking for places they can go elsewhere. That was part of the rationale for the Gulf, but it also creates an opportunity for Europe now if it can capture some of these, build out some of these projects if it wants to. Australia now is kind of angling to try to attract multi gigawatt projects in Australia. But I think the domestic politics of this are quite, I think they are most advanced in the US just because that's where the build out has been and has had the biggest imprint.

I would say it's still unclear to me how much of the backlash is driven by specific real kind of issues with local projects, whether it's impact on energy, on local energy prices, whether it's air quality, pollution, all the kind of litany of complaints that people have, or whether it's more of a kind of symbolic backlash against AI writ large. And data centres are kind of one of the leaders, you could be one of the kind of visible signs of this technological revolution that a lot of people don't feel included in or represented in and see this kind of threatening to their way of life. And so distinguishing how much of the data centre backlash is a kind of real local political issue that you can deal with by being a good citizen to the community, versus how much of it is just this kind of broader resistance to AI writ large. I think that's unclear for now, but I do think the way the build out unfolded in the US, NDAs, backroom deals, no transparency from some of the wealthiest companies in the world just trying to ram these projects through, I do think that has bred some of the backlash that we're seeing.

AS
I agree. And that's my view, and I come from the energy industry into the data centre industry. So my view is it was naive from data centre companies to grow and assume that they can deploy large projects as they used to deploy small ones where they can be super secretive, they don't need to say anything to anyone. But at the time, the biggest of those projects were 10 megawatts. And the 10 megawatts doesn't make a dent on a community. If you go to a gigawatt, it's a completely different system, it's a completely different beast. And it requires a much closer relation with the local community in order to get social license to it. I want to mention one study that scared me on the outlook of the geopolitical unfolding. And the Stanford 2026 AI index, where they were measuring the confidence that people had on AI, had 84% of respondents in China being excited by AI, and that contrasts with only 38% of people in the US. At the same time you had another poll, the Edelman poll, which said 87% trusted AI in China versus only 32% in the US. Now this is scary, right? Because you only have two big players. One of them has grown very fast, it controls most of the capacity and the chips. The other one has almost limitless power to enact and has a very willing population to push through. And Europe is nowhere between these two, it's years behind on these two, so how will the world unfold?

JC
So that there might be some general proposition you could make to explain those quite remarkable stats to do with confidence in the future generally, that the ability of a society to look forward and feel like life is going to get better for me and my children and their children. And the lack of that feeling is a powerful political sensibility in the US and indeed here in the UK and across many other of the liberal democracies. And it's a problem in of itself, but here it does appear to be attached to a fear of the new, and maybe also a fear of the dominance of a small group of technocratic institutions that happen to be called companies, but have more power than most countries.

SWL
I would say there is a general trend when you look at public polling on attitudes towards AI that across the global south, across developing countries, there is more excitement about technology. You can tell lots of different stories for why that might be the case, but there's lots of different kind of theories you can tell for that, but it generally is a kind of a broad trend across the board that in advanced democracies, industrialised societies, people are a little bit less excited by AI than they are in developing countries. But I agree that this technology is attaching itself onto broader political narratives, broader currents, and public opinions are going to be intertwined with that. As you said, these companies are powerful political actors in their own right, and use of them is going to be shaped by their broader political leanings one way or the other.

AS
I would challenge you to do a small intellectual exercise. So let's imagine we are in the 1830s and the world is we speak globally as we are speaking right now. And instead of AI, we're talking about the steam engine. How would you feel, we're in Europe and everyone's against the steam engine, because it pollutes, it makes noise, it doesn't make sense. The US has been driving it, but it's fading down, and China says, bring it on, we'll build all the steam engines that we can do. What would happen to economic reality across the globe?

JC
I mean, listen, I think we've arrived a point in our discussion, we probably need to pull some strands together. But one of them obviously is just about power in every sense of the word. Who has access to the compute power? Who has access to the resources required to make the compute power? Who has access to control the access of others to the compute power, and then how is that power distributed? If it's already a large gap between the very powerful and the rest, how do the rest organise to increase their power in order not to be utterly dominate and kept out of an extraordinary transformation in the whole of the economy, not just one or two industries.

So it's perfectly proper that we have conversations about sovereignty, sovereign use of power, international cooperation and treaties, defense. And in the middle of it, there are private companies that raise money from shareholders and borrow from banks and have to which, of course, we haven't had time to discuss which is a parallel risk, and yet they are all interconnected. And we've got to figure out how best to understand those connections and find a way of making the transformation alluring and not terrifying.

SWL
And I do think that some of the domestic politics of this in the US create opportunities for other countries to increase their own power when it comes to these things. So just to give a kind of concrete example right now, if it becomes significantly harder to build a data centre in the US because of domestic politics, but also good constraints and so forth, that creates opportunities for other countries that can bring power and computing power to the table. So to give a concrete example, let's think about Australia. If Australia can go to one of the US labs, the US companies and say, we can give you access to a big data centre that we can build relatively quickly, we can credibly say we can get this up and running in two years or some realistic timeframe. That then gives them leverage over the US to ask for concessions in return, to ask for guarantees on access to the best models in the future, to make sure that Australian companies are included in any future Project Glasswing style restricted access program.

So when Anthropic is considering which 30 companies should get advanced access to these models, if they're dependent on a big Australian data centre, then Australian companies will have a pretty good chance to say Australian banks should be in that project. And so this is one way where you're seeing some middle, you talk about how can other companies, how can other countries try and carve out some leverage, some power, there are lots of different, maybe they band together, they form a block, they do some sort of international treaty, there are different pathways. But one of the ones that we're seeing some countries gravitating towards now, is trying to use their ability to bring data centres to the table as a bargaining chip, as leverage to ask for concessions from US companies or from the US government.

Now that doesn't guarantee that you would be exempt from a capricious administration in the future shutting you out. Again, these companies are very powerful, but I think, at least in my view, they're still clearly less powerful than US government. But it does mean that you create a influential voice within the US political system that will lobby on your behalf. Because if one of these companies doesn't want to lose access to its data centre in Australia, then they will lobby the US government to say, let's include the Australians in any advanced project in any research and access program, let's not shut them out of us because you know, we really need this data centre by providing us.

And so then you can see some of these, that's just one example of a way in which another country can try to carve out some leverage, some actual bargaining power to bring to the table when it's negotiating with the US using access to computing, using access to sites. And this goes back to the discussion earlier about siting, using access to siting, not so much the chips are still the US provided, US dominate, coming from US dominated projects, but using access to sites as a source of leverage over the companies and potentially even over the US government. So, yeah we should debate how much leverage that actually gives you, but it certainly gives you more than nothing with these data centres just in Texas.

JC
And Afonso, so let's just go back to Europe for a moment if you may, I don't know whether you want to do this or not, but can we just try and imagine how Europe would catch up, find leverage, find a way of really taking away some dependency, building some resilience in its capacity to compete?

AS
For Europe, it's a survival challenge to embrace this. We have an aging population, a lot of it is retiring in the next decade, we have productivity going down. And unless we embrace this as Europeans, we're going to stay behind. How that can be achieved, right now, you have a wide, completely different approach from each country, Spain just took some very harsh measures that discourage data centre investment, the UK is trying to attract saying we're okay, we're better than the rest of Europe, and then other individual countries are trying to do in their own way, enablers of data sensor because they're starting to perceive how risky it is to depend on data centres and indirect way to R&D from US or China to evolve. So for Europe, it's a survival issue.

JC
I think we will nonetheless still have the kind of resistance that we've talked about in the States in Europe. I mean, you can see there's a big Google facility going up in Austria where the power demands are pretty much the same as Vienna. As soon as people get those sorts of scale, they start to panic or worry that this can't possibly be happening. And then there was the Xlinks project that some of us are familiar with and it's all very exciting, and it's supposed to land in Devon and build a massive great data centre there, and already the Daily Telegraph is publishing scare articles about what that's going to do. And then we've talked briefly about the Green Party, so I think we should anticipate there will nonetheless be resistance despite a parallel feeling that we’ve got to catch up. And how that gets balanced out, Lord only knows, but that is the tension that we're grappling with.

And it will have all of these elements we have talked about. You'll have to have a blend of security conversations that are done with the right people, both in the military and in military defense and all the people understand how to make any infrastructure resilient from attack. And then you'll have to do the same thing with a cyber security community. And then you'll also have to deal with all of the local issues over access to energy and water and the like. And in the meantime, figure out how you're going to get access to the scarce resource of the chips and figure out where those dependencies could in some way be eased by negotiation probably at this stage, I can't see much else that Europe has got. Maybe, you know, one substantial manufacturer in Netherlands, but that's that's it really. So building a plan has got to include all of these elements, hasn't it?

AS
Completely agree, completely agree. And to your point, Europe needs to figure out, what does it want? We are on the cusp of one of the most transformative general purpose technologies that have come across. And with every general purpose technology, you have a step change on how companies, governments work, but you also have a step change on power consumption. I love the steam engine, the steam engine analogy, because if we look at how much power the entire globe consumes pre-seam engine, which was roughly 3,000 terawatt hours, the globe of biomass, and in a couple of generations, it went to 12 to almost 13. So there's a step change whenever you introduce general purpose technology. So to ignore it is trying to ignore that this is a step change technology.

Now what worries me that the steam engine did not bring is the severe imbalances it can create in terms of global, because what AI brings is the ability to deploy additional workers, additional productivity to the people who are embracing it. And I had this discussion in one community that I was speaking to, and they were pretty much against it saying, Oh, but what about our kids jobs and so on? I told your kids jobs are dependent on whether this country can adapt to AI or not, because otherwise, at the pace that this is going, we'll be outsourcing everything we do to China.

JC
This is a good point to conclude I think, folks. Bearing in mind, it's all part of Cleaning Up’s program of work on the energy transition, electrification, a point you just made Afonso is really a vital one to dwell on as we part, which is that this whole development of the technology is creating staggering demand for power, for electricity. And more demand for electricity is, as a matter of fact, creating more demand for clean electricity, not universally and not 100%, but it is, as a matter of fact, creating more demand for clean electricity. And the speed of deployment is on the whole and works in the favor of clean technologies, and the same with understanding storage better and understanding how gas turbines might support the electrification of a very substantial demand like a data centre. And more demand for power is actually not necessarily a bad thing for the energy transition, improving a point of what these technologies can do. And that sometimes gets lost in the fear of the change, does it not?

AS
I completely agree. My first project was all about getting a large scale data centre completely with as much renewable power as possible. And you see already some initiatives that bring your own energy type of initiatives that started in the US, some European countries are trying to adopt it as well, are actually good enablers of additional clean energy into the system. And I would add one thing that the fact that you have a large consumption centre into a grid on intermittent power tied with best, can actually be a benefit to the grid itself as if properly managed and be a true enabler of the next stages of the energy transition.

JC
Sam, would you like a concluding remark? I realise that that is more of a nuance rather than a kind of clear cut for and against, but what about you? How would you round things off given our discussion on these interrelated security geopolitics and data centre theme?

SWL
Yeah, no, I agree. I agree with everything you just said, I think it can be easy to frame this debate exclusively in terms of trade offs between different matters. And they're all real trade offs. So trade offs between national security and preserving building out compute for financial security purposes, versus preserving things the way they have been or between not trying to drive, trying to keep energy use down and trying to capture the tax revenue that might come from attracting data centres. And so there are real trade offs here, but I think there are also ways in which you could do these build outs that both serve local communities as Afonso talked about in the way that he's trying to do these build outs, that also further the energy transition, that also secure a country's kind of geopolitical position moving forward. And all good things naturally go together, I think there are going to be real trade offs between different values and you are going to have to choose on some margins between different priorities. But there are ways of doing these build outs that I think can mitigate some of the downsides that we've talked about, whether it's on the energy side, whether it's the local political backlash, and they can show up a country's position and can reduce some of its vulnerabilities, those will always remain to some extent.

JC
Very good. Well, I just want to thank you very much for joining us on Cleaning Up and having this vital conversation and sharing it with an audience, hopefully out there in the world grappling with these issues and trying to figure out what's what makes sense. So it's been a great pleasure to have Afonso and Sam on this program, and we look forward to continuing the conversation in no doubt in other episodes down the line when one Michael's in charge, take care.

AS
Our pleasure, thank you.

SWL
Thank you very much for having me.

JC
Well, thank you very much to Sam and Afonso for a very good, very thorough conversation about a really tough subject connecting data centres, security and geopolitics. I hope those of you out there enjoy that and found it useful to you, and we look forward to having your company on future episodes of Cleaning Up.

ML
Cleaning Up is proud to be supported by its Leadership Circle. The members are Actis, Alcazar Energy, Arup, Climate Imperative Foundation, Copenhagen Infrastructure Partners, Cygnum Capital, Davidson Kempner, EcoPragma Capital, EDP, Eurelectric, KKR, Mitsubishi Heavy Industries, National Grid, Octopus Energy, Quadrature Climate Foundation, Schneider Electric, SDCL and Wärtsilä. For more information on the Leadership Circle, please visit cleaningup.live. If you're enjoying this episode, please hit like, leave a comment, and also recommend it to friends, family, colleagues, and absolutely everyone. To browse our archive of around 250 past episodes and to subscribe to our free newsletter, visit cleaningup.live.