Loading summary
Transcript26 lines
- [00:01]
A
Your film is now ready to be shown. Good morning. I'm Justin Hendricks, editor of Tech Policy Press. We publish news, analysis and perspectives on issues at the intersection of tech and democracy. When the essay Europe 2031 was published in June, it set off an immediate debate in Brussels and beyond. The report, a fictional near future scenario co authored by a group of AI researchers, argues that Europe risks economic and political irrelevance if it fails to compete at the frontier of AI development. Within days of its release, the United States government ordered Anthropic to restrict access to its most advanced AI models for non US users, an event the fictional scenario had projected for 2029. In this podcast discussion, Tech Policy Press Deputy Editor Ramja Jahangir is joined by co Author of Europe 2031 Maximilian Nigela and Frederica Kaltener, who is an author in a three part series for Tech Policy Press that argues that Europe's dependency problem runs deeper than frontier model access to debate what Europe is actually getting wrong and what it would take to change course.
- [01:10]
B
Here's Ramsha One viral essay has dominated AI policy conversations in Brussels and beyond over the past two weeks. Europe 2031, a sprawling fictional near future scenario, argues that Europe is sleepwalking into technological irrelevance, that it accounts for roughly fifth of global GDP but remains almost entirely absent from the frontier of AI development. Its prescription is ambitious and contested massive compute investment, including in partnership with American hyperscalers and a coalition of democratic allies to negotiate access to advanced AI. The critics are equally forceful. They argue that the report mistakes deeper dependency for pragmatism and that Europe's real problem isn't a lack of data centers, but a structural entanglement with US platforms that no amount of investment alone cannot bind. Today on the Tech Policy Press podcast, we bring those two positions into direct conversation.
- [02:09]
C
My name is Maximilian Nigele. I'm a research affiliate in Technical AI Governance at the Oxford Martin AI Governance Initiative, which is a part of the University of Oxford that many deals with quite a questions of AI governance in all different directions International Air Governance Technical and have been affiliated with RAND also since a couple of weeks ago, where I led the research area on European Frontier AI at the RAND center on AI Security and Technology, but then left my job recently, among other reasons, also to work on the Europe 2031 project, which came out a few weeks ago and which I worked on basically the last two months or so. Excited to be here.
- [02:49]
D
Excited to be here too. I'm Federike Kaltoiner. I work since many years on Tech policy, and in the last few years, specifically on the discrepancy between the reality of the AI market and some of Europe's ambitions for AI, including as an advisor for the AI Now Institute. And in this function, we've also published a series on tech policy press that analyzes the European AI market. And I'm very excited about this discussion.
- [03:16]
B
Thank you so much, both of you, for this very timely discussion. Maybe we could actually start by introducing Europe 2031, a very, very viral essay that was published, I think, around two weeks ago, and it's been the talk of the town. So, Max, would you like to introduce that to our listeners?
- [03:33]
C
Sure.
- [03:33]
B
Happy to.
- [03:34]
C
Europe 2031 is a kind of scenario piece, as we call it. It's a bit like a story. It's told from the perspective of two people, a young European Commission policy officer and her entrepreneur friend who's moved to Silicon Valley. And they sort of have a conversation throughout the scenario. The scenario spans between 2025 and 2034, when there's an epilogue, basically. And the whole topic of the piece is just European AI policy and the way in which Europe basically falls behind the big AI powers due to its lack of agency, but also partly due to just incentives that are not aligned between the various actors that are involved. And the story sort of unfolds slowly beginning in 2025, where it's still based on real events. So the first couple of chapters are all real stories that many of us have heard about in our policy, sort of like if you're in the policy bubble, you know, about these events that we talk about. And then at some point mid-2026, it turns into a something that we call a disciplined forecast or a disciplined imagination. So it's fictional, but the story is told in such a way that it could have happened based on our best judgment of what we think, where the trends are going. And, yeah, many people maybe have been shocked by it, it turns out, rather badly. So in 2031, basically, the European economy is kind of hollowed out. The Americans and the Chinese make offers to buy up European industry. There's basically no geopolitical agency left in Europe. And in the end, Caroline, the young policy worker who used to work in Brussels at the European Commission, kind of emigrates to the US to live with her sort of entrepreneur friend there. And nobody really knows what's happening with Europe, but it's implied that it's not a positive situation. And the reason why we wrote this piece is because we are kind of like a ragtag group of people, think tankers, investors, researchers of all kind, who all came together because of a shared diagnosis. And that shared diagnosis was that European AI policy was not going well. And we tried to change it, but somehow nothing stuck. And we tried to write papers about it, we tried to explain it to policymakers, but you just needed a kind of viral or slightly subversive moment to really point attention to that. And that's certainly my motivation going in, but maybe as the last thing that I want to say, so this was written by eight different people. And none of us are, in a way, like a think tank or anything like that. We're all like different people. We have different opinions. And so when I speak about the project and everything, then that's my position and not necessarily the position of everybody else in the group. And yeah, we also didn't get any financing or funding for it. This was just all based on our personal savings and, like free time, basically that we invested. And yeah, I have to say I was super surprised by the impact of the piece. I personally didn't expect it to go that viral, but it did, and that's amazing. And so we're having lots of discussions, which was exactly also the intention. So looking forward to that.
- [06:41]
B
Thanks. Max. You mentioned obviously in the report in the essay that you that Europe would lose access to Frontier AI models by 2029. And it's interesting that sort of happened within, I guess a day, few days, or maybe a day after publication of this essay. Could you talk a little bit more about when that happened, what happened? And then do you read it as evidence for the same conclusions that you pointed to in your essay?
- [07:07]
C
Yeah, I can remember when it happened. I was basically somewhere in London at a conference and remember waking up in the morning and my phone was full of messages and people telling me, oh my God, oh my God, something happened. It's amazing. Like, they. They looked, a prediction came true and I didn't really understand what was happening. And then I checked on the news and then lo and behold, indeed, the US government had restricted access to one of Anthropic's most powerful models out of national security reasons. Basically, I tried to sort of figure out what actually happened behind the scenes. And I think it was in fact a mixture between just wanting to teach Anthropic a lesson because there has been the simmering conflict between the US government and Anthropic. But then also there were real national security concerns around the model potentially being jailbreakable with some specific methodology. And so they restricted access. This is something that I think many of us in the scene, let's say in the AI governance bubble, had already expected to happen at some point in the future. That's why we also wrote it in the scenario. That it happened so fast was definitely a surprise for us. We didn't see that coming and it also happened in a slightly different way than we had predicted. But I think this general pattern that as models become more powerful, they will be restricted and also as compute becomes more scarce, governments will take prioritization decisions, I think that was definitely in the cards and indeed something that we've recently seen now which also basically vindicates. One of the predictions in the scenario was that China now restricting access to certain open source models, which I think had a bit of less attention. But that's also something that happened. And of course we were happy because I think that definitely boosted the virality of the essay. It came exactly at the right point. So in a sense we got very lucky with it. And yeah, it somewhat vindicated our prediction, but I think this was something that not only us, but many people in the bubble sort of saw coming at some point.
- [08:59]
B
Fredrike, you have been running a series on what the AI market of Europe looks like in partnership with Tech Policy Press already and speaking to some of the structural dependencies of the European stack. So what does that dependence on foreign AI actually look like in practice? And then a follow up to that question that ties to what Max was talking about is when economic competition over AI becomes a geopolitical competition, who should control access to the most advanced systems?
- [09:30]
D
There's so many questions at the same time, maybe let me start by saying what I appreciated about the Europe 2031 work. When reading it, you feel a real sense of urgency. You're left with a real sort of sense of concern. And those are two things that I share. I also thought it's been incredibly well done. This is how our policy papers should be written. Very engaging and it definitely started a discussion. My concern is that it uses a single scenario to make an incredibly controversial recommendation that I'm worried about will make Europe even more dependent than it is already. That's where I'm coming from. And I can take a step back and talk about what our thinking is about the European AI market and what this dependency looks like and what what I'm concerned about. So what has really been interesting in the last few years, Europe's AI sort of industrial strategy has been characterized by a large degree of magical thinking. So the strategy focused on two, in a very simplified way, on two different aspects. So on the one hand there's been an effort to invest in compute for training large scale models, obviously with insufficient funding, but that's been one part of the strategy and the other one is boosting AI adoption. And what we wanted to achieve with the series was to argue that Europe is structurally dependent on a very small number of US firms across the AI stack in ways that are really difficult to untangle. And this sort of stack by stack, supply and demand thinking that we've seen from the Commission doesn't just fail in making Europe independent, it also risks making Europe more dependent. So we start this series by looking at the application layer and what we notice there is that if the strategy is to boost the number of AI startups in Europe, we have to take a look at what these startups actually are. And many of the most celebrated AI startups in Europe, not all of them, but a significant share, sort of focus on the application layer. And in many cases they build on proprietary models, let's say models by anthropic to make AI useful and to build on top. There's nothing wrong with that approach. It actually is. It's not easy, it's kind of difficult, but it's incredibly risky as a strategy for a company and also as a sort of industrial strategy for Europe. It doesn't really make us more dependent because as Max sort of mentioned, it's not just that models can be turned off, it's not just that models at the frontier, access to these models can be revoked, but also all kinds of models, if you're building on top of them, you're dependent on another company's terms of services, on their pricing, which is currently heavily subsidized. And also for Europe it means that the more AI is used, the more money is being melt primarily elsewhere. So that's sort of the first piece and the second one focus on cloud. And our argument here is that Europe's strategy is primarily focused on expanding COMPUTE for training. So for building European models, it hasn't really looked at where these models are being run. And we take the case study of Depot, that's sort of like the counter example to these application layers. AI startups. DeepL is the Cologne based translation AI company, has their own proprietary models, proudly European trained on European COMPUTE run on European service. And they recently entered an agreement with aws. And in this piece we take that apart and ask why is that? That has been interpreted as the company's lack of commitment to Europe, but in reality it's actually a brutal business decision for A company that wants to serve their models globally, AWS is the best choice. And there's sort of like a gravitational pull of US hyperscalers that mean that lots of different sort of layer by layer interventions in European AI industrial policy end up making Europe more dependent. So interestingly, in a way, the reason I'm concerned or the sense of urgency I feel stems from partially a similar analysis in the year of 2031 proposal. And that is the fact that if Europe is just pushing blanket AI adoption in the absence of tackling its dependency, there's a real risk that we will be even more dependent.
- [13:53]
B
Max, I kind of want to come to you on a lot of things that Frederic has pointed out. But most importantly, when we talk, and this is the broader debate with digital sovereignty in Europe also, what do we mean by that? And a lot of people have described that problem, but there's also this tension of are we talking about the same problem when we talk about digital sovereignty and the Europe's AI problem? So how do you define this? And also, just to go back to what Frederick said, that a lot of people have spent time arguing that too much independence is important and not enough time building capability. And the other side is that that's the wrong trade off to accept. So what's your argument in this? And yeah, how would you describe sovereignty? And is it achievable?
- [14:44]
C
Yeah, maybe. Let me first come back to some of the points that you made, Frederica, and then I can come back to the digital sovereignty point. So, yeah, first of all, thanks for the kind words. We're happy that the scenario sparked some discussion and was sort of emotive. And I think we actually pretty much aligned on being worried about European dependence. I think Europe 2031 is very much a piece about exactly that, about European dependence. And so it's interesting that we, that we shared the same goal, but maybe have a different diagnosis on how to alleviate that goal. And so I'd be interested in figuring out where we can converge there. My read of what you're saying is that we have this recommendation that Europe's dependency partly is just a result of a compute shortfall. And maybe this is a good point to actually introduce how I think about digital sovereignty. I think sovereignty in the AI space is very much a result of mutual dependency. I sometimes like to distinguish between competitiveness and sovereignty. Competitiveness is just the degree of value creation that's happening on your soil and the economic value that you can reap from that. I think sovereignty is more the question, can somebody else pull the rug from under you. Even if you have a lot of value creation, even if you have an amazing industry, somebody else can just turn off the knob basically and take all of your capabilities and value creation from you. And I think it's that latter sense in which I'm just worried that we are running into a very strong one sided dependency as Europe. Because if compute is really the substrate of the future, if this is the thing that we need to run AI systems in our economy, in robotics and everywhere, then if just one side has all of the compute, then that's going to be bad from a geopolitical perspective. And I think it's there where we came in with our recommendation of, look, need to build compute one way or the other. We currently don't see that Europe, European companies are building compute because they don't have a business model. Sure we have some gigafactories, but these are run by sort of not AI companies, but other companies and they don't want to build as much compute. So we need to build the compute somehow. And we're not saying that we should build it only with American hyperscalers, we're saying anybody should build it. But it's clear that we need more of this. And if European companies are not going to build it, then we should also consider letting US companies build on our soil because that gives us the sovereignty advantage in the sense that if we have more compute on our own soil, then in the future, if the US government comes around and wants to restrict access to models, we have a very strong bargaining chip because we have a lot of compute which in theory we could shut down if we have physical access to that. Yeah, please, maybe just one more point because I think we actually agree on a lot of stuff, but I want to maybe prompt you with that as well. I think we totally need to strengthen our own ecosystem in Europe. And I think the way we do it is by just creating good enabling conditions for companies to grow. So like scaling capital, bringing the best talent here, making it as easy as possible to build. And that's how we're going to get AI companies on European soil. And so, yeah, curious how you think about those points.
- [17:51]
D
I think it's worth really spelling out what the compute recommendations mean. So the Recommendation is by 2031 that Europe should that about 15% of global compute should be on European soil. That sounds quite modest. However, if you do the calculation and the math, what that actually means in practice it would mean adding the, the entire electricity consumption of Germany to 2024 in a context where there are already conflicts of energy prices, where there's local resistance against data centers. I'm not saying you should never do that, but I think it's really worth stressing that this is quite, quite a controversial wartime effort build out of compute that's being demanded. But what I find even more interesting is the assumptions behind the scenario. So I think Europe 2031, it sounds really reasonable and pragmatic, but it actually isn't scenario planning. It's one single scenario that relies on a number of technical assumptions, but also relies on the number of economic assumptions that all need to be true in order for the scenario to actually play out the way that it does and for the recommendations to make sense. One assumption is that there is a sort of frontier that models, larger models, keep being more capable and that this trend doesn't stall. That's a hypothesis. It's what's been observed the last few years. This may continue, but it may also not continue. So that's one assumption. The second assumption is that being at the frontier creates an advantage that's decisive. Right. So in this scenario, models do not commoditize. So there, there will be a real difference between the models at the very top or the one model at the very top and everyone else behind. And it's not possible to catch up. And also, and I think this is the most sort of important premise that also it is at the frontier where most, if not all economic value will be generated. And all of this happens not in decades, but in a couple of years. And even the most important thing for me is already if you look at how people are using models now, it's not just that models keep getting more capable, but that the additional cost of using larger models, and especially models at the frontier, are worth their additional cost. So they generate more economic benefits than they cost. And all of this could be true, but I think it could also really not be true. Right. There's the scenario where models commoditize, where there are multiple models that sort of people use interchangeable or what we've seen recently, computers currently heavily, heavily subsidized. So currently it does make sense to use the largest model for every possible task because you're not actually really paying the real price. Sure, if the hypothesis is that prices go down, that will change, but if they're not, then you end up in a situation where it makes a lot of economic sense to be very discerning about which model you use for what kind of use cases. And if all of these conditions don't hold, we're in a very different kind of market that looks very different and where very different interventions make sense. And I think for policymakers in particular, especially since they know so little about AI and understand the AI market so actually not enough. It really is important to say Europe 2031 is a scenario, it's a hypothesis and it's also a bet. There are many, many other scenarios that you should think about and it would be really important to figure out why. What kind of reclamations interventions make sense in each of them. So that's sort of, I think, where we differ quite drastically. And then I think more politically it's sort of like. Because the scenario like it's communicated with a sense of urgency and inevitability. What really concerned me when reading was sort of that public opinion is seen as sort of something to be managed, right? So people, there's protest against AI, there's protests against data centers in the scenario, but it's not fully being wrestled with. The assumption is that people may not like this right now, but whoever has sort of control of the frontier will have make so much economic, will make so much money and this will benefit the economy so much that then ultimately some of that wealth will be redistributed. And I think that's again a massive, massive assumption that first of all this much economic value will be generated in the first place, but also secondly that people will benefit. So my concern, and I think it's interesting, I understand and I think academically it's quite smart to distinguish competitiveness from sovereignty. However, I do think what are we doing all of this for, right? AI is not an inevitable force that we need to react to, but these are highly political decisions that need to be made. And where Europe is right now and what sort of the kind, what the continent really needs is a robust, long term healthy economy that also meets many of the other goals that we've decided that we want to tackle. From the environment to climate change to inequality. All these are kind of issues that Europe wants to tackle. And for this we need a robust and healthy economy that generates money that people actually benefit from. And if you define sovereignty as that the dependency is not just the risk that somebody to turns off a model or turns off inference or turns off access to compute. The risk is also that most of the money is being made elsewhere, that you are dependent in ways. And I think this is the dynamic we really outline in our piece. The problem of hyperscaler dependence is that the hyperscalers operate along the entire vertical market. So if you're dependent even in industrial AI, if you, if you're dependent on aws, you're dependent on your competitor who also is venturing into industrial I and robotics. If you're dependent like the startup Lovable, right? If you're the vibe coding Startup Lovable builds on anthropic models it's built on, it builds on primarily Google Cloud, you're dependent on companies that want to enter the same markets that you're also in. And that is a really long term risk for a healthy economy. And I think that's one of the reason why sort of this single minded or quite narrow focus on compute is quite a risky bet. And then finally, I think even if you, if you agree with all the premises that the scenario lays out, the only protection against compute being weaponized or shut off are sort of like how the scenario describes it, security guarantees. And I'm incredibly inconvinced that this is sufficient. Right. We have seen that in the geopolitical moment that we're in, the United States administration has made quite irrational decisions that don't make economic sense for themselves either. So the idea that if we have more compute on European soil, even if most of this is owned effectively by US companies, legal securities will help us sort of like prevent a scenario where computers being turned off. And I'm really not convinced that this is the case. It's almost the equivalent of that we had with Russia and Ukraine for a long time. We're so entangled with Russia, we import so much of the oil that there never will be a conflict because it doesn't make economic sense for Russia. Turns out economic sense is just sort of one of the reasons why countries act seemingly irrational because there are many more other factors to consider. I could talk more about this, but I think this is sort of my main concern. And where we differ, I do want
- [25:25]
B
to go back to the tension before coming to you, Max. I think when I said that you both are describing a problem, but are you describing the same problem? But I think there are two questions here. One is obviously to build scale in compute because Europe is behind, and then the other question is who controls the infrastructure and the distribution and the ecosystem. So I think there are two problems we're talking about and I wanted to point that out. Max, do you want to address some of the concerns Fredrike pointed out?
- [25:54]
C
Yeah, happy to. I think you started out with saying it's a big proposal, it requires building lots and lots of stuff in a very short amount of time. And that's absolutely true. We say at the Beginning of the scenario at the very top of the website that the challenges by Frontier require the most ambitious political agenda in post World War II Europe. And we really mean that. That's not just a catchy sort of phrase that we use, is actually what we mean. And we think that AI will be similar to the Industrial Revolution. And in the Industrial Revolution we needed to build lots of roads, lots of railways, lots of infrastructure that at the time were really appalling to a lot of people and they didn't like it. But in the end we just had to do it. And nowadays we take it for granted that we have this infrastructure. So I think there will be big trade offs, I think there will be difficult decisions around these infrastructure buildouts. But we think that is what it takes. Regarding the scenario, you say that it's a scenario, it's a hypothesis, it's a bet that's based on assumptions. And that's absolutely correct. I would never claim it's going to be the reality because we can't predict the future. We try to make a best guess on what the future will be like. But absolutely it's based on assumptions. And I guess we can now discuss specific assumptions behind that. Will the trends continue, for example? I mean, that's a longer discussion. Maybe briefly on the point whether models will commoditize. I think that's interesting. I agree that it could be the case that models commoditize increasingly. I think even in that case it will remain the case that compute capacity will remain important because it will still determine the degree and the intensity to which you can actually run models, even if these models are commoditized. So that's maybe a brief reply to that point. But I mean, I completely agree with you that there are different scenarios. And I also agree with your point that we need to prepare for different scenarios. We're not claiming that this is the way it's going to be and that's the end be all and end all of everything. I think there are definitely different scenarios and we should prepare for each of them. And actually that's something I've been arguing with policymakers. I think we need to do more contingency planning, in fact, for different low probability scenarios that occur. So I'm with you on that point. And maybe also I want to say that we unfortunately had to make very difficult choices in writing this scenario because the future of AI is going to be extremely complicated and there are a million different things that we would have wanted to put in there. So, for example, we got a lot of criticism from people that we didn't focus more on the national security risks. Something that's very close to my heart. I think those are a thing I care about quite a lot, like biological weapons, cyberattacks, potential for loss of control. And we are actually wrestling quite a bit whether we should make this more prominent and the same for the whole public opinion bit and control who makes the rent. I think that's a whole nother discussion, which I personally am very worried about. I don't want to end up in a world in which a few tech oligarchs own all of the capital and all of the compute and public opinion is just being ruled over and democracy is hollowed out. I think that's a huge, huge risk and problem that I see and that I would have loved to weave more into the scenario. It's just that from a storytelling perspective, we wanted to make sure that this is about European sovereignty independence. And I think it just works better if we don't introduce like a million different strands in there. But I think you, you observed correctly that this is a prioritization decision that we made. But I want to just claim here that there are lots of other things that we think are extremely important in AI, and I think definitely the wealth redistribution point is a very, very big one, which we completely gloss over here. But that's just the way we planned the piece. It's a piece of agate prop. We wanted to start a discussion around sovereignty independence. And so this is just a design choice, but it's great. It's like, good that you point this out because I think it really is worth a much, much bigger discussion than this. Then, like, I think you made a few points around, you know, AI being this inevitable force. I think I completely agree with you as well. I think it's not inevitable. I think it will be shaped. I think the whole reason why I work in air governance is because I think humans need to shape the way this technology is developed and it needs to be in the interests of humans. And that doesn't mean that there won't be hard decisions. I mean, you mentioned climate change. You mentioned the environment. We will need a lot of electricity. The big hyperscalers in the us they are already building gas plants behind the meter, which is horrible for the environment. Which is horrible for the climate. But I'm not saying. But I also don't want to say that we can do everything. I think one mistake in European AI policy in the past has been that we have been claiming, oh, we can just save the climate and do everything and save the economy and this kind of everything bagel idea that I think has been very damaging and I think there has been a reluctance to actually explain hard trade offs to the voters. Maybe. Yeah. I think on the dependency point, on the hyperscaler dependence on the hyperscaler market and security guarantees and this idea that we run ourselves into dependency. So I think one thing I very much want to agree with you is that we need to build more indigenous capacity and more indigenous. We need indigenous hyperscalers, we need indigenous AI companies and so on. It just happens to be the case that we don't have them at the moment. It is very unfortunate. But our best companies, they don't go to the European hyperscalers. In many cases they go to the American hyperscalers because they just, in the current competitive market that we're in, they're just offering the better services and the American AI models are just better at the moment. And I think that's not a reason. I think that's not because Europe hasn't been taking care of its sovereignty enough or because we haven't done an AI sovereignty strategy by the Commission or something like that. I think this is just plain and simple economic policy. It's just the same competitiveness policy that Mario Draghi talked about that everybody has been talking about for years on end. If we had the best talent, the best capital markets, the best conditions for our companies to grow and scale extremely fast across all of Europe. I think this is the way we can solve this problem. I think the way we can solve this problem is not by making specific predictions about which models we should buy or maybe even ideas around. Oh, we only need to buy European. I think it's mainly just the economic policy that we need to fix. And I'm curious how you think about that point.
- [32:29]
D
Federice Maximilian said so many things I'd love to respond to. I think what's really interesting about obviously like I'm not criticizing that you made a compelling narrative focused on one scenario. Sure. And of course there are many things you didn't consider and you made the choice not to consider other scenarios. What I'm really interested in is that the scenario creates a sense of inevitability and it's really designed to back up the policy recommendations. And these are massive compute build out loosening of labor protections. I also noticed in there and that raises lots of questions. And I think it's also important to consider other trajectories and whether the same recommendations would hold. So what's Interesting. Like if models commodify, where is the money being made? Most of the money will be made at the compute layer because the models themselves are becoming, maybe not completely, but they'll be much more interchangeable. So if we're then in a scenario where all of the compute layer is essentially owned by US companies and I can explain a bit more about the market logic that without further intervention would lead to a scenario where much of this sort of infrastructure layer is owned by a very small number of US companies, then they will continue what they're already doing, which is rent seeking. And then the question is, what does that mean for the European economy? I think what I also wanted to address is the idea, I think you said in the current competitive market people choose American models and they choose to go to US hyperscalers. And of course that is true. People make the decisions that they make. And I also think people or companies, the public sector, there are some scenarios where you would choose a vendor based on, let's say sovereignty criteria. But in most cases that's not how companies or individual make make decisions. So it's really interesting to look at. So what does the market for AI models and, and cloud actually look like? And that's not me speaking, that's the Competition and Markets authority in the uk, that's the FTC in the United States, pre Trump, that is the European Commission. They've all analyzed this market and come to the conclusion that it is not just concentrated, that we know, we know that it is dominated by a small number of companies, but that at least in the case of cloud providers, that there are is strong evidence that they are also abusing their market position. And what isn't really understood widely and I agree with you that there's like that the interventions that we've seen from Brussels are really falling short. What isn't understood is that you can't simply add more cloud providers to the market. Like cloud isn't an sort of like a good that can be increased, decreased, tripled, interchangeably. But you need very different kinds of cloud services for different uses. And the reason why currently US hyperscalers benefit so much from us is that all the large proprietary models come bundled with the three dominant cloud providers. And that creates an economic condition where European providers don't have an incentive to offer the specific workloads you need for AI inference because there's simply not enough demand. There isn't enough demand for open, open weight models of sufficient scale for it to make a business sense to invest in this capacity. That's just one example. And if we're talking about like a wartime effort to increase compute and to invest so much money in this infrastructure, it's really interesting what isn't mentioned in the recommendation and what I think should be the central recommendation is that Europe needs to much more actively shape this market because this isn't a free and competitive market. It's a market where a very small number of players have such a dominant position that they can abuse their power and that makes it really difficult to compete. For example, should AI companies be allowed to vertically integrate in the way that they are? Do we want the same companies to operate in cloud, to offer models and, and to operate at the application layer? And I don't think that makes any kind of sense. So the more, let's say competitive, the more interoperable, the more open this market is, the better for Europe. That's for me the most important recommendation before any kind of investment and then going back to the scenario, thinking what is the complete queued build out actually for? So in the way that the scenario is written, all this compute build out is so that Europe can run US models on US compute and then Europe can compute in sort of niches, robotics, industrial AI. But the scenario doesn't wrestle with the fact that both Amazon and Apple, lots of companies are also investing in industrial AI and robotics. So I'm not really sure if this is the niche, if this is not an empty niche that Europe can fill. And the other question is these kinds of users don't need the same kind of compute that we're talking about when it comes to sort of frontier models and deploying them at scale. And this is where I think the recommendation gets really, really controversial. We're asking people to invest a wartime effort. To me that sounds like, and I've heard this mentioned in policy rooms, we have to invest people's pensions, we have to mobilize all the public sort of available funds that we have. US hyperscalers won't just come to Europe.
- [37:55]
C
Just to clarify, this is not something that we're saying in the scenario. We're not saying.
- [37:58]
D
I know you're not saying that, but that's how I know. But the question is like, where's the money going to come from? And those are the recommendations that I'm seeing. So if it's not public funds, then US hyperscalers, there's a very tight market for GPUs, why would they come to Europe and not build build compute in the US or somewhere else? Need to give them incentives and we're already seeing in the US that the incentives that are being given are sort of lacks or environmental protection subsidies. So this is a massive, this is a controversial political ask and I'm not really sure what Europe gets out of it. Europe gets to run US models on US infrastructure in the hope that Europe can compete in industrial AI or robotics. And the entire scenario also assumes that there will be no market correction, that the current trajectory of AI that we've been seeing the last few years will continue as it has been before. And I think that's a bit of thin evidence for sort of like the weight and the scale of the recommendations.
- [38:59]
B
I think there's obviously a lot to unpack here, but maybe before we wrap up, I would like you both to answer basically what should be a priority for European policymakers right now if let's say Europe cannot compete across every layer of the AI stack, which layers are strategically non negotiable and what should be the priorities?
- [39:23]
C
Yeah, really very brief reply. So I think it's interesting, I'm actually open to the idea that Europe should take more active steps to shape the market. I think that is not necessarily exclusive to some of the recommendations that we make. I think it's an interesting point. I think one maybe important thing I want to highlight here is that we are not recommending that public funds or people's pensions are being spent on compute. So one thing that we ask for is that more of the capital, the global capital, US capital, including that is already going into compute. More of that should be going on compute on European soil because it will go into compute anyway and it's just better if it does so on European soil than in other places. But it is true that we need supply side reform for that. On the pension funds bit, this is actually something that's fairly, I wouldn't say it's controversial. I think that's something that has been asked for many times. I think the idea here just that we need more of the pension fund money across Europe, that we have to be invested in innovative companies and like scaling capital and venture capital just like the US does, which is I think one of the main reasons why the US economy is like you know, pushing ahead, especially in this tech sector. We're not saying that people's pensions should be spent just on compute. This is just like a sort of venture capital point. And on the question that you mentioned, Ramshay. So I think I'm not sure if I want to make a specific bet or prediction on which layers of the stack we should be competing In I'm in a lot of these discussions all the time in roundtables and I always hear this point of, okay, what is Europe's strength? What should we compete in? Which layers should we focus on? And I always find that a bit confusing because nobody in the US ever talked about this. Nobody in the US ever sat down and said, okay, which layers of the stack will the US now compete in? All of this is private companies, private capital that do things that innovate and they happen to result in a very strong positioning across the stack. And so I think before we make any predictions or bets or say, oh, we definitely need to focus on the whatever industrial AI layer or something like that. I think we just need to implement basic economic reforms that cut across all of these layers. The 28th regime is a good example. I think that's just great. As I mentioned, we need more scale and capital, including maybe from pension funds, maybe all of the big money pots that we have, just all of the basic stuff we need to make it easier for senior engineers in Silicon Valley, for example, but also anywhere to work in Europe. I think Germany recently has introduced to discuss exemptions on labor law for especially well earning employees. If you're a millionaire, you should be able to work in an extremely fast moving startup and not have lots of bureaucracy around being employed. These kinds of basic things, just get the best people, give them a lot of money, let them build. I think this will enable us to complete all across the stack, including the layers that will be important in the future because we as policymakers, or at least I can't predict what that will be, I think the companies are better at that. And so we should make sure that the companies can actually make these bets. And that would be sort of like my first basic recommendation before saying like any specific layer or something like that.
- [42:45]
D
I think it all boils down to do you think of AI as a normal technology, as something we've seen before, or do you think of AI as an exceptional technology? And even if you think of it as an exceptional technology, do you think most or all economic benefit will come from either owning or deploying the frontier? Or do you think there are a lot more options? I think this is what shapes. Depending on these answers, you end up with very, very different policy responses. I'm more in the camp of AI as a normal technology, as in, of course AI is useful for many interesting tasks, but it also isn't useful for anything and everything and under any conditions. And diffusion may take longer than we think. So this is just where I personally fall in. But regardless, I'm also with Maximilian, as in policymakers should make focus on things that are useful, regardless of how the trajectory continues and how the market continues? And I think this is where the sort of like the massive compute build out is a speculative bent. This isn't the same as investing European pensions in startups and companies. This is essentially using European pensions as high risk venture capital. You can do that. But it's a really controversial proposal that needs more than sort of like managing people's opinion. This would need massive support from people because if the bet doesn't turn out, I'm really concerned that we'll experience a crisis of political legitimacy. People are already not really excited about AI and data centers. If this is being pushed down their throat and the bet doesn't turn out, yeah, democratic institutions will suffer. So what are the things that policymakers should be doing in a condition in which the market on the one end is heavily concentrated, but there also is a degree of technical uncertainty and also economic uncertainty? And we're still writing this final piece in which we'll lay this out, but we're thinking along the way along two ways. So one is, I think we need real scenario planning in policymaking and we should really take a close look on some of the trajectories that are much more beneficial to Europe than others. Right. So this sort of like the one company owns the frontier, the frontier is decisive, nobody can catch up. That's devastating for Europe. The commodification thesis is much more beneficial. It leaves much more room for leverage. So Europe should do everything in their power to create the conditions under which it would benefit from a commodification. Because even if there are sort of more models, if people use smaller models, if people use open weight models, you need to create conditions in which people can seamlessly switch, in which the market is truly operable, in which the market is contestable, and in which the market is open. And currently the market isn't that. So creating the condition for truly competitive markets is one of the most important things that Europe can do. And I think finally, like in all this thinking, there is a risk in treating AI as an exceptional technology because it means surrendering lots of other policy goals that we've already sort of democratically agreed on. And the goal should be like long term economic prosperity from which protects the environment, which benefits workers, which benefits people. Of course AI plays a role in it, but the question is really how big of a role should it play? Should everything be about AI or should AI be seen as sort of one technology that Europe invest in or like. Or should it be seen as sort of or should the main goal be to create really good conditions for a prosperous innovation ecosystem where all kinds of ideas can prosper?
- [46:31]
B
Thanks so much Frederic and Max for your conversation about that gets very that gets at something real about where Europe stands. That's it for today. There's obviously a lot more to discuss and we hope to continue the conversation through contributions and the space. Thanks to Europe 2031 and Frederick's series on Europe's AI market with an AI now institute and Tech Policy Press are in the show notes. So those of you who haven't read these pieces, I urge you to.
- [47:11]
A
That's it for this episode. I hope you'll send your feedback. You can write to me at JustInEchPolicy Press. Thanks to Ramsha, thanks to my co founder Brian Jones and thank you for listening.
- [47:29]
D
Tech policy press.