Global Affairs Expert Webinar: AI and Geopolitics

Sarah Kreps, the John L. Wetherill professor in the department of government, adjunct professor of law, and director of the Tech Policy Institute at Cornell University, leads the conversation on AI and geopolitics.

These webinars provide an opportunity for college and university educators and students to discuss global issues with CFR fellows, Foreign Affairs authors, and other leading experts. To register for future invitations, please complete this form or email [email protected] with your name, title, and academic affiliation. 

Speaker
Sarah Kreps
John L. Wetherill Professor in the Department of Government; Adjunct Professor of Law; and Director, Tech Policy Institute 
Cornell University

Presider
Irina A. Faskianos
Vice President, National Program and Outreach
Council on Foreign Relations

Transcript

FASKIANOS: Thank you. Welcome to today’s session of the Fall 2025 Global Affairs Expert Webinar series. I’m Irina Faskianos, vice president of the National Program and Outreach here at CFR.

Today’s discussion is on the record, and the video and transcript will be available on education.CFR.org. If you would like to share these resources with your colleagues or classmates, you can do so. We will circulate after the fact. As always, CFR takes no institutional positions on matters of policy. 

We are delighted to have Sarah Kreps with us today to talk about AI and geopolitics, a very big subject. Dr. Kreps is the John L. Wetherill professor of government, adjunct professor of law, and the director of Tech Policy Institute in the Cornell Brooks School of Public Policy. She is most recently the coauthor of the thirteenth edition of International Relations, published by Pearson earlier this year, in addition to seven other books about technology, international politics, and national security. Dr. Kreps is a nonresident senior fellow at the Bookings Institution and a senior fellow at the Bitcoin Policy Institute. She served as Stanton nuclear security fellow here at CFR. She’s a life member. And she also served as an active-duty officer in the U.S. Air Force.

So, Sarah, thanks very much for being with us for this discussion. I thought you could set the stage and talk about the ways that you see artificial intelligence currently shaping global geopolitical dynamics, and what you’re thinking about, and worrying about, and see as exciting, and opportunities that we should seize?

KREPS: Yeah. Thank you so much, Irina. It’s great to be here. This is such an expansive and dynamic topic, AI and geopolitics. So what I will do, rather than trying to cover everything, is touch on a few areas that I’ve seen evolve firsthand. And by way of background, my perspective on this starts with my time, as you mentioned, in uniform. So before I was an academic, I was a military officer. And I worked with emerging technologies that were just beginning to change how militaries thought about information and command and control. And that experience—those experiences taught me the way in which innovation in defense is never just the hardware. It’s about trust. It’s about doctrine. It’s about the way humans interact with that technology, and how that connects to decision making. 

And so when I moved into academic research, I carried that perspective with me. And back in 2018 I became one of the early academic collaborators with OpenAI. And they—we were at a think tank in DC and a couple of people from OpenAI were giving a talk on this thing called GPT-2. And this—they weren’t talking about national security, but I just, in my mind, connected dots and thought, if you can generate a lot of authentic-seeming information, there would be a real risk for AI to do everything from flood elections with synthetic propaganda to—you know, my mind spun with all kinds of other things as well. 

And so six years later, or I guess seven, electoral misinformation hasn’t disappeared but now it seems small compared to today’s challenges. So, you know, AI arms races among states and firms, autonomous systems on the battlefield, deepening global divides in compute and infrastructure. So I think the questions have moved from whether AI can generate fake news to how AI is reshaping geopolitics. And so to walk through that, I think we have to start with infrastructure and access. And it’s, you know, chips—you probably hear these—you know, chips, Nvidia, clusters, datacenters. But they’re all really important and foundational to who leads and who lags in this both state-level and firm-level arms race. So I’m going to start there, on the kind of infrastructure and access. 

So you might have heard about graphic processing units. And GPUs are what—they specialize in, matrix operations and run thousands of lanes at once. So back in the 1990s and I guess early 2000s, Nvidia was producing these GPUs that were used in the gaming industry. So it was a successful company, but nothing like it is today. Because what happened then was these GPUs that were very useful in that kind of graphic space, because you needed a lot of compute power and thousands of lanes at once, that’s exactly also what you need for artificial intelligence and machine learning. 

And so the reason why this is an infrastructure story is because you need to—you need power, you need cooling, you need permitting. And chips are foundational to all of that, and then you have these electricity needs every month after. So what that means is—and what makes this very interesting and not just about geopolitics—is now you have multiple layers of governance and private and public actors. So you have utilities. You have municipalities. You have local politics, state politics. And then you have this whole global picture too. And so frontier AI depends on scarce physical inputs, like GPUs. And so OpenAI is running around forging deals with Nvidia and AMD, these chip companies. 

And what they’re—and it turns out that this creates a compute divide. So it’s a hierarchy that is based on access to chip supply and access to being able to supply those at scale. And so it’s both at that international level, but then within societies. And, again, this is just—I’m trying to do a quick survey—but you can see how within societies this same dynamic or pattern with—a divide could hold. So access to advanced models, access to training data, the skill to steer them. So maybe you’ve heard of prompt engineering and fine tuning. So how you interact with that model will determine—and you’ve probably heard input versus output, so garbage in, garbage out. But conversely, if you can interact meaningfully with the model, you’re better able to get meaningful outputs. And so it turns out that a small number of firms and institutions and individuals can afford to have this access at all these levels. 

And so this gap, or this divide is about—is shaping who gets to create knowledge, who gets to consume it. And so in education those with compute and expertise can build the models that define the terms of the market and the conversation. So this AI divide then is hardware and energy, who can host and power the clusters, but also the knowledge and interface. So who’s understanding, adapting, and directing these models. And so there’s real agency in this. It’s not just about the cables and cooling, but it’s about who’s controlling the inputs, who’s understanding these models. But also, what I’m seeing in my local jurisdiction as they’re talking about a datacenter, but it’s about understanding these multiple levels of responsibility and jurisdiction. 

So with that picture in mind, the arms race makes a lot more sense. So I referenced both a state-level and firm-level arms race. And so—and usually that conversation takes place in the context of the U.S. and China. And that has to do with what I talked about just now, which is who can afford hundreds of billions of dollars in compute and resources? And it generally falls into those two camps, the U.S. and China. And so both of them are pouring resources into AI. And both have viewed this in various strategy documents as a strategic technology or vehicle of power. And so they have both articulated a vision of leadership, global leadership, as critical to economic growth, and military advantage, and technological prestige.

So important in that is the control of inputs. So the talent, the training data, the semiconductors. And so this is why U.S. policy in the last few years—and this cuts across two different administrations—but has been the set of export controls that are reflecting the importance of the certain technologies in this whole process. So Taiwan, for example, produces the most sophisticated chips. And they do it—have generally done it in Taiwan. And so there was—has been a question of resilience, about whether—about a Chinese invasion of the island. And so these become real chokepoints. But so do—and this has come up in the last couple of days—the Chinese have a monopoly of what are called rare earths. And so those are chokepoints. And so we—as the U.S. engages in export controls on the advanced GPUs to China, China then in return is restricting gallium and germanium exports. 

So that’s the state level, but there’s also a firm-level race. So it’s—you know, you hear about these tech companies. They’re locked in a corporate competition—OpenAI, Google, Anthropic, Meta. So they’re all chasing these models. They’re competing for scarce chips and concentrating the compute power in a handful of platforms and providers. And so that same dynamic is now spilling into civic life. And here I’ll get to AI and elections, which is where I started. And I don’t want to downplay this, but I do want to say that my views on this have changed a little bit. And I think one of the articles that was included in the materials for today was a somewhat alarmist piece in the Journal of Democracy. 

And my concern, stemming out of the 2016 election, mapped on to these new technologies, was, wow, you could—a nefarious actor could generate huge amounts of AI-powered manipulation—so, hyper-personalized, microtargeting, deepfake candidates, synthetic voices. And I thought that that might—that certainly, in theory, had the potential to manipulate the public—the mass public. And so—I and—and I think the worry is still there, but for a different reason. I think what we’ve seen is, frankly, a flood of AI generated slop—so cheap spam, low-quality content. And I think the information is just overwhelming people rather than persuading people. 

But that also has a downside, which is a collapse of trust. If nothing—if people just assume nothing’s authentic, then people just might not believe anything at all. And the insight that I—that I really stand by in that piece is that democratic stability depends not only on truth, but a baseline of shared reality. And so if AI is challenging that baseline by making it easy to doubt every photo, video, or voice recording, then that really puts us on shaky grounds. 

Let me turn for my last kind of substantive point here to a different geopolitical space, which is the battlefield. And so this is obviously a really important one. So militaries are increasingly integrating AI into everything from logistics to surveillance to decision support and targeting. And so there have been big debates on what this means and which countries are using these systems, how they’re using it, and what it means for decision cycles and escalation and accountability. And we’ve seen this in both—in all—in a few different contexts recently. So Ukraine has seen AI-enabled target identification and autonomous drones. We’ve seen systems—decision support systems being used in the Gaza conflict. And the U.S. and China are developing AI command systems, but they have different attitudes toward human oversight. 

And so I think the open question here is whether these institutions are building safeguards, or how these speed and efficiency pressures are shaping or undermining human control. Or really, and this is a project I’m working on right now, how different types of individuals interact with these systems. And so when I was interviewing NATO officers this summer—and they all are using AI in totally different ways. And they—I mean, and it’s so multifaceted. But it’s different countries that are overriding outputs. 

You know, so it gives this—its AI system a prompt, it gets an output. I had five different people tell me they were relying on different confidence levels to make a decision. Some were overriding at this threshold. Some were overriding at this threshold. And so you get—what’s central in the military, which is unity of command, you no longer really have that because everyone’s using these systems in different ways. And so I think there are a lot of open questions in that battlefield space.

So just to wrap up the initial comments here, if we’re putting these threads together, it’s, I think, a really complicated tapestry of international, national, even local-level dynamics. We have private and public. We have states and firms. They’re all competing for compute and chips and market share. And then at the societal level, there are questions about how AI is straining the integrity of elections, the questions about the labor markets and how it’s—the impact of AI on unemployment. And then on the battlefield, there are these open questions about how autonomy is promising speed and precision but might risk escalation and the lack of accountability. And then at this both global and society level, there’s the compute gap and divides in compute and infrastructure in ways that can entrench inequality, both across states but even within states. 

So I’ll just leave it there. And I look forward—I hope I teed up enough controversy and questions that we can have a fruitful conversation. 

FASKIANOS: Thank you, Sarah. That was great. And really great context for us to have this conversation. 

(Gives queuing instructions.)

So I’m going to go first to Dan Drezner, who’s at the Fletcher School at Tufts University.

KREPS: Hi, Dan.

FASKIANOS: Dan, do you need to unmute yourself? 

Q: There we go. 

FASKIANOS: There we go. 

Q: Sorry. I didn’t get the instruction to unmute. Professor Kreps, thank you very much. Appreciate it. Very interesting overview.

My question is about whether or not we are in at least a temporary AI bubble in terms of the amount of investment being, you know, plowed into this technology, without it necessarily leading to an immediate payoff. And if we are in a bubble at least in the short term—and I’m not trying to deny the utility of it as a general purpose technology in the long term—how do you think that will shake out in terms of the sort of short term economy? And will it lead to an even stronger consolidation, as a whole bunch of firms wind up having to go out of business?

KREPS: Mmm hmm. (Laughs.) Yeah, Dan, great question. And, as you know from our previous conversations, I lean optimistic on a lot of fronts. But I do worry. I share your concern about whether we’re in a bubble. And I guess for that question, it really depends on how we think of—what we think a bubble means. So if we think of it as the valuation of firms, or GPU prices, or investment flows, I think they certainly look bubble-like. These VCs and corporate spending is—seems a little bit—and we’re old enough I feel like we can use these expressions, like irrational exuberance. (Laughs.) And there is not yet a revenue model that seems to justify this amount of spending. And so I share that—I share that worry. 

And I guess the thing is that we don’t—and I think we’re starting to see people question. And I saw Nate Silver comment on this the other day online. Well, I’m not—really not getting productivity gains from this. And I think it’s—or seeing. I think it’s uncertain what those durable advantages of this will be. I do think that some of what’s being funded, where there’s not even a clear revenue model, would give me reason for concern. So then the question is, I think this is right—and this might not be the worst thing—is the consolidation, which I think, for the reasons I mentioned earlier, are likely to happen anyway, which is all of these—this is such a revenue-intensive industry that it makes it difficult for—it makes it difficult for small startups to be part of it. 

Which I think will lead to a consolidation and kind of a merger dynamic, in which these smaller startups eventually get bought out by the bigger ones just because they’re the ones that have the capacity and the chips to be able to continue. So I’m not one—I don’t feel like I could speculate on whether it’s a bubble or not, but I do think that the consolidation in this industry is probably overdetermined. And I think we will see a lot of startups that go out of business because there wasn’t much that they were standing on.

FASKIANOS: Thank you.

I’m going to take a written question from Sarah Tenney Sharman, who’s an associate professor at The Citadel: Could you comment on the grid capacity differences between China and the U.S.? What impact could this have?

KREPS: So, yeah. I mean, I what I’ve been following is more the U.S. I guess my sense of what is going on in China is a bit more limited. But it seems, from everything I read, that it’s pretty capacious. And I do think that this is the correct question to be asking, which is what is the grid capacity? Because I think it’s something that—you know, that there’s a bit of a lag effect in this space. That there was so much focus on AI and getting the models down. And then, OK, for the models we need the GPUs. Oh, we need to go and hoover up all the GPUs. And then it’s, like, well, how do you turn on those GPUs? Well, you need power to do it. And so they don’t—China has a national grid planning model, because they can do this in a large-scale way which makes it fairly effective. 

I think the U.S. grid is considerably more fragmented because it’s—and less centrally planned compared to China. So there’s a lack of concerted planning because we’re such—this fragmented federal system. And so I think there is a lot—I think that’s a bigger bottleneck in the U.S. than it is China. So they’re deploying this new infrastructure, but there are big regulatory and permitting delays. And I was telling Irina before we got on the call that we have one right here on our lake in Ithaca. And there’s a board meeting today in the local village to discuss this high-performance datacenter. And there are so many concerns in the local area about what this means environmentally, what it means for energy prices. 

And it’s interesting to see the company’s perspective on this. They clearly were behind the eight ball on this because there was a—and this is like me as a political scientist talking and being intrigued by this. That there was a public comment period that opened up, I guess it was in August. And there were seven pages of comments. Not a single one of them supported the datacenter. And this, I should say, is a decommissioned coal plant on the lake. So if you’re in the middle of the lake you see this complete eyesore that has been decommissioned. And what this energy company—or this, you know, TeraWulf is trying to do is just resuscitate it as a datacenter. And they’re getting so much pushback. But they were completely behind in trying to do the kind of local level ground game of understanding what these local jurisdiction dynamics were. 

And I think that speaks to the question, which is that if you can just brute force override local opposition, like you can in China, energy is not going to be the limiting reagent the way it is here, where you have a town board of five people who don’t understand, you know, technology, chips, and longer-term consequences. And I don’t—I’m agnostic on whether it’s a good idea or not, but they don’t know—it’s a high-level question that they’re not trained to be versed in. But it’s representative. It’s emblematic of what’s happening around the country, where you have local town boards that are weighing in on questions that have national and international consequence, because, again, without those datacenters, you can’t do that AI leadership that the United States says it wants to do.

FASKIANOS: Thank you.

I’m going to take the next question from Chad DeWaard, who’s at Culver-Stockton College.

Q: Hi.

My question was that, with advanced—like, China, they put out—they put out DeepSeek last year. And it’s a much more—it’s a highly advanced AI model, but it was built, to their to their claim, extremely cheap compared to a model like GPT-4 or -5. Do you think that as technology and just code advances, and they make models cheaper, that it will lessen the need for this sort of geopolitical, I guess, grab of resources?

KREPS: Yeah. I mean, I think that’s a great question. And I think one thing to note is—and this sort of speaks, again, to my earlier comment about when you’re sort of centrally planned, like China, another feature is that we don’t really have full transparency over what’s going on behind the scenes. And so but I think it’s fair to say that DeepSeek was able to achieve pretty high performance at—even if it wasn’t six million, which they had said—I think a fraction of the training cost of U.S. models like GPT-5. And I think what it shows is that there is a way in which scaling laws deliver smaller and smaller gains per dollar kind of at a certain point. Kind of the marginal improvement in the accuracy or the reasoning would require—or they require exponentially more compute. 

And so I think, in a way, that gives—that gave them a clear second mover advantage, because they could optimize based on the experiences of the U.S., and, you know, some of these open-source models, and just understanding the technology that the U.S. had invested—or, U.S. firms had invested huge amounts of money into. So I think it does suggest kind of a—both a diminishing marginal return, but that it is lowering, in some ways, the barriers to entry, because that means that potentially states and firms outside the U.S. and China could field competitive systems. And that could, I think, democratize some of the field. I think it still will depend on scarce resources, though. You know, which is—like, it’s clear. And that’s why I think OpenAI is trying to buy chips not just from Nvidia, but now AMD, which is—you know, these—that seems to be, again, a limiting reagent, as is the—as are the datacenters themselves. 

But I do think if the U.S.—if the Chinese teams can match U.S. model quality—and this isn’t exactly what you asked about—but with smaller and older chips, then it does raise questions about some of these export controls and whether—and I think that’s maybe in part what we were starting to see unfold with the U.S. and allowing the release of some of these chips to China. Which is, maybe these hardware sanctions weren’t really effective anyway because this is more about engineering skill and model efficiency, which is a lot harder to regulate than it is the hardware itself.

FASKIANOS: Thank you.

I’m going to take the next question from Stephanie Forbes, who’s a graduate student at the University of Maine: Could you expand a little bit on how the AI race is expanding to regions like the Global South, and how that’s something that’s contributing to the compute divide you mentioned?

KREPS: Yeah. You know, and it’s not even the Global South, but it’s something that I’ve seen in Europe as well. Which is that, especially in light of the European Union’s AI Act where it was a much more, kind of, precautionary policy of being more risk averse, I think there’s a sense at least within part of the EU that maybe that was too aggressive too early. And that this—kind of the genie might have been out of the bottle, and maybe instead of trying to regulate it, you know, they push talent and resources out of the EU. I think the Global South is a slightly different case, but a very interesting one as well. I was just reading about Argentina and a $25 billion datacenter. And so, you know, we’ve seen some of the—and I think Indonesia as well. 

So it there’s a mobility of capital that I think we saw with globalization in certain industries over the last thirty years, but I think as this has become a much more kind of globalized industry it’s natural to see that talent and resources will flow to places where it’s more economical or, frankly, probably where the—where these projects can be strong-armed. I don’t think that Indonesia and Argentina will have the local jurisdiction resistance to datacenters the way I’m seeing them in my town here. And so I do think that there’s an interesting story there, but I do—you know, in terms of where projects and capital will flow. 

But I think in the shorter term, it is—it does seem to be the case that AI is an accelerant in the hands of those who are using it. You know, and I think the vote is still out on that. And Dan, I’m sure, has thought—Dan Drezner has thought more about this, I’m sure. But, you know, it’s sort of surprising that, in light—or, some people have observed what’s happening in the U.S. economy of, like, all time, you know, stock market highs amidst a tariff war, and speculated that this is just because the AI innovation engine is so much on fire. And so if that’s the case, then it is suggestive evidence that there is that growing divide between those who have access to that technology and those who don’t. 

So, yeah. So I think it is an interesting question about the Global South and whether they can, or want to, jump on that train. I mean, it seems like there was a sense in Europe that maybe that’s not a train they wanted to be on. And one of the things I say in a book that’s coming out next year is that there’s no right or wrong choice on this, but there should be an awareness of what the tradeoffs are. And so just to contrast the U.S. and Europe in this case, you know, I think Europe said: We don’t know what the risks are, but we’re more wary of the downsides than we are lured by the upsides. And the U.S. seems to be leaning into the upsides. And so I think where the Global South can—they can learn and decide what their model is, and where they’re leaning on, kind of, the risk versus opportunity continuum.

FASKIANOS: Thank you.

I’m going to go next to Ryan Kostanecki, who’s at Macomb Community College in Michigan.

Q: Hi. Thank you for this talk today. And I’m sitting in a class with my students, so it just happened to fall at a good time. 

And I asked my students if they had any questions. And I had a student ask a question, which you kind of went into but if you could expand on a little more, about just kind of resource management and global inequality. Like, we’re told that AI is supposed to, you know, democratize and, you know, like, we hear from the CEOs all the time that it’s going to spread wealth to many people and be this kind of panacea. But even your point earlier about, you know, trying to convince people in your community that it might be in their interest to build a datacenter, or, you know, to keep America competitive, but what does that mean to the person that’s sitting across the lake and considering, you know, reactivation of, like, a coal plant in their community, right? So how does—how do you think about these things in terms of, you know, the global inequality, resource management, and things like that, from a local level all the way to a geopolitical level?

KREPS: Yeah. And I don’t want to be—I don’t mean to have conveyed that I was trying to advocate one way or the other on this datacenter. I was kind of asked to think about an AI innovation hub that could do workforce development for the region to support this, but then I met with board members. And to your point, these board members just do not trust these corporations. And so in one of the town halls that was held, this industry member from this company that would be doing the datacenter, said, oh—because someone said, well, how do we know that this isn’t a fly by night or a bubble, and in five years this—there won’t be any demand from this, and now we’re going to have this thing just sitting there? And the person from the company said, oh, well, we have—one of our investors is Google. And so Google is a long-standing company. So we’re pretty confident. 

But what was really interesting about that statement, which was meant to legitimize this whole thing, I think actually just completely does the opposite. And there’s sort of this lack of understanding about local politics and their concerns, to think that, you know, citing Google is going to make people feel better about what you’re doing. And I think what that gets at is just this—you know, I think that this is a difficult-to-understand technology from cradle to grave. I mean, so I think people might associate the technology, oh, this is an old coal fired plant. And the real devil is in the details. And one of the people I was talking to on this project, this environmental energy engineer. I said, so what do you think about this datacenter? And she said, well, there are right ways to do it and wrong ways to do it. She said she was wary of, you know, just having a geologist within this industry, because the worry about whether they’re just going to rubber-stamp things. But there are environmentally sound ways to do it. 

And so I think those are the—and I guess the concern would be that if you are moving too quickly, maybe these concerns haven’t either surfaced or been addressed. And I think that probably—that seems to be at the root of these more local debates. And where you could see, as I said earlier, if you’re in the Global South, and you are in a community that just feels like it needs the revenues, are they even weighing the cost? And so I think there you could see, much as we’ve seen with other, you know, mining projects in the past, where they might be exploitative, because this, you know, multinational can come in and just kind of wave $20 billion in front of a not very affluent community and say, hey, we’re going to come in and improve things. But maybe they don’t have the layers of scrutiny that we’re—that, you know, you can see at the local level in the U.S. 

So I think that the divide then cuts across many different dimensions of this, whether it’s, you know, again, that kind of opportunity and risk. So—and I think part of the problem is there’s so much uncertainty on both of those. Are we more likely to see that there was, in retrospect ten years from now, that this was a massive opportunity? Or are we more likely to see, wow, that was a really big risk, and we jumped right in? But I think that my takeaway is always that it’s important to be framing the questions, because if you don’t frame the questions you can’t possibly understand the answers.

FASKIANOS: Thank you.

I’m going to take the next question, a written one, from Niks Andžāns, who’s a graduate student at the University of Alabama: Are democracies inherently at a disadvantage in AI development because their commitment to ethical standards and regulatory oversight slows innovation compared to more authoritarian regimes?

KREPS: Hmm. Well, good question. I think—so, you know, I think there’s certainly something on the—as I said before—the centrally planned economy front, which is that you can—the Chinese—the CCP can say, go do this. We’re giving you billions of dollars. You go do this. Do it quickly. But I see more risk—because I actually think that in the end—and we’re—you know, I want to separate democracy from a capitalist economic system. Because I think capitalist systems are very competitive, and they’re very creative, and innovative. And so where I would see the bigger divide is actually on AI battlefield questions, where if we think there’s an advantage to AI on the battlefield being able to do things more quickly. I think the ethical considerations there are more likely to hold back democracies with the rule of law than they are the nondemocracies. 

We were actually just talking about that in my PhD seminar earlier, the question of whether, you know, regime type and battlefield behavior. And so I—and there tends to be a correlation between democracies and battlefield ethics. Not perfect correlation. But I think, by extension, we’d be likely to see that in terms of rapid AI adoption by non-democracies than in a democratic system.

FASKIANOS: Thank you.

I’m going to go to Ted Alden, Western Washington University.

Q: Thanks very much, Irina. And thanks for an excellent presentation here. 

I’m sort of torn between two questions, so I’ll ask them both quickly. And you can—you can pick which one you prefer. One is, my understanding is the Chinese are using more open-source models for their development of AI, as opposed to the more closed shop approach here with OpenAI and the other competitors. I wonder if you could comment on the—on the relative value of each of those. Actually, I’ll just leave it at that question. I had another question but it’s come out in some of the other responses. So if you could respond on the sort of open versus closed model for developing these AI systems, and the advantages and disadvantages of each. Thanks very much.

KREPS: Yeah. That is a great question. Yeah, so I think one of the reasons Chinese firms have been more reliant on open-source models is that they have faced a hard compute ceiling because of U.S. export controls. So if you open source, and that means that, you know, you put your code in a repo, and so everyone can access it, teams can start from weights that are already trained and then fine tune them on local data, with—and what that means is it requires far fewer GPUs, which the Chinese don’t have access to. So that allows these Chinese groups to be reaching higher performance while still staying within its domestic chip limits. So then the open source is a way to basically be a force multiplier, given the limits to their hardware. 

And so within that then, you know, you have this thriving internal ecosystem around these open weights. And so these tech giants like Alibaba, Huawei are publishing at least partially open models to attract developers. So but, you know, what that—the contrast in there—the reason why that can work there is that, you know, again, if there is a way in which the sort of centrally planned innovation model there—I’m not saying profit motives aren’t a factor, but I think the challenge in the U.S. space is that so much money has gone into—invested into these models, these frontier models. And there is a kind of winner-take-all feeling, that there’s no—not only no incentive to be as open, but on the contrary more incentive to be closed about it. So, yeah, that’s what I would chalk it up to, just as—you know, just the way in which the limits to the access on hardware almost forces them into that approach.

FASKIANOS: Thank you.

I’m going to go next to a raised hand from Walton Brown-Foster who’s at Central Connecticut State University. If you can accept the unmute.

Q: Thank you.

FASKIANOS: OK. We can hear you, I think.

Q: Oh, I’m sorry. That was accidental. I am driving. I do not have a question.

FASKIANOS: OK. No problem. I will go next then to Annelise Riles, who’s at Northwestern University. 

Q: Hi. Thanks so much for that, Sarah.

I have a question about the battlefield piece. So I’m interested to know what your take is on the prospects for success of any of the current attempts at arms control in this space, treaty or otherwise. And related to that, I was interested in your comment about the fact that the Americans and Chinese bring in human decision making at different points or in different ways in the battlefield context. And I heard Herb Lin give the Blair Lecture here at Princeton last week, in which he relayed that the Americans and the Russians have such a different conception of what AI is to begin with that it’s very difficult to even have a dialogue about arms control. And that we need almost a comparative study of concepts of AI before we can even get that conversation going. So interested in your thoughts about what—to say more about what the difference is between the Chinese and American perspective. Thank you.

KREPS: Thanks, Annelise. Good to hear you. 

So, yeah. I am not optimistic on any verifiable agreement, let alone a treaty, that would constrain battlefield AI, at least in the near term. And, you know, this actually is part of my forthcoming book as well—(laughs)—on disruptive technologies. Is that, you know, looking back at nuclear it took from 1945 until 1963—so almost twenty years and a Cuban Missile Crisis—to mobilize people to the view that people like Einstein and other arms control advocates had been pushing for, for almost two decades. And I think there are a couple of challenges that make this even less likely, more difficult, at least in the short term. 

One is that I think there is disagreement, as you had indicated—has been articulated by people like Herb Lin—that people disagree—that countries disagree on the scope of the problem, or the problem itself. And, you know, is this—and sort of touching on this, but the difference between the EU approach to regulation and the U.S. The EU is much more concerned and the U.S. is leaning into this. And so it suggests that some countries are treating this as not a taboo at all. And so finding common language is really difficult in that regard. 

I think a second real problem with this is the dual use problem. And that connects in two ways. Not only—I mean, I think a big part of this is that these technologies are coming from a civilian space, whereas there was no civilian use of an ICBM during the Cold War. How do you regulate and draw a clean legal line between permissible and prohibited systems for something like ChatGPT? Which is something that, you know, people—like, military people going out to a battlefield could just plausibly take their iPhone and use, you know, this system on their phone. So how do you regulate in that space? 

And then I guess another problem is that there’s rapid, rapid technological change. And so what we saw in the Ukraine war is just an accelerated development and feel the feeling of drone and autonomy combined. And so it really creates a strong operational incentive to keep pushing those capabilities. And then sort of lastly, and maybe related to the dual use problem, even if you could agree on definitions, how do you verify compliance? And so there was an example from the Libya context a few years ago, where the UN did an investigation and they saw that a missile had been fired from an autonomy-capable drone, but they couldn’t tell whether the drone was in autonomy mode at the time. 

So how does that get parsed? Is autonomy a matter of the software, the human procedures, the mission authorization? Or is this a separate hardware signature? So the inspection and monitoring of that sort of thing is then further complicated. So, yeah, not very optimistic about whether there’s at least a near-term kind of global agreement on the horizon.

FASKIANOS: Thank you.

I’m going to take the next question from Gustavo Oliveira, who is an assistant professor at Clark University Graduate School of Geography: Are leading actors in the military, states, and companies in U.S., China, and EU concerned that environmental disruptions or increased environmental regulations might limit AI development—e.g., restrictions on energy use, purpose, and destination of rare earth mineral and refining, et cetera?

KREPS: Yeah, great question. And I think it gets at these multilayer considerations. So, I don’t think until now that has been a limiting factor. And if I understand the question right about these environmental issues intersecting with AI, I think the biggest ones that come to mind are the ones that we’ve touched on already with respect to energy and infrastructure and sustainabilities. And so I do think that that electricity supply is going to be the most binding constraint on expanded U.S. computational capacity and AI dominance. And so far, again, there’s no—there can be almost no national strategy—energy strategy on this, because everything is decentralized. 

So when I was over in—I was in Buffalo a few weeks ago. (Laughs.) And their infrastructure company there is called National Grid. I had never heard of National Grid. We have NYSEG here. If you live in DC you have a totally different one. If you live in Boston, you have a different one. So all of these then are, again, completely decentralized. And given the energy demands of these large AI models, there are certainly going to be environmental impacts. And the question then I think is how—is that tension, and how that’s managed. The high density compute demands straining the cooling and power and transmission with this very decentralized energy grid we have, and the national layer motivation to lead in this space. 

So I think—and this is not a normative view—I just think the reality suggests that these environmental considerations may be a constraint on all of those other things. And, you know, what an optimistic view would suggest, coming back to this local question that I’ve been involved with, is that that could push these companies to clarify how they are engaging in this energy space in ways that are not, you know, having an adverse effect on environmental considerations. I mean, I do think at the least there is—you can’t do it in a carbon neutral way. And so, you know, on the other hand, I guess—(laughs)—if you’re not riding a bike to work then, you know, already it’s a continuum.

FASKIANOS: Gustavo, I did not see your hand up. Is there anything you want to add to your question?

Q: Thank you, if I may. There is a follow up. 

So it looks like—and Sarah’s conclusion reinforces—that national security strategies for AI dominance are at complete odds with national security strategies around climate change mitigation and adaptation, which also have, of course, national security considerations. I wonder if—are these conversations happening just in separate policy spaces or separate branches of the military? I can’t imagine that highly advanced militaries like the U.S., China, and EU don’t recognize the tensions for national security of runaway climate change and AI dominance predicated upon doubling down rather than, you know, finding alternatives to the irrational exuberance that Sarah mentioned of the way that this technology is being developed right now energy wise.

KREPS: Yeah. I think that’s a great observation. And I think part of that stems from—I think there are a lot of causes there. But I do think there tends to be kind of a siloed approach to these problems. You know, you have the climate people concerned about climate change. And people who are doing—who are concerned with AI. And often they don’t—you know, a lot of national military strategies treat—still separate AI national security from climate and sustainability. And I do think that those two things are in tension with each other. And I think it’s a question of, you know, how do—you know, how does—how do those get mediated? 

And my prediction is that—or, what I see happening generally, but I think with the constraints that you’re seeing—starting to see crop up at the local level. And you see so many articles. There was something in the Wall Street Journal about the xAI datacenter in Memphis causing these kind of local reverberations. So I think that so far, the AI arms race at the firm level has pushed and accelerated so quickly that it was almost, like, this question of datacenters hadn’t even cropped up, and the implications of it. And I think that’s starting to catch up in ways that are tilting the scales a little bit, just because the sheer energy needs. I mean these AI models are just, you know, eating energy. And not only is the grid not equipped to deal with it, there is then also this big climate consideration that I think has not received as much attention because of all of the kind of framing of this being in this national and firm-level arms race. 

FASKIANOS: Thank you.

I’m going to go next to David Johnston, who is at the Oregon Institute of Technology.

Q: Good afternoon. Thank you so much for this extremely fascinating discussion. 

I just want to let you know that I’m a physicist. So this is—geopolitics is not my background. But I’ve been very interested in AI since it came out—you know, or since ChatGPT came out back in November of 2022. My question is kind of around, actually, what you mentioned, Sarah, about the—I don’t really want to say parallels, but the, you know, regulation around nuclear power and nuclear arms. And I guess I’m concerned that it doesn’t seem, anyway, that the—that the government is really trying to constrain or challenge the dominance of these extremely large companies, and just letting them kind of run wild. Do you know if there are any discussions about, like, you know, how can we incentivize other groups to—I’ve heard a lot of talk about small language models and how they might possibly be the future. And they wouldn’t require these immense amount of resources. So I’m kind of curious what you hear about on that front.

KREPS: Yeah. I think it comes back to my skepticism about whether AI governance and regulation will take off. And it’s because in a nuclear context, which you mentioned, essentially, the government owned the means of production. You know, Lockheed Martin making missiles, right. So it can control that. But the government’s not owning the labs that are developing the frontier models. You know, my understanding is that—and, actually, you know, one of the things that had come up a couple of years ago when there was a big AI summit in England, at Bletchley Park, was that there—the government was in conversations with OpenAI, not—that these were considered national security attributes. And that—not that OpenAI had any—everything’s so proprietary. But that those could not be—those model weights could not be released because this was now considered a national security asset. 

But generally, you know, the government doesn’t really control these private firms. And in fact, they’re relying on these tech firms for the capabilities that they need. And so I think in that sense, instead of being adversarial or highly regulatory, I think they’ve been really almost, like, a partner, because they haven’t—you know, limiting that corporate dominance would mean weakening its own technological base. And it doesn’t have any incentive. And in fact, it has incentives not to engage in that. So I think that explains that—kind of the light touch to regulation for those two reasons, both that it’s its dominance depends on that, but also it has less control over that means of production.

FASKIANOS: We are at the end of our time. We have so—an overwhelming number of raised hands and written questions. I’m so sorry that we could not get to them all. But Sarah Kreps, thank you very much for covering such a large amount in a in a little square of time—postage stamp of time. So we really appreciate it, for you being with us. And we will circulate the transcript and video after this so people can review it. And, again, my apologies that we couldn’t get to you. But thank you for all your good questions. Really much appreciated.

Our next Global Affairs Expert Webinar will be on Wednesday, October 22 at 1:00 p.m. (EST) with Jeremi Suri, who’s at the University of Texas at Austin, on the evolution of American political and institutional norms. So I hope you will join us for that. I encourage you to learn more about CFR paid internships for students and fellowships for professors at CFR.org/Careers and visit CFR.org, ForeignAffairs.com, and ThinkGlobalHealth.org, for research and analysis on global issues, and education.CFR.org for free expert-informed teaching and learning resources. Again, thank you to Sarah Kreps. 

KREPS: Thank you. 

FASKIANOS: It’s great having you back. And thanks to all of you. And we look forward to your continued participation in this series.

(END)