The Limits of the American Way of AI

DAN KURTZ-PHELAN

Ben, thank you for joining me and for the essay you and Teddy Collins wrote in our new issue, that one is called “The AI Grand Bargain.” You had also, I remembered, written a piece for Foreign Affairs, in 2020, not so long before you went into government. That one was called “The U.S. Has AI Competition All Wrong,” and I was struck in reading that piece just how much of it foreshadowed both the debates we’re having right now in this space, but also the last administration’s AI strategy, which I imagine is not entirely an accident.

BEN BUCHANAN

Thanks for having me, Dan. Yeah, I think it’s fair to say that piece did do a little bit of foreshadowing, and I take great pride in it not least because my understanding is that is the first time we got compute into Foreign Affairs as a noun.

DAN KURTZ-PHELAN

That’s right. It was a major copyediting crisis, but we got through it.

BEN BUCHANAN

It was a sign of things to come, I think, of the importance of compute, or computing power if you prefer, to national security policy.

DAN KURTZ-PHELAN

Exactly, and I think you were the first person to explain that to me, so I appreciate that. I want to go back to that essay, but I want to start with the new one. You and Teddy make an observation early in the new piece about how different artificial intelligence is than most major cutting-edge technology of the past, especially those that came to decisively shape national security and geopolitical competition and the set of issues that we focus on. I’m curious how you characterize that difference, the kind of public-private balance, and why that, in your view, is so fundamental to understanding both where we are, but also some of the challenges that we’re facing?

BEN BUCHANAN

That’s right. I think this is the core intellectual through line to AI policy today and many of the most fascinating questions, scholarly and policy, in that AI in our view is the first really revolutionary technology, at least for the last hundred or so years, that is primarily coming from the private sector, without U.S. government funding or direction. And if you look at space or nukes or the early days of the Internet, the early days of microprocessors, GPS, radar, so much of this, there’s a very strong government hand, usually the Department of Defense, in some cases the intelligence community, that is really animating that technological development. It’s the private sector doing the work in many cases, but they’re often doing it at the behest of the U.S. government and for U.S. government contracts.

And the consequence of that level of the U.S. government involvement in those technological revolutions is that the U.S. government almost by definition has a little bit more understanding because someone in the U.S. government is managing that account and also [has] some capacity to shape where the technology goes. And it was so clear to us going into the White House in 2021 that the U.S. government didn’t have this in AI. If you go back historically, it did, 1950s, 1960s, 1970s, 1980s, absolutely. But the current deep-learning moment that’s propelling so much AI progress is not one that’s done at the behest of the U.S. government. And there’s a lot of benefits to that. [It] saves money for the taxpayer, but it also makes policy harder and in some sense poses a new set of policy challenges, as you said, about the relationship between the public and private sector.

DAN KURTZ-PHELAN

If we look at, let’s say, the Internet, to take a relatively recent example of the ones you listed, what has changed between, I don’t know, the early 1990s and the last several years of AI? Is it something about the technology that explains that difference or something about the way government and the defense industrial base, defense research base, work that accounts for it?

BEN BUCHANAN

I think the Internet is a technology that very much comes out of the government, the Department of Defense, [Defense Advanced Research Projects Agency], [the Advanced Research Projects Agency Network], all of that. And I think one of the consequences of that is you did have a U.S. government in the Clinton administration that understood fairly quickly what the Internet could do, and I think made pretty savvy policies around that. So that I think is probably an undisputed historical fact at this point. Of course there’s debates around Internet policy, but I think in general, the activities that they carried out were tech savvy and good. If you look at AI, we can really date the deep-learning moment to 2012 or so. Since then, it has grown up entirely in the private sector. And the scaling that I’m sure we’re going to talk about, the huge amounts of investment, all of that is private-sector capital put to work by private-sector labs for the most part, not even at universities, just a few companies that are really creating this technology that, at least in my view, will have profound implications for national security.

DAN KURTZ-PHELAN

If you could go back to 2012 or the years after that, is there something different government should have done that would’ve given us a different balance now, or is there something inherent in the technology that explains the difference?

BEN BUCHANAN

No, I think it’s good that this technology is private-sector developed. I don’t have an issue with that. I think a lot of what we tried to do was to build out government capacity so that when it came time to actually make policy around this, we were ready. And we tried to hire through the executive order and elsewhere, what we called the AI talent surge amounted to almost 1,000 people who are AI experts coming into the government. So, I think it is not a lament on my part that this grew out of the private sector, but it is a recognition that when something becomes this powerful, of course there will be some role for government and the like in applying it to national security. And we wanted the government to do that in as tech savvy a way as possible. And if you buy my argument about this coming from the private sector, it’s probably a natural corollary that it’s going to be a little bit harder for the government to use it well because it didn’t invent it.

DAN KURTZ-PHELAN

There’s a fairly striking claim at the beginning of the new essay. You argue that, while the U.S. lead in artificial intelligence looks “unassailable,” to use your words, it’s in fact far from that when you compare, especially to where China is, and if anything, our position is eroding and the prospects don’t look so great in years ahead. What are those concerns? What does the kind of picture, the superficial picture, of where we are miss about the trajectory as you see it?

BEN BUCHANAN

There’s no doubt in my mind we have the winning cards in our hand. We are in an incredibly fortunate position in large part due to the work of the private sector that we’ve talked about. We have control of the computing supply chain. We have a talent base that draws on the entire rest of the world. We have capital markets that no country can match. We have a lot of really good cards to play. We just have to play them, and we have to play them in a way that uses the strengths of the public and private sectors together.

And my concern is that we are not doing that, or we are not doing that as well, or we will do it with the same strategy we did in 2014 or 2018 where the government’s just not really engaging in AI in a meaningful way. And strategies that work then in the early years of the deep-learning era may not be as robust for 2026 and 2028 and 2030. So that is my concern, but I have total confidence in the leverage and the cards that we have. I just want us to use them well.

DAN KURTZ-PHELAN

If you project the pessimistic scenario forward, what does that look like? What’s the concern that you see?

BEN BUCHANAN

Well, if everything goes bad, we’ll start to see significant erosion on a few fronts. We will fail to build the energy to capitalize here in the United States on what we could do, which means we’ll fail to build the data centers, which means we’ll fail to scale. If we screw up the talent policies, we’ll fail to draw the talent from around the world that is fundamental to AI innovation, AI progress. If we fail to use our advantage in chips, we will let the Chinese buy high-end chips, which they’ll use to power their military and their AI companies. Maybe we’ll let the Chinese buy high-end chip-making equipment, let them build the domestic chip-making industry so they’re more independent from us. So there’s definitely a lot of downside risk and a lot of pitfalls ahead, but the theme of the essay is, we can avoid these downside scenarios. We just have to get it right on policy.

DAN KURTZ-PHELAN

For those of us who are not steeped in this and not working on this every day, the release of the Chinese DeepSeek model earlier this year, I believe, was a startling moment in that it looked like they had managed to create an impressive model, one that looked impressive to those of us who are nonexperts, with much less capacity, much less spending. Did it look that way to you? How did you read that release?

BEN BUCHANAN

Yeah, it didn’t look that way to me. I think a couple of things are true here. First, DeepSeek talent is extraordinary, and Teddy and I used to write a weekly letter or roundup to the national security advisor on what happened in AI that week in industry. This is a consequence of when industry was driving the bus, you’ve got to figure out what’s going on. And we wrote about DeepSeek many times starting in 2023, and we knew DeepSeek was really good, their talent’s really good, they’re really smart people.

DAN KURTZ-PHELAN

And just to be clear, you’re in the White House, you’re the kind of White House AI czar at that point.

BEN BUCHANAN

In the White House, yeah. I was a White House special advisor for AI, and Teddy was the director for AI on the National Security Council. So, we would write this. We were studying the papers. Teddy gets a lot of the credit for doing a lot of the technical work there. And we noticed, in 2023, DeepSeek doing impressive stuff, DeepSeek code and so forth. So, no one was surprised about the technical capabilities of DeepSeek’s researchers. I think the system DeepSeek released, really the impressive one was in December called V3, the R1 one that got attention in January was just kind of a natural extension. That was basically on trend with where American AI companies were, maybe a little bit behind where American AI companies were, but that’s what you’d expect of DeepSeek’s talent given the computing power that they had. And that system was trained on American chips that were either smuggled in or that were legal to purchase at the time.

And DeepSeek’s CEO said the number one constraint DeepSeek faced was not talent or money, but chips, and that makes total sense to me. So, I think the DeepSeek data point is really in line with the broader thesis here. In fact, the broader thesis of that essay you mentioned that I wrote in Foreign Affairs in 2020 about computing power being the fulcrum of AI progress, and our reaction at the time, what we told people then and what I think has been born out is, okay, DeepSeek did this once, they smuggled chips once, but the computing-power demands are going to grow exponentially and DeepSeek will not be able to keep up doing that. Certainly not if they use Chinese chips, and probably not if they have to continue on smuggled chips, assuming we enforce the export controls. And DeepSeek’s been very quiet since January.

DAN KURTZ-PHELAN

Was the extent of the smuggling of high-end chips a surprise to you?

BEN BUCHANAN

I was worried about it the whole time, and I think this is something where we have to make sure the U.S. government can enforce the policy. Otherwise there’s no point having the policy. I think some of the chips DeepSeek bought were legal. We had not yet controlled them, and there were huge debates and so forth about that, the H20 chip in particular. We don’t know exactly which chips they used, so I don’t know the breakdown of smuggled versus legally purchased. But we do know they were U.S. chips, and we do know, at least based on news reporting, that they have since tried to train on Chinese-made chips and they have failed, largely because the chips are not very good.

DAN KURTZ-PHELAN

Does the U.S. government have the capacity to enforce those restrictions effectively? Where does it even happen, or is this a new capability that needs to be built?

BEN BUCHANAN

Export control enforcement is not itself new. We have the Bureau of Industry and Security in the Department of Commerce. They’ve got a really important job. We advocated for more money for them in the Biden administration. We got some more money, but the idea of, “We export control certain technology and therefore we need to enforce that,” that is a Cold War idea. I think it’s fair to say we should continue to give them more resources and hold them accountable for enforcing these controls in an era in which we’re controlling a broader range of stuff. So, no one here is more supportive of beefing up BIS and making sure we can do the enforcement than I am. I think it’s fundamental. And it should be a bipartisan priority. Whatever the controls are, they should be enforced.

DAN KURTZ-PHELAN

And to be fair, you can look at plenty of moments in the Cold War, I think about A. Q. Khan and the development of the Pakistani nuclear program. Plenty of moments when export controls proved pretty leaky back then as well.

BEN BUCHANAN

It’s a tough business. It’s a tough business. I think the advantage of export controls on chips is you really need a lot of chips, so they’d have to get a lot of them to do things at scale, but we tried to set policies in place that would cut down on smuggling and smuggling routes and the like.

DAN KURTZ-PHELAN

I want to dig into a bunch of these specifics, but first stay at a somewhat more abstract level. I think all of us who are, again, not experts in the national security world throw around the terms race and AI lead without really knowing what it means. As you understand the race or the contest or however exactly you would characterize it, what exactly are we talking about? What does it mean to win? What do we talk about when we talk about AI competition, to put it in one way?

BEN BUCHANAN

Yeah, I think I prefer competition to race because race implies there’s a single finish line, and I don’t think that’s the case in AI, but I think this is a competition that has multiple facets. The first, and, in my view, really essential part of the competition is the competition to develop high-end AI capabilities. And sometimes this is called the competition for the frontier, and that is vital because those are the systems that propel the technology forward. Those are the systems that enable greater capabilities in the industry and the like. So that, I think, is what we see a lot of attention on, and that is the most compute-intensive part of the AI competition where I think American companies have done well.

Another area of competition is what you might call the application. So, taking that high-end technology and applying it to things like national security. And one of the examples we give in the piece is that it is not necessarily the country that invents the technology that applies it well in a national security context. And we talk about how the British and the French made a lot of key innovations early on for the tank, but it was the Germans who figured out actually how to use it. So just because Anthropic and OpenAI and Google do great in AI in a private-sector economic competition does not mean it is the American birthright that we will have an edge in applying that to cyber-operations in a national-security context. So, there’s a key part of what we worked on. A key part of the competition is making sure that we’re applying it well.

And I think the third part is what you might call diffusion. Just saying, “Let’s get this technology out across the world so that we can compete in markets overseas,” so that when someone in a developing country is querying an AI system, they’re using Gemini from Google or ChatGPT from OpenAI or Claude from Anthropic, rather than DeepSeek. And that’s also a fundamental part of this competition.

And in my view, there are things that you can do that are robustly good across all three. One of the big ones is don’t sell China chips and chip-making equipment because, well, chips and chip-making equipment are fundamental to the frontier kind of tautologically given the role of computing power. They’re fundamental toward applying that to national security applications, and they’re also fundamental to competing. DeepSeek needs to have the chips to perform inference and run their servers and train their systems and the like, and that’s what they need to do to compete in the developing world and beyond. So, those are the three parts of the competition. And I think we have a strategy for each, but then there are actions that are good across the board.

DAN KURTZ-PHELAN

It does appear that China has been more effective in those two second arenas of competition, that in some ways the Chinese system makes it easier to enforce the use of AI across the military and national security system, that it’s been more proactive about diffusing chips, at least in the global South. Is that impression right?

BEN BUCHANAN

I don’t know if that’s right. I think it is manifestly clear China is behind the frontier. So, competition sector one I think they’re clearly behind. I think the military context is hard to say in public. You could imagine an argument that goes, a centralized military will be better at applying something in a top-down sense. That may be true. The Chinese military is a cipher in many respects, and there are aspects of it that are very impressive, and then there are aspects of it where it’s riven with corruption and so forth. And frankly, I’m not a [People’s Liberation Army] scholar, so I would not want to get out over my skis in saying that. And then the third one, in terms of the Chinese diffusion, I don’t think they’re particularly good at that. They’re certainly not good at diffusing the chips because they don’t make that many chips and they don’t make very good chips. [U.S. Secretary of Commerce Howard Lutnick] testified this year that China is only making 200,000 chips this year. That’s not enough for a single data center. So, in terms of the competition, [the fear that] China is going to put a ton of data centers around the world, that is just not a serious threat right now, especially if we do our job on cutting off chip-making equipment. DeepSeek is definitely more competitive. I would imagine American companies are making a lot more money than DeepSeek is and so forth, including at lower price points, but I still think I’d rather be us than them on the diffusion aspect of the competition.

DAN KURTZ-PHELAN

The essay is about the limits of the American way of AI. I’m curious [about], as you’ve observed, the Chinese way of AI, if we can put it that way. Over the last several years, are there other elements of it that you find worth imitating, that you admire, in terms of their effectiveness?

BEN BUCHANAN

They’re doing a great job on energy. They’re doing a really extraordinary job on energy. And I think the stat here is, you can look at how much energy is being added to the grid as a percentage. And from 2005 to 2021, the net new energy add to the U.S. grid was something like 0.1 percent. [It] got a little better under Biden, bipartisan infrastructure law, big push on electric vehicles and heat pumps and AI data centers and clean power and all that. So, a little bit better. China is adding 12 percent per year, compounding. Even a developing country like India, my understanding is adding three or four percent. So, the energy story here, it’s a case where the Chinese are just doing extremely well and that will benefit them. And if we don’t get our act together on energy, we will struggle.

DAN KURTZ-PHELAN

I was struck by a line in the 2020 essay you wrote, I’m quoting you here, “Fears of an AI arms race between the two countries abound—and although the rhetoric often outpaces the technological reality, rising political tensions mean that both countries increasingly view AI as a zero-sum game.” [The] two countries here, of course, being the United States and China. Is it a zero-sum game? Is that the right way to think of it?

BEN BUCHANAN

I don’t think it is a zero-sum game. I think there is an opportunity for win-win collaboration between the two countries, and there’s a line in the president’s executive order that I’m very proud of.

DAN KURTZ-PHELAN

[U.S. President Joe Biden], just to be clear. Yeah?

BEN BUCHANAN

Yeah, President Biden’s, the first AI executive order. Something to the effect of, we are willing to engage on AI safety and cooperation not just with our allies and partners but also with our competitors. So, I do think there is a space for robust engagement between the United States and China, as two very significant leaders in AI, to manage the risks for all of humanity and to have positive-sum outcomes. I don’t think we need to give the most advanced technology in the world in order to realize that, and I wouldn’t want to trade one for the other. But I do think there’s a space here in which we can have positive-sum outcomes.

DAN KURTZ-PHELAN

And I know that we saw at least nascent talks on [the] AI and nuclear weapons complex between the United States and China. Is there any other meaningful cooperation on that front, or is that really the only place where we’re seeing any traction?

BEN BUCHANAN

Well, I don’t know what’s happened since I left, but both Teddy and I and a number of our colleagues met with the Chinese in Geneva in May 2024, which was a follow-up to the meeting between President Biden and [Chinese leader Xi Jinping], I believe in November 2023, where they talked about AI and agreed to have a dialogue. And I think we tried to say, “Look, we think there’s room for collaboration here,” and so forth. I wouldn’t characterize the Chinese as incredibly open to it, but I think the agreement we eventually came to on not having AI and nuclear command-and-control systems could be described as common sense.

But I think it’s really good that the two countries at least can show the world they can do common sense. And there’s nothing I can tell you [that] you don’t know about diplomacy, Dan, but sometimes you’ve got to start small and build from there. And had we had a Harris term, I think this is something that certainly I was committed to, and I think a lot of people in the White House and the administration would’ve been committed to, is saying, “Let’s engage with the Chinese, and let’s find the things where it affects everyone and we have shared interests, and let’s work on those.”

DAN KURTZ-PHELAN

And was the progress there possible simply because it’s so easy to imagine a doomsday scenario involving a, I don’t know, AI hallucination that starts a nuclear apocalypse. Is that a kind of clarifying idea?

BEN BUCHANAN

Yeah, I just don’t think anyone’s Team AI and Nukes—at least I wasn’t. But just because an idea is obvious, as you know very well, it does not mean that it happens in foreign policy or international diplomacy. So, I think the State Department, people like Tarun Chhabra, Seth Center, get a lot of credit for shepherding that from common-sense idea to actual reality. And of course, that goes up the chain to [former U.S. National Security Advisor] Jake Sullivan and others. And I think it would’ve been a foundation on which we could build, and frankly, on which I hope the Trump administration builds, and this is not a partisan thing.

DAN KURTZ-PHELAN

We talk about this in such bipolar, U.S.-Chinese terms, as with many other issues at this point. Are there other relevant players here as you look at [it], whether it’s American allies, the Europeans, or Japanese, or Koreans, or adversaries?

BEN BUCHANAN

Yeah, I really look at this as our network of democracies and our alliance of democracies joined together. And as your question implies, there are countries like Japan and the Netherlands that are world leaders in making chip-making equipment alongside the United States. Of course, Taiwan is the world leader in making chips. Korea has a very important role. Germany has a very important role given some of the inputs to the chip-making equipment. So, I think there’s a constellation of democratic countries here that we tried to pull together to assemble a united front, for lack of a better term, or a joined group of nations that had a commitment to how this technology could be used in a way that benefits democracy and manages stability and so forth.

DAN KURTZ-PHELAN

When you talk about first arena of competition, the frontier models, again in popular conversation, there’s been a debate about AGI, about artificial general intelligence, or superintelligence. Do you see it as a useful way of thinking about the issue or a distraction from the real challenges?

BEN BUCHANAN

I don’t love the term AGI, and I think there’s two things that are both true here. The first is, you don’t need to believe in AGI or superintelligence, I think, to buy the policies that we did. And the policies that we put into place and the way we view this, I think, is robust to a wide range of AI outcomes. And even in the world in which AI progress stopped tomorrow or next year, there would be implications of AI for cyber-operations, and we’d want to make sure that America comes out in that application and invention of the technology for those national security purposes. So, you don’t have to believe in AGI or superintelligence, I think, to say, “Well, we shouldn’t sell chips to China. We should make sure we’re building energy at home. We should get the world’s talent here,” and so forth.

That said, I personally believe that we see extraordinary AI progress, that a lot of the things that were hypotheses or theories in 2020 and 2021 when we were going in are now close to borne out, and there is good evidence that that progress will continue. There are really clear and measurable ways in which AI progress has accelerated in ways that we expected, but couldn’t know for sure what happened over the last few years. I see no sign of that slowing down. I think there’s some evidence that it’s actually going to speed up as we see AI systems accelerate AI research.

One of the most significant research papers from 2025 is a system called Alpha Evolve from Google that came out in May or so. And in that paper, Google showed that you could use AI to speed up AI progress, and they in fact used AI to find a better way of doing matrix multiplication, one of the core mathematical algorithms, essentially a more efficient way of doing math. And humans hadn’t improved matrix multiplication in something like 56 years. So, the AI system found a new way to do math, or a more efficient way to do math, that I think shaved the time off of that mathematical operation by something like 23 percent. So, it’s early, but you can begin to see signs of what so many AI researchers are looking for, which is this loop in which AI systems accelerate AI progress. Now there’s a question of how far away the ceiling is, at what point does it plateau? All that stuff is very real, but I think it is the case here that we should at least be thinking seriously about AI systems that continue to get profoundly better.

DAN KURTZ-PHELAN

And what’s the threshold that would allow you to call it superintelligence or AGI, or at least what we would think of in the popular conversation as one of those things?

BEN BUCHANAN

Yeah, the definitions are always shaky, but I think for AGI, what you’re pushing at there is the generality. So, the common definition is a machine that can do any cognitive task, or almost any cognitive task, as well as almost any human. And then a superintelligent machine is a machine that can do any cognitive task better than any human. But we should be honest and say these are fuzzy definitions.

DAN KURTZ-PHELAN

And there’s, of course, a growing skeptical view that all of this is driven by some combination of self-important grandiosity on the part of some of the private-sector leaders and maybe a desire to distract from near-term regulatory debates to focus on these kind of sci-fi scenarios. We won’t get into motives there, but it’s an interesting question.

BEN BUCHANAN

Yeah, and I think it’s fair saying, when I was in the White House, I had domestic responsibility as well. And the analogy I often used was the decathlon, which is, there’s a lot of events in AI policy, and America has to do all of them. So, at the same time we are thinking about export controls, and at the same time we’re thinking about energy, we also need to think about what are the near-term applications, positive and negative, of AI today? There’s a lot of upside in science; there’s some risk as well. And of course, there’s risk of discrimination, and how are we using AI in law enforcement and so much else. So, one of the reasons I love AI policy is it is so all encompassing and we have to do all of the things.

DAN KURTZ-PHELAN

You’ve mentioned energy a couple of times. If we go back to the essay that you and Teddy wrote, this is one of the limits to the American way of AI that you focus on. Why is this so central? It was really the first constraint, the first fear that you focused on. Brian Deese, I would note, the former head of the National Economic Council, your colleague in the Biden White House, wrote a piece for us about a coming electricity crisis. What worries you and what will it take to get really serious about the need to overcome those constraints?

BEN BUCHANAN

That was a great piece, and Brian is a great thinker on AI and the broader economy. I think what worries me is, I have made the argument now for years in your journal, that computing power is essential. We’re scaling computing power to scale AI systems. I think that’s pretty well proven at this point. Well, we’re also scaling energy, and if we’re going from a training cluster to train the AI system in 2020 that maybe was a couple thousand chips to training clusters in 2026 or 2027 that are millions of chips, and the chips themselves are larger, even though the chips are on balance more energy efficient, they’re still drawing more power. And you’re just going from an energy draw to, I don’t know, maybe in 2020 it was kilowatts, or, low end, megawatts to gigawatts. And if you look at what OpenAI is talking about or Anthropic [is] talking about, they’re talking about individual data centers of five or ten gigawatts of power. That’s as much as something like 24 U.S. states, I think, for five gigawatts. And we’re talking about an industry draw that’s something like 50 gigawatts, which is as much as essentially many countries in the world. So, there is just a mathematical conclusion here, which is, if you buy that computing power is going to continue to drive AI progress and you buy that companies are going to have the money to buy tons of computing power, then they’re going to have to find the power and the electricity. And I think that is going to be fundamental. And if we sit here in four years and we’ve really plateaued AI progress, odds are it’ll be because of a failure to build the energy.

And it’s worth saying, I think the companies need to pay for this. This is not the case where it would be acceptable, and we say this in the essay, for this to be done on the backs of citizens or taxpayers or ratepayers and the like. This is the kind of thing where the companies are making giant financial bets. They’re spending hundreds of billions of dollars and they should foot the bill for this.

DAN KURTZ-PHELAN

You can already see a political backlash starting to take shape in places where data centers have been built.

BEN BUCHANAN

Yeah. And I mean, Northern Virginia where I live, my understanding is something like 40 percent of the power in Virginia, electricity in Virginia, is going to data centers. Now, Virginia is an outlier because it has a history of building data centers, but I suspect this is going to be a very significant political issue going forward. And I have a lot of sympathy for people, ratepayers, who are saying, “I don’t understand. Don’t do this on the back of my pocketbook.” But I also believe we need to have a set of policies that makes it possible to build a lot of power, and the company should pay for it.

DAN KURTZ-PHELAN

Do you see any willingness on the part of the companies to do so?

BEN BUCHANAN

Yeah, I do. I mean, I think it probably varies by company, but I do think that. We had conversations with the companies and a lot of them said, “We would pay for it. We the companies would pay for it.” And they also said things like, “Well, we want to pay extra. We’re willing to pay a premium to make it clean power.” And one of the things that’s in the essay and also in our second AI executive order on infrastructure, is my hope for this is that we can use their willingness to pay for power to catalyze a next generation of electricity generation. Of course, advanced solar, advanced wind, the economics there are quite good, but also things like small modular reactors a little further out, and things like advanced geothermal where I think there’s a lot of promise.

So, in an ideal world, if we’re getting to an upside scenario, we’re sitting here in four or five years, we’ve built the power we needed to train the stuff and build out the data centers and the infrastructure. We have not done it on the backs of ratepayers, and we’ve used the company investment to catalyze a next generation of clean energy that everyone will benefit from.

DAN KURTZ-PHELAN

And I would note that between your piece and a couple of pieces that Brian Deese has written in our pages, there’s a pretty comprehensive blueprint about what you need to do in order to get there. Whether our political system can manage to actually implement that is a separate question.

BEN BUCHANAN

That’s right. I think this is the case where the cards are in our hand. To mix metaphors, the path is clear, and it’s just a question of, Can we get the system, private and public sector, to do this, and state and local? And there’s a lot of moving pieces.

DAN KURTZ-PHELAN

One of the other vulnerabilities, one of the other limits, to the American way of AI that you focus on in the essay is security and espionage. What are your concerns there?

BEN BUCHANAN

Essentially, we are inventing a technology with very, very deep national security implications, at least if you buy our thesis. If you buy our thesis as well, well, it’s coming from the private sector for the first time. A lot of its predecessor technologies grew up in eras of secrecy where the secrecy was kind of built in, for better or for worse—the nuclear programs of course, some of the space programs, many of the DOD programs. There are downsides to secrecy in cost innovation and the like. So, it’s not surprising to me that companies have been able to be very dynamic in part because they’re more open and collaborative and so forth. But we are going to have to find some mechanism in which they preserve that openness and that collaboration and that innovation, and yet also make sure their secrets are robust to what I suspect will be very significant foreign espionage threats.

And this is not speculative. We’ve already [given] a couple examples in the essay, but we know that there are cases of foreign intelligence officers or agents of foreign intelligence trying to penetrate American companies, trying to steal trade secrets and the like and take them back. And of course, I think China’s history here in a wide variety of areas beyond AI is pretty well documented. So, we should expect that, almost tautologically, as AI gets more important, other nations will recognize that and will use their full tools of national power, including foreign intelligence, to try to get secrets. Insofar as the secrets are important, which I think they will be, we will have to protect them wherever they are, including in the private sector.

DAN KURTZ-PHELAN

Even before AI became the focus of this conversation, there was not, as I see it at least, there was not a lot of success in getting pretty important private-sector actors to take their own cybersecurity seriously. It took a while to get there. That does not seem to suggest that it’s going to be easy in this case either.

BEN BUCHANAN

I don’t think it’s going to be easy, but I do think if you were to rank American companies by their cybersecurity posture, you’d have the banks doing quite well. They’ve learned hard lessons, I think, and have done well. And you probably would have companies like Google that are near the top, and hyperscalers and the like. So, it is probably fortunate that the companies that are inventing this technology do have some pedigree with security. What I would like is I would like the government to bring the things that only the government can do, in terms of threat intelligence and the like, to assist those companies. And this was a theme in the national security memorandum that President Biden signed in October 2024 that Teddy and I worked on extensively. It is to say, well, let’s see what we can do here to preserve the dynamic ecosystem, preserve the way you’ve done cybersecurity in the private sector, but also bolster that however we can with the kind of support that presumably only the government could muster.

DAN KURTZ-PHELAN

And I suppose the difference with some of the more traditional kinds of cyber-penetration is that you’re dealing with telecommunications infrastructure that was built 50 years ago in some cases.

BEN BUCHANAN

That’s right. Google, I think, has learned hard lessons, and companies like Google—I’m using them as a stand-in for the broader industry—have learned hard lessons already. And I think we have a foundation there. I want them to build what they’re going to build, and I want the U.S. government to help however it can.

DAN KURTZ-PHELAN

There’s another side, the flip side to this national security vulnerability, which is the way that the U.S. national security apparatus integrates AI into its own operations. As you put it, “Right now, strong partnerships between AI firms and U.S. national security agencies are few and far between, and those that do exist are in the early stages.” What is so hard about this? What are your anxieties here?

BEN BUCHANAN

There are a number of pieces of it. The first, of course, is the U.S government has to get it, and you need to have people in the U.S. government who understand what this technology can do. There were some, but I wouldn’t say it was the norm. And again, I don’t want to blame them. This is coming not from their area. They’ve got real jobs. It’s not their job to read AI research papers, but it is fundamental that the U.S. government figures out what’s going on in the private sector and applies it where appropriate.

Second, I think there is just an ossification of military and intelligence, a hardening of how it is done, that I’m sure you experienced when you were in the government. Sometimes that’s there for good reasons, sometimes that’s there for bad reasons, and that just makes it [so] that the military is not early adopter, or the DOD and the intelligence community are not often early adopters of technology that we’re going to have to overcome. And I think we actually made good progress on that. And again, a lot of hardworking civil servants get credit there.

I think probably the third is figuring out what are the guardrails that are appropriate for this? And this is an area where the United States needs to distinguish itself from China and needs to say, “Look, there should be real limits on how we use this technology. And it is not acceptable to use AI to build the domestic surveillance state.” Even if we want to use it to improve our cyber-operations against the Chinese, which we should, it’s not okay to turn that on our own populace. So, in some sense that is a break, but it’s also a break of a reasonable form.

What I found is that what everyone benefited from was clarity, in that everyone in the system wanted to know what the technology could do. They wanted to know how to apply it to the mission, and they wanted to know what the boundaries were. And that was what the national security memorandum was meant to do. I think we did a lot of the big-picture stuff, in my view, in a way that was right and set a good course. The work that remains to be done is now translating that to the rest of the system. Getting people at the line-officer level to buy in to make this a priority, and then changing a lot of government procedures, procurement and the like, to make sure we, the government, are buying technology from a broad range of companies and the like that are actually inventing this. And this is something that is called out explicitly in the national security memorandum as a priority, but I don’t know [if] that has actually happened, and certainly not in the three months that we had after that document was signed, and it needs to continue. So, it’s really important to me that we have a competitive ecosystem in AI that’s not just the same couple companies that are doing this. It’s especially important in a defense contract where you do have just a couple of big primes. So, that, I think, is a final piece of this that definitely remains to be executed.

DAN KURTZ-PHELAN

In the piece, you pull these elements together into what you call “a grand bargain between the tech industry and the government.” Describe that grand bargain. What does that look like?

BEN BUCHANAN

Yeah, grand bargains are always easier as taglines than as policies, but I think the analysis Teddy and I did is that, well, there’s some things here that only government can do and there’s some things here that only the private sector can do, and it is really in America’s national security interest for everyone to work together. And just because something makes sense doesn’t mean it will happen, but we think there’s enough alignment here of interest that the country could push this with savvy policymaking. So, obviously there’s the robustly good things: holding the Chinese military back, controlling chips, controlling chip-making equipment. Then American companies need talent, so bringing in high-end talent from around the world, things like the O-1 visa and the like, to get them the researchers they need. The stat that always stuck with me was, in 2022, when the study was done, of the top 100 AI scientists in the United States, something like 70 were born outside the United States. So, there’s clearly a deal to be done on talent. Obviously on energy, we’ve talked about this quite a bit, building out the energy here at the company’s expense to make sure they can scale the infrastructure in a way that they need to, providing security for the companies or security assistance for the companies, and then in return getting the companies to help with the application of AI to national security, which I have found companies generally are willing to do. But it’s a question of making sure that the fit is right. So, we think there’s enough things on each side of the ledger here that there’s a deal that can be done that will be to everyone’s interest, most of all of the United States.

DAN KURTZ-PHELAN

You spent three or four years at the government side of this in the White House. You’ve now been working with companies, advising companies on the other side for the last several months. Has your view of this grand bargain changed your sense of willingness to engage in it or challenges that you didn’t appreciate sufficiently when you were in government?

BEN BUCHANAN

I don’t think the contours of the bargain have changed. I think I feel more urgency now because the technology is clearly continuing to make a lot of progress. And there were worlds in which it could have plateaued in 2024 or 2025, and those are not the worlds we are living in. The scaling law of computing power yielding capabilities continues to hold. We are starting to see signs of a recursive loop in which AI makes AI research more tractable and more efficient. So, the data that has come in over the last nine months since I left the administration, or ten months, whatever it’s been since I left the administration, only reaffirms my view here that this is an urgently important matter.

DAN KURTZ-PHELAN

Let’s go to that moment ten months ago when you left the administration. I’m interested in how you see the continuities and discontinuities between the Biden and Trump approaches on these issues. At a high level, from a distance, it appears that we’ve gone from a sort of controlled and deliberate approach under Biden—know what you’re deploying before you deploy it, some regulation, a more restrictive approach to chip diffusion and what the rest of the world has access to to across the board, accelerationism. In some sense, I think that’s reflected in the Trump AI action plan. Is something like that, right? Is there more continuity that that story misses?

BEN BUCHANAN

It’s hard to pin down what the Trump administration AI policy is because there’s a bunch of different conflicting actions. So, the rhetoric is certainly very accelerationist, and yet if you look at the energy policy, with some exceptions, it isn’t. They’re canceling transmission lines, that would be the largest transmission line in the country. They’re canceling solar farms in Nevada that would produce five gigawatts. A pure accelerationist I think would say, “We’ll build as much energy as possible.” And that was the rhetoric that [U.S. President Donald Trump] talked about, I think, in his convention speech, building electricity for AI and so forth. So, that’s a case where I just think it’s been a little bit hard to figure out what the policy actually is. And my suspicion is you just have different people making different policies and there’s not a ton of coherence.

I think in the Trump administration, again, the rhetoric was, he said, staple a [green card to every diploma], or something to that effect. That hasn’t happened. We’ve seen a very, very different immigration policy that I imagine will have effects on AI researchers’ willingness to come to the United States.

And then on stuff like the chip controls, they’ve been all over the place. They took, in my view, a very admirable step of controlling the H-20 [chips]. They have in the AI action plan a really well-written call-out to the importance of computing power for AI competition with China, the need to deny this resource to the Chinese. They’ve reversed the controls on the H-20. They’ve been all over the place on what their posture is going to be on more advanced chips. They haven’t really done, as far as I can tell, anything on controlling chip-making equipment, which is an area where you need kind of continued maintenance to keep with changing times. So, it’s hard to point to a single coherent policy there.

In many cases, I think, the underlying rhetoric is not terrible, but I don’t necessarily see a clear through line of here’s how they’re approaching AI and here’s what they’re doing to win the race to the frontier or win the competition to the frontier, to apply this to the military where we haven’t seen very much. And then to diffuse American models around the world so that American companies outcompete DeepSeek. I just haven’t seen that kind of coherence yet.

DAN KURTZ-PHELAN

I want to ask you about differences in views on compute and how to maintain that edge, but I suppose it’s worth first just really lingering on this point that you keep coming back to. Can you just explain for those of us who are not experts in this why compute to you is the center of this, and what the components of that are?

BEN BUCHANAN

Yeah. A key part of the AI process is what’s called the training of an AI system. So, this is the part in which you basically show the AI data, you teach it about the world, and you build the system that you’ll then go off and use for new problems and the like. And this is a compute-intensive process. And one of the things that we have observed as an empirical observation, and that led to the piece I wrote in 2020, is that the capability of the resulting AI system scales with the amount of computing power used to train it. Assuming you’re also putting data in, but that’s usually not that hard, the more computing power used for training an AI system, the more capable the AI system is as a result. And we call this the scaling law. And in the sense that it’s not a law of physics, it’s an observation, computing power will get more efficient every two years.

So, this has continued to hold for probably, at this point, 12 or 13 orders of magnitude of computing power, where you keep increasing the scale of the computing power and then in a very predictable way, the AI systems increase in capability as a result. So, that is one really obvious point for the importance of computing power. In 2024, we began to see a little bit more nuance to the scaling law, 2023 and 2024, in which you could do other steps on top where you could also scale computing power. Basically using reinforcement learning, another kind of AI algorithm, on top of your AI system, and you could scale that as well. So now we had two things that you could use computing power to scale, and as you scale each of them, the resulting model got more capable.

And then we also noticed what you might call a third scaling law, or a third piece of the same scaling law is, once you had the trained system, you could then apply it to a particular problem and you could use, at the expense of more computing power, you could essentially get it to think for longer on that problem. In the same way that if you gave someone ten minutes to write a Foreign Affairs essay, it probably wouldn’t be great, but if you gave them a year, hopefully it’d be a little bit better or substantially better. As AI systems learn to linger and think about problems, they can often perform better. That’s great, but that just comes at the expense of more computing power. And you can see this if you use ChatGPT or Claude and you put on the extended thinking mode—it thinks for longer on a problem, you tend to get a better answer, but it uses more computing power. So, those are three reasons, all bundled together in this notion of compute scaling, why the more computing power you have to throw at AI training and use, the more powerful AI systems are as a result.

DAN KURTZ-PHELAN

And that comes down to how many chips you have, how advanced they are, and how much electricity you have to put into them essentially?

BEN BUCHANAN

Essentially. Yeah. And I think I’d add one component about how well the chips network together, which is a strength of U.S. chips as well.

DAN KURTZ-PHELAN

You’ve mentioned the importance of keeping the most advanced chips in our hands and in the hands of our allies, and out of the hands of China especially. There’s a different view that you hear from parts of the Trump administration. I take your point about incoherence in different parts of the administration having different views. What is, if you can give the most credible version of the alternate view, how would you describe that?

BEN BUCHANAN

I think the reality would be different if China could make a lot of chips. And sometimes I hear the argument [that] China can make a lot of chips, or if we control chips with them, China will indigenize their chip production faster, and therefore we should take a different course than you guys did. I just don’t think, as a factual matter, that’s right. I think as a factual matter, China in 2014 recognized that chip making was hugely important for national security. They began spending $200 billion of public and private capital to try to indigenize their chip-making industry, essentially put the pedal to the metal on that. The Trump administration, the first Trump administration, [former Deputy National Security Adviser] Matt Pottinger and others, I think, get a lot of credit for controlling advanced chip-making equipment to the Chinese to inhibit this progress, and the Chinese really didn’t make a lot of headway.

So, we are really outperforming the Chinese on chip making, probably by a factor of ten to 100x, depending on how you want to measure it. And it is fair to say that is the crux of this whole enterprise, and in the world in which China caught up, they would really have a lot of advantages given their talent and their electricity production. So, maybe some folks in the Trump administration think China is closer than they are on chip making. My view is they’re pretty far behind and we can take smart action by denying them chip-making equipment to hold them back.

Just to dwell on this for one more second, I know this is a foreign policy discussion, but it’s worth just going into the technology of chip making. My view is, this is the hardest thing we do as a species. This is just incredibly technically impressive work. Remember, the United States really can’t do this right now. The [2022 CHIPS and Science Act] helped some, but it’s really just a couple of companies in the world, really just [the Taiwan Semiconductor Manufacturing Company] that can make this at any kind of scale. You’re talking about incredibly precise lasers, vacuums thousands of times more pure than the vacuum of space emitting radiation with incredible precision to make these chips. It is a process that you spend ten minutes looking at it and you’re like, wow, okay, this really is something that is hard to do. And then you recognize spending ten minutes looking at it gets you 0.1 percent of the complexity of the process. And this was a technique for chip making that was built over decades by a network of American and other democratic companies. China’s going to have a very, very heavy lift indigenizing that all from scratch, especially if we’re yanking the chip-making equipment from them. So, that probably is the steelman case for the Trump administration. I don’t find it compelling as a factual matter, and I think we should all agree that we shouldn’t be selling chip-making equipment to the Chinese. We should make it harder.

DAN KURTZ-PHELAN

The other change between Biden and Trump, as I understand it, is fewer restrictions on the kinds of chips and the number of chips that could go to other countries in the world who are not necessarily close allies or adversaries to the United States. Is that a reasonable change? How do you understand that debate about diffusion?

BEN BUCHANAN

Yeah, this is something called the diffusion rule, which I would argue is distinct from the diffusion part of the competition, though it’s related. And this question essentially is, in a world of constrained chip supply, where every chip that’s being made is being bought, how do you want to determine where those chips go? And our view, I think was, we wouldn’t necessarily use this phrase, but pretty close to “America first” in saying we want American companies and the companies of our close allies to get first crack at the chips. And we want to make sure that we are, in addition to that, cutting off the smuggling routes through which these chips might be smuggled to China and the like. So, if you look at our rule, the diffusion rule, it allows for really large numbers of chips to go all over the world, but not so many chips that they could build out whole AI training ecosystems and the like. It gives essentially a preference, if you’d like, or an advantage to trusted American companies that we know will not smuggle the chips and that will meet security standards and the like. We create a pathway for companies from other countries to reach that status. So, this is not a protectionist policy, but it is saying we want the United States to lead in AI, to dominate in AI development. We want American data centers built all over the world, but we want to make sure that those chips are not being smuggled to China. And we don’t want to repeat some of the mistakes of the past of just offshoring what we imagine will be a very vital industry to nations that are willing to foot that bill. So, there is a mix of very tactical cutting-off-smuggling reasons for why we did what we did, and then also strategic reasons to just assert a very important role for the United States, even as we work with allies and partners all over the world.

And concretely, the rule would allow millions of chips to go to even countries that are not close allies, the vast majority of them through American companies, but also a large number to their own domestic companies. And again, there’s a pathway there for those companies to meet security standards and the like and get even more chips. So, it is not like we were saying only chips for America, and that’s it. We were saying, let’s have a managed policy for positioning America the strongest, and America and its allies the strongest, in a world of chip scarcity, where chips are increasingly important.

DAN KURTZ-PHELAN

The flashiest announcement on Trump’s, I believe, his first foreign trip to the Middle East in the spring was the news of a major data center in the United Arab Emirates with billions of dollars of American chips going into it. Did you see that as a good deal, kind of the right way of doing this? Or does that raise concerns to you?

BEN BUCHANAN

The devil’s in the details. Our policy would’ve permitted American companies to build big data centers in United Arab Emirates or in Saudi and other places that wanted them. So, in that sense, it’s consistent. Our policy also would’ve let us sell large numbers of chips, though not as many, to local firms in those countries. So, our posture was not “We’re giving them nothing.” The devil’s in the details on the Trump thing, and frankly we haven’t seen any details. And I don’t know how many licenses have been granted and I just don’t know how they’re operationalizing this. For better or for worse, the Biden administration put this all on a 200-page rule that lays out, it’s complex, sure, but it’s all there. So, we know what the deal was, and I don’t have that same kind of clarity on the Trump administration posture here.

DAN KURTZ-PHELAN

I mean, I think a lot of people following this debate without any of the expertise that you have, say, “Look, I remember when people were telling us that China would never become, would never get to the cutting edge of other technologies in the past.” And all of a sudden, we woke up one day and saw that it had. Does that worry you? Does your assurance here, I know you’re saying we should be empirical about this, but surely that precedent would be somewhat worrying?

BEN BUCHANAN

Yeah, I mean, I think the precedent cuts a couple of ways. First, insofar as some people in the Trump administration, like Howard Lutnick, have made the argument that we have to get China addicted to our chips so that they don’t indigenize, I think that the precedent of history here absolutely cuts the other way. In every single other industry, [electric vehicles], solar panels, whatever it is, the notion we’re going to get China addicted and they are not going to indigenize just doesn’t hold water. They’ve done it in every single industry, as your question suggests. So, in my mind, we know their intentions. Their intentions are to indigenize, their intentions are to kick out western suppliers, to have their own supply chain for chips.

So, with clarity on their intentions, then the question is, will they succeed? And I think my view is maybe someday. I would never doubt their ingenuity, but our job is to make it as hard as possible. And the good news is I think chip making is incredibly difficult intrinsically, such that we really can, with targeted action, buy a lot of time here. How much time? Is it five years? Is it ten years? Is it 20 years? I don’t know. It depends on the threshold, depends on the scale and so forth. But there’s no doubt in my mind about Chinese intentions. There’s no real doubt about Chinese capabilities. It’s just a question of what can we do to slow them down? And the good news is I think we can do a lot.

DAN KURTZ-PHELAN

I want to close with a perhaps somewhat more personal question. You’ve been working on AI and its intersection with policy and national security since the time when most of us in the foreign policy world knew of this only as an obscure and esoteric issue. If you step back and reflect on the last decade or so of your experience studying and working on AI and how it shaped policy and geopolitics, what has surprised you at that time? What about where we are now would’ve shocked you if you could go back and tell yourself ten years ago?

BEN BUCHANAN

I think I am surprised by how much this has become a mainstream issue. I think, intellectually, I expected something like that just because I had very strong convictions about where this was going to go from a technology perspective, and that’s part of the reason why I spent so much time on it as far back as 2015 or earlier. And the same is true, I think, for my co-author in this piece, Teddy, who, you know, we were studying this together way back when. But I think, emotionally, or as someone who has been a part of the process, it is just surprising to me to have not one but many conversations in the situation room about the intricate details of chips and chip making and where AI is going, or some number of conversations in the Oval Office about how this is affecting America’s domestic competitiveness and what this means for the American people and the foreign policy dimensions. It’s nerd stuff, and it’s surprising to me how much the nerd stuff that is near and dear to my heart has gotten picked up by a group of people who, if they’re nerds, they’re a different type of nerd.

DAN KURTZ-PHELAN

Well, Ben, thank you for this conversation and for the many great pieces you’ve done for us. We’ll look forward to more. And I should also thank you for, five or six years ago, forcing us to have the hard copyediting conversation about whether compute could be a verb.

BEN BUCHANAN

It is a metaphor for many, many things about where the world is going.

DAN KURTZ-PHELAN

That’s exactly right.

BEN BUCHANAN

Thank you, Dan, for having me in your pages then, and again alongside Teddy this time.

DAN KURTZ-PHELAN

Thanks a lot, Ben.

BEN BUCHANAN

Have a good day.