The artificial intelligence revolution is deeply linked with geopolitics. It’s well known that a small handful of countries and companies control the manufacturing of the highest-end semiconductors. But when you add in the scramble for the critical minerals that are needed to manufacture those chips; the data centers that house them; the land, energy, and cooling required to run those data centers; and the subsea cables that channel data and power, one realizes how the infrastructure that powers the AI economy crisscrosses the entire globe.

On this week’s episode of FP Live, I spoke with Jared Cohen about the shifting geopolitics of AI. Cohen has written about the topic extensively in FP. He’s a co-head of the Goldman Sachs Global Institute and previously worked at Jigsaw, Google, and as a member of the State Department’s Policy Planning Staff. We spoke on the morning of Tuesday, April 8: Certain references to tariffs at the time may be overtaken by events. The full discussion is available on the video box atop this page or on the FP Live podcast. What follows here is a lightly edited and condensed transcript.

Note: This discussion is part of a series of episodes brought to you by the Goldman Sachs Global Institute.

Ravi Agrawal: You’ve thought a lot about the concept of chokepoints. What are they exactly? Why does AI have so many potential chokepoints?

Jared Cohen: Well, AI software needs to run on AI hardware. And that means data centers. Data centers have tens of thousands of component parts representing a complex web of supply chains that rely on bilateral and multilateral commercial relationships across jurisdictions and different companies around the world.

Data centers serve as AI factories. Let’s use the analogy of the human body. The brain in a data center context can be the advanced chips and the high-end technology. You should think of the power that drives and makes data centers work as the blood flowing through the body. And think about the component parts, the tens of thousands of parts which you’ve mostly never heard of, as the skeleton.

So when we think about chokepoints, are any of those component parts highly concentrated, or even entirely concentrated, on a single supplier such that if you took that supplier out of the supply chain you’d have cascading bottlenecks? Are any of those component parts concentrated in a single country or set of countries, which would make them subject to disruptions from geopolitics trade barriers? Are any of the critical component parts impossible to reorient to a different location because of that complex set of relationships and contracts and web of commercial touchpoints around the world?

There are many instances where a single shock to the system creates a supply chain shortage. It’s rare that a supply chain shortage begins directly at the component part or the end product. It’s usually the subcomponent parts that are constrained. And then those cascading bottlenecks ultimately manifest in a part of the supply chain that everybody’s heard of and is reportable on the news.

RA: I like the human body analogy. How much of the AI value chain is suffering from the shock of [U.S. President Donald] Trump’s tariffs? Semiconductors were not part of the April 2 round of tariffs, but is the overall pie still affected?

JC: We’re bracing ourselves for more information about future sector-specific tariffs. And everybody tends to focus on semiconductors. But semiconductors only represent one piece of the larger semiconductor story.

The United States has about $82 billion of semiconductor imports. Those largely come from Taiwan, which is facing a 32 percent tariff; South Korea, which is facing a 25 percent tariff; and China, which is facing a 54 percent tariff. But that’s the direct import of semiconductors.

I would argue that $82 billion is not as important as the much larger aggregate value of the goods that have semiconductors packaged into them before they get imported to the United States. Those are indirect semiconductor imports. The United States imports $521 billion of machines that have semiconductors packaged into them before they arrive in the United States, $478 billion of electronics that have semiconductors packed into them, and $386 billion of vehicles that have semiconductors packaged into them.

RA: That’s something like $1.4 trillion, if my math is right.

JC: Yes, but it’s probably broader than that rough taxonomy. But the point is, tariffs on those categories is also a tariff indirectly on semiconductors. This is what I mean by cascading bottlenecks; the second- and third-order effects matter a lot when you’re talking about AI infrastructure because of how complex it is and how many different component parts make up a data center.

RA: That explains some of the recent panic in the stock markets.

I want to look more closely now at some of the chokepoints you’ve described. You had this great essay in FP recently about undersea cables as a massive but under-covered vulnerability in the AI economy. Talk to us a bit about that.

JC: I point out in the article that 80 percent of global trade happens above the ocean. But 95 percent of data flows beneath the ocean. This undersea flow happens on a technology that first emerged in the 1850s. So a pre-Civil War technology is highly relevant to the AI conversation. And so you have 750,000 miles of undersea cables that transmit $10 trillion of financial transactions a day, an unimaginable volume of instant messages every single day, offshore power every single day—and perhaps most importantly, the exacerbating reality that national security secrets are also included in that. So it matters a lot. Just to put it in perspective, 750,000 miles of undersea cables is enough to go around the Earth 30 times. These undersea tables are made of fibers as thin as a human piece of hair and can go as deep as five miles down into the ocean. So they’re kind of our lifeblood.

But there are 150 cable breaks a year. Historically, those cable breaks happen primarily because, like any hardware, they naturally deteriorate as they age. Sometimes there’s an accident. But what’s changing is the scale and what’s driving that breakage.

Now, it’s affected by geopolitics. Just to give a couple of examples, back in October, a Chinese vessel in the Baltic Sea dragged its anchor over 100 miles and severed a cable connecting Sweden, Lithuania, Germany, and Finland. NATO actually had to intervene to seize the ship. In January, a Chinese vessel severed one of the 15 cables that connects the island of Taiwan. But not everything is geopolitical; shortly thereafter, one of the Taiwanese-governed islands, Matsu, had two of its cables break down and get severed due to natural deterioration. All of this can easily get conflated.

But it’s not just state-sponsored attacks. Over 90 percent of the Europe-to-Asia data traffic happens through the Red Sea. With the war happening in the Middle East right now, you’ve had multiple instances of cables getting cut in the Red Sea. The Houthis don’t have undersea cutting capability, which they claim to, but they have sunk a couple of commercial vessels whose anchors inadvertently severed some cables. And so the entire landscape under the water is changing quite remarkably.

RA: I want to move to critical minerals. China has long controlled many of these minerals but now has also become a kind of monopoly player on the processing front. How does that play into the geopolitics of AI?

JC: I want to preface this by saying there are a number of supply chains that the United States deems geopolitically important, including the 50 critical minerals and 17 rare-earth elements. I’m concerned that, even in the best of times, these highly complex value chains are not well understood. A voracious appetite to diversify a supply chain may get out politically speaking in front of the economic realities of what’s possible. So if you look at critical minerals and rare earths, most of the focus is on mining and where the minerals are. So you’ll hear that a new lithium mine was found in this place and a new nickel mine was found in that place and this will transform our dependence on China for critical minerals and rare earths.

The problem is the value chain is much more complex than just where the minerals are. Once you mine the minerals, you need to purify the metal, chemically treat it, refine it, and process it. And 92 percent of all refining and processing capability exists in China. It’s highly concentrated. There are only five non-Chinese refineries in the entire world. There’s one in Nevada, one in France, one in Western Australia, one in Malaysia, and one in Estonia. So that’s another reason why this is a very difficult space to break into. So you have ESG [environmental, social, and governance] challenges—it’s one of the dirtiest processes in the world. And you have a high concentration of the refining and processing capabilities in a single country. The United States, permitting wise, can’t get its act together fast enough to permit refineries and processing facilities. But also it is not a field that is attracting large amounts of human capital. The big mining programs are no longer actively training the next generation of talent. And even if you address all of that, you have price manipulation threats from China, which controls so much of the market.

RA: As we keep looking at these different chokepoints, chips are the best known or at least most discussed element of this. What is your sense of how much longer the United States, through export controls, can stay ahead of China, at least on the highest-end level of chips?

JC: This is an issue today, but not before, because we used to run our data centers largely on CPUs, which worked really well for AI cloud workloads. But generative AI, and AI workloads with the high-end chips, represents a paradigmatic shift in the space.

People talk about whether or not the use cases will ever justify the spend. People obsess over open weighted models versus closed models. We follow what’s happening inside of these companies with a reality show fascination. But none of these known unknowns matter as much as the one that I think is most urgent: the nations that will determine where all this AI infrastructure is going to be built. Right now, the United States has a dominant position on this. There are 8,000 data centers worldwide, roughly half of which are in the United States.

But the problem is, for AI workloads the chips are ultra high density and require a concentrated source of power. Power demand in the United States has been roughly flat, or even declining, for two decades. We’ve been offloading baseload power. The difference between AI workloads powered by GPUs and cloud workloads powered by CPUs is that intermittent power doesn’t work.

RA: And what’s the difference between baseload power and intermittent supply?

JC: So intermittent power is wind and solar. It’s subject to the weather, so you can’t rely on it 24/7. Whereas baseload power—nuclear, coal, natural gas—is 24/7. And you need 24/7 power to run these data centers.

But that’s only part of the problem. The data center architecture for AI workloads is quite different. The chips get very hot. It needs to have liquid cooling, an entirely different HVAC system, etc. And the United States only has 3 percent vacancy in its data centers around the country. But even if you didn’t have a vacancy problem in U.S. data centers, you can’t easily retrofit existing data centers. From an engineering perspective, they’re not designed to handle these loads. The United States has enough baseload power, but that baseload power is in places like Texas and North Dakota and the data centers are in northern Virginia and Silicon Valley. And so the political complexity of moving natural gas from those locations to those data centers is quite difficult and prohibitively expensive. So the United States doesn’t have enough powered land to meet AI demand over the next couple of years on its own. And it doesn’t have enough differentiated data centers. It can do a lot of this. But my view is if the United States wants to maintain a dominant position in the most advanced AI, it’s going to need some kind of an overflow option to provide that excess capacity.

Read More

Containment Can’t Win the U.S.-China Tech Race Alone

Washington needs to foster innovation to beat Beijing.

Analysis

|

Stefanie Kam Li Yee

Build Allied AI, or Risk Fighting Alone

Cutting-edge systems need to be developed in tandem.

Argument

|

Becca Wasser,
Josh Wallin

American AI Is High on Its Own Supply

Why hasn’t the messianic urge for “efficiency” hit the tech industry?

Argument

|

Bhaskar Chakravorti

RA: The geopolitics of this are fascinating. Where could you get the energy needed for strong, nonintermittent baseload power? And I’m curious how the Middle East or the Gulf states, which are energy rich and have the ability to move quickly, fit into these calculations?

JC: Before I answer that, Ravi, I’ll just give a sense of the scale of the energy needs. So, in 2022, the United States had roughly 17 gigawatts of power that was powering its data centers. Estimates from our Goldman Sachs investment group say that by 2030, that’s going to be more than double to 35 gigawatts of power. That is an extraordinary jump in a country where power demand has been flat for two decades. Today, U.S. data centers use 3 percent of U.S. power and electricity. We estimate that by 2026, that number is going to look more like 8 percent.

The United States has three kinds of countries it could rely on to build out some of this capacity. You can keep it in the Western democratic world: Canada, Australia, the Nordic countries. That’s a little complicated right now because of what we see playing out. But even if it wasn’t complicated, those countries have the same challenges as the United States in moving baseload power to data centers.

Places like Malaysia and Indonesia have tons of cheap energy; a lot of these data centers have been built there in the past. The problem is geopolitical; if you train frontier models in those countries, there’s nothing to stop that capacity from going straight to China.

So then that brings you to a third option, which is the Middle East. The Middle East to me is one of the more attractive hedges for the United States because you have an unlimited amount of capital, tons of cheap energy, controlled regulatory environments, unlimited powered land, oceans to build data centers near for liquid cooling, and—perhaps most importantly—you have the sovereign ambition to do it. Now, the criticism levied by the administration—which put Middle Eastern countries into a tiered category where they were capped at 320,000 GPUs, which is not enough to realize all of their ambitions in the AI space—is how do we know that the Middle East will stay onside and so forth? And that’s certainly a gamble. But if you look at the United Arab Emirates, Saudi Arabia, and Qatar, there’s been a paradigmatic shift in these countries with younger leaders taking over a decade ago and the return of great-power competition between Washington and China instead of a war on terror as international framework.

In a lot of respects, these Middle Eastern countries—the wealthy Gulf countries—are the ultimate geopolitical swing states. They control hugely important parts of the supply chain. They have abundant, flexible capital that they can deploy as they see fit. They have agendas that they pursue, independent of Washington and Beijing but are needed by both Washington and Beijing. And while they can’t always play both sides, they’re flexible, agile, and increasingly asserting themselves economically. AI represents a once-in-a-generation opportunity for them for geopolitical mobility. If they’re going to pursue it anyway and are trending away from their path leading up to 9/11, and we need an overflow option, we should do everything possible to help accelerate that path.

RA: You brought up this idea of a geopolitical swing state, a term that you coined. But if you set up partnerships with them and they build data center capacity, what is to keep them from swinging to China at a later stage?

JC: Geopolitical swing states have quickly found that AI represents a ceiling on how much you can swing.

This gets back to the point about chokepoints. If you’re a geopolitical swing state reliant on critical components controlled by the United States, then the United States has a big say on how much you can swing. And because of export controls on high-end chips, the United States has forced the hand of countries that initially wanted to build a cocktail technological ecosystem of Chinese and American technology. When it came to advanced technology, they had to make a binary choice.

The concept of a geopolitical swing state represents a new moment in history. These countries have certain attributes that position them to take advantage of the tension between the United States and China and achieve enormous mobility, which they will maximize to their own interests. The swings that they make or don’t make will be in service of clearing geopolitical brush to maximize their interest as well as their usefulness to both countries. But there are some issues where maximizing their usefulness exclusively to the United States serves their interests because the United States is the only country that has the 2-nanometer chips that they need.

RA: It seems as if, instead of swing states changing how the geopolitics of AI works out, the geopolitics of AI is restraining some of the behavior of swing states.

JC: Yeah, it’s a fascinating thing. The entire world found out about generative AI at the exact same time. On Nov. 30, 2022, the public release of ChatGPT created this unique moment of technological simultaneity that we’ve never experienced in history before. We’re revisionists, and we like to think of the internet as emerging more abruptly than it did or social media as emerging more abruptly then it did. This was abrupt and simultaneous, even if the lead-up to it was not, in terms of the ubiquity of understanding. And while there was simultaneity, there wasn’t parity in terms of who was best positioned to win from a commercial or sovereign perspective.

But to me, since that moment of simultaneity, the illiberal use cases of generative AI matter a lot. There is chatter about how it can be used for disinformation and misinformation, bio and chemical warfare, more traditional warfare, etc. All of that is true. But the geopolitics, at least right now, around AI are anchored a lot around the physical movement of goods or physical infrastructure that’s required to facilitate the movement of data or to run AI software. So it’s a little bit of a back to the future.