As details emerge from the trade negotiations between the United States and China this week, one thing seems clear: Rare earths were an important part of the discussions. China has a monopoly on the production and processing of the minerals used in the production of high-end magnets and chips. In response to U.S. President Donald Trump’s massive tariffs, Beijing’s new restrictions on critical minerals ended up bringing the two sides back to the table.
The battle over access to rare earths is part of a larger competition between Beijing and Washington on artificial intelligence. Who is best placed to win it, and what will that mean for the world? On the latest episode of FP Live, I sat down with the two co-heads of the Goldman Sachs Global Institute, Jared Cohen and George Lee, both of whom follow the geopolitics of AI closely. 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.
RA: George, at a high level, where’s China at in its race to catch up with the United States on AI?
GL: What’s been fascinating is the generative AI revolution has provoked a pivot inside China. The surge of confidence, investment, and focus in this area is really fascinating. If you go back to 2021, [Chinese President] Xi [Jinping] imposed a series of crackdowns on what was then the leading technology ecosystem in China. When we emerged from the COVID-19 [pandemic], with the rise of generative AI, China evinced some ambivalence early on. One can understand that in a more closed semi-authoritarian regime, a less controllable emergent machine is somewhat threatening. So, China imposed rigorous regulations around this space.
What’s changed is the emergence of a highly capable model from China. It expressed its own native capabilities and captured the attention of the global ecosystem around China’s ability to compete and lead in this space. That provoked a new policy response in China to lean into this technology and integrate it with its historical strengths in data, robotics, payments, etc.
So now we’re in the sprint mode of a real race for supremacy between the United States and China. And it’s really emerged as a critical vector of competition between governments.
RA: Where does DeepSeek fit into this, Jared? My understanding is that it didn’t shock computer scientists or insiders in the AI world, although it did shock the U.S. national security community. Why is that?
JC: There are a couple of reasons. One, there was a perception that robust export controls on China, particularly around GPUs, were limiting their compute power such that it was impossible for them to run large language models at the same scale. There was a sense that they had an uphill battle when it came to generative AI. But necessity drives innovation, not just smart computer scientists—and China has both. Part of what spooked everybody with DeepSeek is that it basically managed to perform at the same level as GPT-4 at roughly 5 percent of the cost. Whether or not it was operating at scale, it was a research milestone that introduced the idea that export controls on China was an insufficient strategy to holding them back.
The market’s reaction was outsized to the reaction from computer scientists, who knew what was going on. But as a result of the market reaction to DeepSeek, you’re also seeing the realignment of the Chinese private tech sector with the state-led system, as George mentioned. At the end of the day, that is the bigger consequence of DeepSeek than a technological or a research breakthrough.
RA: And, George, it strikes me that the Chinese system might have an advantage in its ability to corral public and private sectors together, whereas the American or even a Western system could have built-in checks that hold it back?
GL: On the one hand, Ravi, the United States and Western economies have thrived through the open, capitalist approach to innovation and problem-solving. Particularly with algorithmic advancements, that’s served us well. But you might jealously eye state-oriented actors like China for their ability to impose long-term plans for some of the predicates behind these models. Those include the ability to take a long-term view on building power resources, modernizing transition, sourcing resources like those critical minerals.
One of the things that was super interesting about DeepSeek is that it illuminated the fact that China can lead and innovate at the algorithmic model level. The technical work inside the DeepSeek-R1 model, the papers they’ve published, reveal some of the most interesting computer science work in making these models smarter, reason better, etc. So it’s clear China’s now at or close to the frontier on the algorithmic front. And they do have the advantages of more command control and consistency in marshaling resources like power, which will be really important here.
RA: The issue of U.S. export controls on the highest-end chips, coupled with China’s control of critical minerals, were both relevant in the U.S.-China trade talks this week. Jared, are export controls doing what they need to do from an American perspective?
JC: The [Trump] administration’s moves show their perception of the limits of export controls in the policy prescriptions. The Trump administration’s criticism of the Biden administration is that they focused on prevention—meaning export controls—and not enough on promotion, which I think is fair. And so, their approach is to simultaneously double-down on preventing China from accessing some of the critical technologies necessary to power AI while also flooding regional hubs with that same technology. It’s a stick followed by a carrot to other regions. The previous administration was less open to doing that latter part in places like the Middle East. One example: On the prevent side, the administration announced that anybody using the Huawei Ascend chip is violating U.S. export controls. This cuts off China from consumer markets that it desperately needs to cover many of the fixed costs associated with this buildout. But simultaneously, they got rid of the Biden AI diffusion rules that capped places like the Middle East at 350,000 GPUs. We’ll have to wait and see how this plays out.
It’s going to come down to the bigger question of whether the United States has the capacity to build the AI infrastructure fast enough to meet hyperscalers’ demand. There’s also a question of how comfortable they will be bringing sensitive IP associated with training large language models abroad and how comfortable they will be bringing sensitive customer data associated with training abroad. So those are open questions.
Now, the tricky part is that this isn’t unilaterally up to the United States. Because the supply chains are so intertwined, and because of the realities of globalization, everybody was comfortable moving supply chains that were dirty from an ESG perspective or had cheap labor to China until COVID-19. After COVID-19, the United States realized that it needed to access strategically important supply chains, including critical minerals and rare earths. The problem is the die has been cast. Everyone focuses on the lithium, the cobalt, the graphite, and the minerals that come out of the ground and gets euphoric when we find them outside China. The problem is, once you get them out of the ground, you have to crush those minerals, chemically treat them, purify the metal, and then, more importantly, you have to refine and process them into magnets and other things. And 92 percent of refining and processing rare earths into metals takes place in China. There are only five refineries outside China: Western Australia, Nevada, Malaysia, France, and Estonia. You cannot meaningfully move that supply chain. We in the West don’t have the human capital to grow that industry because we’ve retired a lot of the programs that produce human capital at universities. There are also ESG regulations. And when you have such a high concentration of the refining and processing capability and supply chain in China, it gives them a unique privilege to be able to manipulate prices.
GL: I’d add one thing, which is that the complexity of these machines can’t be underestimated. Jensen Huang, the CEO of Nvidia, recently said that their current NVL72 system, which is their atomic unit of computation today, has about 600,000 parts. Their 2027 next-generation machine is going to have about 2.5 million parts. Now, he didn’t specify how much of that was foreign source. But that supply chain is intricate, complex, and global. And so, it’s unrealistic to believe that we can completely reshore, onshore, dominate, and protect an ecosystem to create this level of computation.
Read More
How the United States Can Win the Global Tech Race
Bans and sanctions aren’t enough to beat China.
Is It Too Late to Slow China’s AI Development?
The U.S. has been trying to keep its technological lead through export restrictions, but China is closing the gap. 
Washington May Regret Overextended AI Chip Controls
Ever-tightening restrictions are boosting Chinese firms.
RA: On that, George, you have a debate between the AI accelerationists on the one hand and then China hawks on the other. This goes to Jared’s point about the trade-offs between prevention versus promotion. When you consider that China has a stranglehold on the critical mineral supply chain, doesn’t that undermine the arguments put forward by people who want to limit China’s AI development at all costs?
GL: It’s certainly constrained. But there are some who believe we’re approaching a milestone called artificial general intelligence, or AGI. One rationale behind the hawk strategy is that it’s a two- to three-year race. They argue we should do our best to prevent China from getting the resources to get there first, because once you achieve that nirvana-like state of AGI, you gain a sustaining advantage. Now, I would debate that but it’s a reasonable perspective. But I agree with you that the idea that we can cordon off China from advancing in this world is illusory.
JC: I would add to that there’s a macro geopolitical question creating a strategic mirage that may bias incorrectly toward some of the China hawks. It’s the idea that if you’re China, engaged in asymmetric competition with the United States, your biggest vulnerability is that the United States sits at the center of a multilateral economic architecture that allows it to overcome those asymmetries and level the playing field. And so, if you look at the current context, one could credibly ask whether, over the next three-and-a-half years, China’s strategy would be to play for time?
There’s a lot of infighting within that democratic economic order: tensions on trade between the United States and its two largest trading partners, Canada and Mexico. There’s no trade deal yet with Japan, the United States’ only G-7 ally in the Indo-Pacific. No trade deal with South Korea, with Australia, with India, or with the European Union. And so, these moments where the United States and China seem to work toward a deal only to have it fall apart in subsequent weeks? This creates a perception of weakness or desperation that, if it gets conflated with the economic circumstances in China, could lend itself toward an incorrect narrative. I don’t know if they are in fact playing for time, but we have to ask that question because if they are, a hawkish approach could, in fact, play right into that strategy.
RA: George, does America lose anything by not being able to compete in the Chinese AI ecosystem? American companies are losing business, of course. But what is the long-term impact?
GL: This is the second-order question around export controls and restrictions. Jensen Huang has come out and said that a $50 billion business opportunity in China is largely foreclosed to him. Second, being unable to deliver U.S. technology into China, reciprocally, the Chinese lose access to the volumes of our consumer market, the global consumer market. But on the other hand, we are forcing them to use Huawei Ascend chips at scale, to navigate away from the Nvidia CUDA ecosystem, which is the software they wrap around their GPUs. Essentially, we’re conferring domestic volume advantages to them that otherwise might have been taken up by U.S. companies. And necessity is the mother of invention; we are causing them to scale up inputs to these models that will allow them to be more prosperous, get that volume, refine, be smarter, better, faster.
RA: Jared, you and I have talked before about what you call the geopolitical swing states, whether it’s India, Saudi Arabia, or Vietnam. How are they triangulating between the United States and China when it comes to AI?
JC: Before “Liberation Day,” I would have said that the geopolitical swing states realize that the limits of swinging with flexibility are around the critically important technologies. And that the United States, because of its advantages in generative AI in particular, had a lot of leverage in terms of being able to push countries to make a choice. At least for now, that is largely still true.
The caveat is, I think, the advantages over time will seesaw back and forth. As George mentioned, whoever gets to AGI first will have a unique posture in maintaining a competitive edge in this competitive coexistence. But countries will be chipping away at areas where they’re falling short for the rest of our lifetime.
These geopolitical swing states don’t block together. They act individually. It’s not a nonaligned movement. They look at their economic advantages and see a fleeting moment. They don’t know how long competition between the United States and China will be a framework for international relations. But they want to get as much out of it as possible.
Trump’s visit to the Middle East told this very important story: The narrative of the Middle East is no longer a story of security and shoring up energy supplies. It’s a story of investment and technology partnerships. And the three wealthy Gulf countries that Trump visited—Saudi Arabia, Qatar, and the United Arab Emirates—got public validation from the president of the United States that they are not just geopolitical swing states. They are major commercial players at the sovereign level in the most important and consequential technology invented since the internet.
GL: These swing states play an exceptional role in the world of this race for AI supremacy. The risk with AI is whether those swing states will be in an open, democratic U.S.-driven ecosystem or in a Chinese ecosystem? This is one of the perils of export control and of a less open approach.
RA: George, is this a case of a rising tide lifting all boats, even outside of the swing states? Or if you don’t have the clout, the money, or the energy, you just can’t keep up?
GL: Yeah, there’s a little of both. On the positive, whether this emerges from the United States, China, or likely both, the declining consumer cost of this technology means that whether you’re producing these intelligent tokens or simply consuming them, they are getting cheaper. So if you’re not on the leading edge of producing AI, you still get to benefit.
At the same time, if you don’t have native expertise, insight, and resources here, you are de facto dependent on others. Critical technology dependencies have real consequences—on defense, on culture. The impact on your economy, of not having your destiny in your own hands, is maybe threatening.
RA: Power is a big part of this. Jared, how have recent advances changed the power needs for the growth of AI? And how does that then play into the geopolitics of competition here?
JC: We’re grappling with hockey stick growth in terms of power demand without having prepared ourselves for that kind of an abrupt change. George mentioned Nvidia’s 2027 Kyber rack designs. These racks are now 576 GPUs on a single server rack that requires enough power for 500 U.S. homes. It requires 50 times the power of server racks that power the internet today.
When you talk about how many gigawatts of power the United States is going to need to bring online in order to meet the AI infrastructure demands, the numbers range from, like, 35 GW to 60 GW. That’s a huge delta in and of itself in between.
Some of the second- and third-order effects of this in the United States is a growing comfort getting back in the nuclear power game. But China is also experiencing the same thing. And one of the things that causes great consternation in the national security apparatus is China’s investment in nuclear for national security purposes. China is a huge investor in coal, in renewable, and in nuclear. So they get the power dynamics. And there’s not the same permitting challenges that we have in the United States and certainly not the same political challenges.
GL: In renewables alone, China added [the equivalent of the] United Kingdom’s power capability in the past year, so they’re building renewables to extraordinary scale. They have 30 nuclear plants under construction today. They have the ability and the willingness to scale coal, which is more controversial in the rest of the world. And this is actually an interesting artifact of their more command-and-control system, which can be both a bug and a feature. Plus, their lead in batteries. They produce 75 percent of the world’s batteries. And so, scaling batteries together with renewables, putting data centers that can benefit from that extremely low cost of intelligence per joule, it’s a very powerful thing.
RA: George, let’s talk about business implications. There’s so much volatility right now in everything you both are describing about the state of the AI race. How do companies navigate this?
GL: It’s inherently difficult. The pace of improvement of the technology is so steep. And as a technologist at an enterprise, you have to make a decision about when and where you shoot your shot. And so, you could move too early in this, make some decisions about deploying this technology too aggressively, wake up, and find the architecture or the leaders have changed. Or you could wait too long and see your competitors have established a sustainable lead over you. So, it’s very difficult.
The other thing I would observe is that it’s very hard to interrupt enshrined workflows in the enterprise. We’re all running experiments, which are beginning to become production projects that are yielding value. But while the technology is on this curve and enterprise adoption is slower, I’m optimistic that it’s inflecting upward. I think 2026 and beyond are the years where we’ll really start to see enterprise impact.