Commerce Sec. Lutnick and Palantir CEO Alex Karp participate in a fireside chat — 9/2/2026

And it has to be culturally specific. If you just take warfare, the German way of fighting was different than the French way, which is different than the British way, which is different, and you see that nowadays. There are technologies that are linear. This is not a linear technology situation. And so it's just crucial that everybody has a private space where you're like, "Is this working? Why is it working? What's not working?"

And then in every country, you have part of the stack you can realistically power. I mean, China is a little different. I think that part will be mostly powered by Chinese companies, but every other country will have to make a decision of what part of the stack will be powered by American companies, what part of the stack will be powered by indigenous companies, local companies, who are those local companies, who are the experts. And that expertise is very, very rare. This is also one of the weird things about this industry: there's so many things—it's like your industry, like you're a master in finance, and there's so many things you know about finance that it would be hard to replicate. Say we switched positions, I am certain I would... That kind of knowledge is just really important to make sure you have sovereign ownership of in your country. If we can help, we will, but ultimately, you'll probably have to do it yourself.

In general terms, you have compute—so I think Jensen was here.

World's expert, talk to Jensen. And then you have basically—the caricature of it is you would have frontier models. In reality, you're going to have frontier and non-frontier models, and non-frontier open models that you can make frontier-like in specific use cases. And so what you're trying to do, let's say you're trying to solve a problem like: how do I optimize the supply chain of my country in a way where I know every single dependency on a country I may not trust? That's literally an issue every single person in this room has—or may not trust tomorrow, or if this part left, what would be the replacement?

Replacing the parts is very difficult because a part isn't a part; the part made by this company is different, even if the same spec. Every single manufacturing company in the world has this. And how do I do it while not exporting my way of manufacturing to another country? Let's just say I am a French company. I have a way of building a nuclear reactor. Nuclear reactors are very complicated. I need to know every single part of my supply chain that is not French or not European, and I also need to know how I would have to rebuild the supply chain if I had to buy a part from someone else, and what would be the cost. Every business in the world has this problem.

So if you want to make a frontier model be able to work on that problem both passively—meaning you know the answer when you ask the question—actively meaning you know the answer when you ask a question, but also passively, you guys get in a fight and you're like—I mean, you're a tough guy, so that's just—and you're like, "Okay, we've got to rebuild the supply chain."

How do you get these models? Without an application layer, they will, A, give you an answer, but that answer, even in the active state, is not 100% accurate. It is an estimation of the answer, and it also degrades over time. So the first problem is: how do you make sure that answer is correct and you can decomp it? How do you take apart the question so you can see every single assumption in the answer, so that if you're a country leader and you have to decide that? In order to do that, you will need an application layer. You'll also need an application layer—which is what we provide, not just... but someone will have to provide it. You also need to make sure that your way of building the supply chain is not being migrated to a lab model company.

In reality, what you're going to do is you're going to bring someone in who understands what part of the data has to be captured in this model. To do that, you're going to have to either post-train a closed model, or as likely currently, you're going to train an open model on the precise questions you need asked and answered. And you're going to do it in a closed environment. That's how you're going to solve this problem, because otherwise you have to worry about—let's just say you tell me cross-country skiing is evil and I pull your... None of these things actually happen in practice. So what does that mean for you? It means you need someone who understands the strengths and weaknesses of every model, also the legal vulnerabilities that you're putting yourself into by using certain models.

You need someone who technically understands the interplay between your proprietary data and a model as it exists and a model as you can train it, and can also set up an architecture that will allow you to bring in more expensive and cheaper models, but also more secure and less secure models. You may not care if a model from America that you don't trust is running on your ad campaign in Argentina, which is your biggest market; you will care if it's running on how you actually make your nuclear reactor work. And so that part of the stack, we are the application layer north—we will put an application—and south to some extent, the integration layer. We will put our ontology in the middle and allow you to train these models so that you can control them and there's no exfiltration.

Whether you use our product or not, you will need this to make these things, A, work, and make them secure. Just to mention, my biggest criticism of regulation outside of America is: please get technical experts to do it. Currently, the application of it is you have companies that are essentially there to deploy large language models. Those companies are exfiltrating, not intentionally, your alpha to the company that provides the model. It is literally the opposite of what you're trying to do. This is not sovereignty; this is anti-sovereignty. Whoever is telling you this works needs to be fired, because they are hurting your country. They're not hurting me, but this is literally the way to export every one of your secrets to some other company. It's only because the people doing the regulation don't actually understand the technical material. This is very hard. Every country has this problem.

I think you've been very effective in kind of tweaking at the edges when people have gotten out of control, in a way that was very effective. But we have the problem: who's going to monitor the development of AI? It's a very hard problem because you need to be an expert, and why would you trust the expert? That's a really hard question. One simple way of asking and answering the question is: is the expert on the payroll? Because a lot of these experts are actually repping a company. I am too, but I'm not repping buying this product.

Part of the problem is—one of the things about the development of nuclear energy and, fortunately or unfortunately, war, was the people who did it actually weren't doing it for the money. And we have a problem now. I'm kind of more left, I think it would be fair to say. I believe in capitalism as the best system because everything else is worse. I think you believe in it as just the best system. There's a slight distinction there. I'm still somewhat of a German Social Democrat, which means things have to work. I am skeptical of migration. I think it was a huge mistake—not controlled migration, but what we got, which was the opinion of Helmut Schmidt. That was Helmut Schmidt's opinion. Maybe we could listen to him. In any case, he was a titan, for those of you who never heard of him.

It's very, very hard to do these things correctly. I'll tell you one of the most interesting things about regulation: when we had CDMA and Europe had GSM—these are the wireless technologies—most people thought CDMA was better, but most Europeans thought the regulation in Europe was better, and so you ended up with... I think Europeans would universally say they have a better product. You can debate it, but this is just a very different revolution where the technology is a breakout technology. You just can't intuit what would be good for your country. You have to really have a deep understanding of where is the IP and alpha actually being monetized. It doesn't inherently mean because a local person is doing it you're safe; that is actually not true at all.

That you need an application. There's two basic things. One—there's a technical version: you will need an application layer which makes the models valuable, which helps train them, which makes them specific, and that will also allow you to select weaker models that might be indigenous and control them. So let's say, as a matter of theory, you have a model from Chile. It's weaker than the model in America. You still might want to use it. Open models are... We have a very important client in America for lots of reasons. They wanted to use open models. The open model is actually, on certain vectors—now again, you can't generalize—on certain vectors, that model is better than a frontier model. Now, it's only certain vectors, but it's the vector they care about.

Then everybody has the same issue as the US government. Every single one of you has... Do you really want to call a large language model provider and negotiate the parameters of what you are doing on your security apparatus, which in every country is different? For a neutral example: in America, civil libertarians believe that you should not purge data because by purging the data, it's hard to see what the government has done with the data. Core to the European concept of civil liberties is that you do purge the data. I'm a little bit on the European side because maybe you have an affair, maybe something like that happens—do you need everyone to know that? So I kind of like the purge the data thing, but I think it's hard to say one or the other is actually better. If you purge the data, you can't really see what the police did. If you don't purge the data, you can, but it's harder to have what might be viewed as a normal personal life, and most people do have a normal personal life at some time in their life—of course no one in this room, but I may have had one.

If you have to negotiate the parameters of the model—it's commonly known we have a big presence in French internal security, and we don't have one in other places. French internal security is very effective at what they do. Do they really need to negotiate when they use a model with an American frontier lab on the parameters that they're using the model? By the way, they might have a way of doing something that allows them to catch people no one's ever thought of. If you export that, you no longer have a way of doing it.

The second point is: every single one of you has a national in this country who is a titan in industry. I told Merkel this, and I told a lot of people: you have to get them to move home. I don't know how you do it, but you've got to get them to come and say, "You are now going to do this for my country, and you're going to set policy in my country." Those technical experts are, with the exception really of China, they are sitting here, and they're sitting here for lots of reasons. The reason why I think it's important to remember how crucial the time is: we don't have time to debate why people are coming here. You could debate it, I could tell you why they're coming here, but it doesn't matter. They are here. Almost everybody who does this is here. You need to get them to come home to your country, sit next to you, and explain what regulation should look like in your terms. What should the stack look like? What model should we use? What software should we use? What chips should we buy? And what is the cost of not... It may not be possible to buy the best stack.

I'm negotiating this a lot of times with countries abroad that, quite frankly, are like, "We don't want to be dependent on the Palantir stack completely," which is a heavy lift because there are certain things that just we have, especially on the battlefield. But I'm helping them to build a plan B, because of course they don't want to be dependent on me. In this case, I'm wildly pro-European despite, I feel like, being slightly maligned, but I'm supplying things to Europe that are not even in my economic interest. What if I get hit by a car? How are you going to like dealing with the next person running the company who's like, "Are you not paying me as much as the gas station?" That's what's going to happen. Developing these things with a technical expert, which you have, can really drill into these issues in hours. A non-technical expert is very likely working for a company that wants you to spend money.

And then the last lesson: you don't need money, you need knowledge. A country I care a lot about is always telling me how much money they're going to spend. I don't want to hear how much money you're going to spend; who's going to get the money? The money doesn't matter as much as who gets it, and that is crucial. The person who is in charge of spending and investing in tech needs to be someone who really is thinking, "Why am I doing this? I could be making even more money in America." And you remind them: because your country is great. That's how you do it.