The AI safety debate has produced plenty of warnings over the past year, but on September 17 one of the industry’s sharpest critics put a price on the risk that no one else had named. If the leading labs grow dangerous enough that society cannot absorb the cost of their failure, Alex Karp said, then the state should take them over.
Karp, the co-founder and chief executive of Palantir, made the comment in a CNBC interview, and the phrasing was deliberate. The question, as he framed it, is responsibility. So long as the models stay in private hands, the people who build them should answer for what they do. Only if the risk outruns any private owner’s ability to bear it would nationalization, and Washington’s control, become the remaining option.
The remark landed in the middle of a running argument about how the industry should be governed. The day before, Mark Zuckerberg had said competition in the market would pressure companies into safety on its own. Jensen Huang had argued that no new laws were needed. Dario Amodei was pressing the opposite case, calling for a slowdown. Karp’s contribution was to question the motives of the whole debate rather than take a side in it.
Palantir occupies an unusual seat in this conversation. The company sells software to governments and defense agencies, and Karp has spent years arguing that the most serious technology should be built with the state in the room. That background gives his nationalization comment a weight it would not carry coming from a venture capitalist or a pure consumer software executive.
The logic is straightforward, and Karp laid it out in plain terms. A company that builds something dangerous should carry the liability for the harm. If the liability is so large that no company can carry it, then the asset stops looking like private property and starts looking like critical infrastructure, which is the point at which the government becomes the owner of last resort.
The comment is less a policy proposal than a line in the sand. Karp has not drafted legislation, and Palantir has no position paper calling for the seizure of competitors. What he is doing is shifting the frame: the argument over AI safety, in his telling, has been conducted as though the companies and their investors were the only parties with a stake, when the public is the party that would absorb the failure.
That shift has been building for a while. Regulators in several countries have moved from asking whether AI should be regulated to arguing over how. The labs themselves have begun to publish safety evaluations and invite outside testing, partly in response to the pressure. Karp’s intervention suggests that the next stage of the debate will be less about codes of conduct and more about who is liable when something goes wrong.
The nationalization idea draws a sharp line between the companies that build frontier models and everyone else. Karp has said before that Palantir does not chase artificial general intelligence the way the frontier labs do, and that distinction is doing work in his argument. A company that builds a tool can be held to account for the tool. A company that builds an agent that can improve itself is a different kind of liability.
Skeptics of the comment read it as rhetoric, a way to stay in the conversation while committing to nothing. Nationalization of a private technology company has no recent precedent in the United States, and the political conditions for it do not exist. The practical question of what a federal takeover of a frontier lab would even look like, from shareholding to management, has no obvious answer.
Karp’s defenders say that is the point. The absence of a precedent is not an argument against the idea; it is the measure of how far the technology has outrun the institutions built to govern it. The warning, in this reading, is meant to be heard before the risk arrives, not after.
The interview drew attention because it named a possibility that the industry has preferred to leave unsaid. Every frontier lab now argues that safety is a priority, and every one of them argues that it can be trusted to manage its own risk. Karp is betting that the public, and eventually the government, will stop taking that claim on faith.
What happens next depends less on Karp than on the labs themselves. If the models keep improving and the incidents keep coming, the question of who owns the risk will move from the conference circuit to the political arena. Karp has put himself on the record early. The companies whose futures he is describing will have to answer him, sooner or later.
