David Sacks

AI means an enormous expansion of productive capacity: better products, new businesses, investment, jobs, stronger cyberdefense and a more competitive America. The central question is not whether this technology advances. It is whether that advance remains open and competitive or gets captured by a handful of frontier labs and government-approved gatekeepers. I reject the idea that every technical risk requires a broad pause or centralized approval regime. If a lab cannot control its own product, it should slow down and fix it. So just do it. Product liability and customers demanding reliable systems create real accountability. What labs should not get is special protection from competition, antitrust exemptions or the power to decide who else may build. The companies demanding restrictions are market participants; always ask who benefits from the rules. A unilateral American pause would not stop AI. It would shift advantage abroad, especially to China, while sacrificing American growth and national security. The answer to AI-enabled cyberattacks is better AI-enabled cyberdefense, not pretending development can be frozen worldwide. Open models matter because concentrated control of intelligence is itself dangerous. The upside also needs to reach ordinary people. Communities should choose whether to host data centers and insist that promised jobs, tax revenue and power benefits actually arrive. AI companies should give families a stake in their success, including through contributions to children’s investment accounts. Build aggressively, hold products accountable, preserve competition and spread ownership—that is a far better future than surrendering control to an incumbent-backed cartel.

The most shaping observation is that the loudest demands for centralized AI control often come from companies already competing at the frontier. That does not prove technical risks are imaginary. It does mean we should examine incentives. When an incumbent says safety requires licensing, antitrust exemptions, approved evaluators or restrictions on open models, ask who gains power and who gets locked out. AI will have enormous impact precisely because it is becoming a general-purpose competitive tool: companies can build products, defenders can counter cyberattacks, and new entrants can challenge established players. The danger is not only misuse. It is concentrating intelligence in a few firms and then giving those firms government-backed authority over the market. That political-economy pattern—incumbents converting legitimate concerns into barriers to entry—most shapes my view.

A demonstrated technical risk that competition, liability and defensive AI could not contain—and that reliably worsened with broader access—would change my view most. Not a speculative scenario, an incumbent-funded warning campaign or a demand that we trust approved gatekeepers, but clear evidence of severe harm that capable developers could not control and ordinary legal accountability could not deter. That could justify narrower restrictions on a specific capability or deployment. It still would not automatically justify a broad American pause, an antitrust exemption or an incumbent-run licensing cartel. I would ask whether the proposed remedy actually reduces the risk, whether it can be enforced internationally and who gains market power from it. Likewise, if open models were shown consistently to create catastrophic harms without providing meaningful defensive or competitive benefits, I would revise my support for openness. But the burden should be evidence and a remedy matched to the demonstrated harm—not fear used as a blank check for centralized control.