--- title: "You Can't Tariff Intelligence" date: 2026-07-24 dateDisplay: "Jul 2026" outlet: forbes originalUrl: "https://www.forbes.com/sites/christiancatalini/2026/07/24/you-cant-tariff-intelligence/" canonical: original rights: full tags: [ai-agi, policy] deck: "Twenty-five companies just asked Washington not to strangle open-weight AI. They're right: AI has a weak appropriability regime, and lobbying won't change that." image: "/images/writing/you-cant-tariff-intelligence.jpg" --- A group of 25 leading tech companies and investors, including NVIDIA, Microsoft, Meta, Andreessen Horowitz, Dell, IBM, and Palantir, is urging Washington not to rush to conclusions on AI regulation. The[letter](https://www.microsoft.com/en-us/corporate-responsibility/topics/open-weight/) is an important wake-up call not to fall for the simplistic and convenient narratives that the AI incumbents have been pushing in recent days to stop the success and diffusion of open-weight AI models. China hawks have taken advantage of the AI debate to compare these open models, which are AI models available for anyone to use and build upon, to historical practices designed to undo American industrial leadership. But the analogy is extremely weak, and the cure might not only be ineffective, but also lead to the exact scenario it is meant to avoid. The simple story compares Chinese open-weight models to "dumping," the practice of selling a product below cost to drive competition out of a market. But China does not get to select the workable business model for AI any more than the United States does. While large countries can try to influence and shape markets, they cannot fight economic "gravity" and the fundamental nature of a new technology. The reality is that AI is subject to a weak appropriability regime, and as much as the leading AI labs would like the problem to go away through lobbying and regulation, the way progress in AI has taken place from the very beginning pushes in the opposite direction. Yes, the government could shield the AI labs from foreign competition the same way it did for AT&T back in [1934](https://x.com/ccatalini/status/2079257647669530637?s=20), but the technology and market are different, and this time we would not just get decades of an inefficient domestic market, but also lose the AI race outright. The letter draws a critical distinction between training on an AI model’s outputs (which is currently protected only by the labs’ terms of service) and unlawful efforts to extract value from proprietary models. The former is a widely used industry practice, and one that has made innovation in AI cumulative and fast. The latter is closer to IP theft and to historical examples of China conducting espionage against GE, Boeing, Micron, and many others. The two behaviors are different and should be addressed with different policy instruments. In fact, the [benefits](https://www.forbes.com/sites/christiancatalini/2026/06/16/nadellas-test-whats-left-when-the-ai-model-is-pulled/) to society of the former outweigh any of the potential costs. The United States currently has a small lead in frontier models, and the companies behind them will likely still be able to monetize them in a number of ways even as competition from open-weight, lower-cost alternatives heats up. Why? Because even a small advantage brings massive economic value in domains such as cybersecurity, biology, and frontier R&D more broadly. That said, as the technology diffuses into a number of workflows, it is also clear that companies will not only be cost-conscious but will also want to ensure that, as they use these new tools, they are not continuously leaking their IP and alpha to a closed lab. These two critical dimensions, cost and [control](https://x.com/ccatalini/status/2070896788408799300), are driving leading enterprises to switch to open models. If the United States makes it harder for US enterprises and startups to use open weights, it will not only drastically slow down their progress in AI, but also make our entire economy uncompetitive relative to the alternative. We should know this well, given that [our choices on internet regulation](https://clintonwhitehouse4.archives.gov/WH/New/Commerce/) are what gave us decades of undisputed leadership and entrepreneurial dynamism. When competition heats up in a sector, the answer is not to ask for protectionism. Historically, while it may give national champions a lifeline, it ultimately backfires unless the goal is just to sustain domestic demand. Here the challenge is completely different: the United States wants to lead in AI to secure not only additional decades of economic growth, but also freedom and choice around the globe when it comes to this new form of critical infrastructure. So if intelligence fundamentally wants to be free, what is the only sensible policy action? It is to embrace the technology for what it is, support cumulative innovation and experimentation—from academia to startups and large enterprises—and help the technology diffuse as fast as possible throughout society. The critics will argue that open weights are dangerous and that we should contain them on safety grounds. The irony is that the argument is completely backwards. We worked through this in our paper on the [economics of AGI](https://arxiv.org/html/2602.20946v2), and it is now clear, especially after the OpenAI-Hugging Face cyber incident, that [open weights favor defenders](https://x.com/ccatalini/status/2078643595742228511?s=20) over attackers. Historically, security by obscurity never worked, and it is even less likely to succeed in a world where information diffuses and is recombined at ever faster speeds. As the signatories point out, "openness may be one of the most important paths to AI safety and security." As I've written before, paraphrasing Linus's law, given enough eyeballs, all misalignment is likely to be shallow. The intuition is simple: would you rather bet on a small, elite group to find all relevant attack vectors, or on the talent distributed across the rest of the world? The choice is simple: do we want to turn the likes of Anthropic into national champions by dictating from the top down what the industry should look like, or do we want to encourage competition and market forces to figure this out for us? The industry is so early that nobody knows where value will ultimately accrue. And the answer will likely vary by application. Right now, AI models are trying to commoditize entire industries and services, AI routing companies are trying to commoditize the models, and the apps are trying to commoditize routing. While nobody knows exactly how value capture will split, the worst possible choice would be to let the government decide that prematurely. It is un-American not only in spirit and approach, but also because it would severely undermine our leadership in AI.