The U.S. government and American tech companies are in the throes of a debate over whether developers of artificial intelligence should be permitted to publicly release the “weights” — the rules or numerical calculations that determine how large language models “learn,” process information, and make predictions — of their most advanced AI models. The concerns behind that debate are real. Those who favor open-source AI or so-called “open weights” argue that it speeds development of all AI systems and software built on them and can help crowdsource security issues and other flaws. But opponents fear that open models could also be “fine-tuned” for purposes their developers never intended, possibly nefarious uses, and that whatever safeguards are built into a proprietary system evaporate once the weights are public.
Yet the current debate relies on assumptions about technology and innovation – and where the “chokepoints” are that safeguard national security and drive economic growth – that may no longer be helpful for shaping policy. Indeed, for the better part of a decade, American technology strategy has been built around a deceptively simple idea: preserve U.S. technological leadership by controlling access to the technologies that make it possible. Export controls on advanced semiconductors, restrictions on AI computing, investment screening, and sanctions lists all reflect this logic. If the United States and its allies control the critical chokepoints of the global innovation economy, they can slow China’s rise and preserve America’s strategic edge.
It is a coherent strategy. It is also increasingly a strategy designed for a world that no longer exists.
The Limits of U.S. Strategy
China’s advances in open-weight AI have exposed the limits of a U.S. strategy built too heavily on the idea of denying them access to American technology and American markets. Despite U.S. export restrictions, Chinese firms have continued to release increasingly capable models, and those models are spreading through the global ecosystem in ways far harder to contain than physical supply chains. On Hugging Face, an open-source community for AI tools, models, and platforms, China has surpassed the United States in monthly and overall downloads of the models that developers use to build AI for their own systems and enterprises, with Chinese models accounting for 41 percent of downloads over the past year. Alibaba’s Qwen Family, a suite of generative AI models built by Alibaba Cloud, has generated more than 200,000 derivative models, a measure of ecosystem power as much as of technical quality. While downloads alone do not translate into products, revenue or enterprise deployment, it does mean that, increasingly, the standards and architecture of AI are being built on Chinese models.
None of this makes export controls obsolete. In sectors such as advanced semiconductor manufacturing equipment, they continue to impose meaningful costs and delays where necessary for protecting uniquely sensitive military technologies and for preserving leverage over the most critical supply-chain chokepoints. What such controls can no longer do is determine who leads in artificial intelligence or across a range of other emerging tech. Export controls may buy time and complicate an adversary’s path to the frontier. But when innovation is diffused around the globe — through open-weight releases, fine-tuning communities, model adaptation, and developer networks that span every continent — it is an approach that is unlikely to lead to success.
Yet much of the current debate still approaches AI competition as a problem of restricting the spread of AI models, technology, and hardware. The question is whether the United States will remain the central platform for global AI innovation, adoption, and standards-setting – and with all the national security implications that implies. Measured against that question, simply shutting down American open-weight development risks getting the answer backward. In fact, with open-source AI likely to become a foundational layer of global digital infrastructure, such an approach likely could not slow Chinese progress, but in fact would redirect the global spread of open AI toward Chinese models.
Most governments, universities, startups, and public-interest institutions will never have the resources to build frontier models from scratch, and many cannot afford to depend entirely on expensive proprietary systems. They will build on open models, adapt them to local languages and legal frameworks, and integrate them into their own software, public services, and industries. As they do so, it will make the United States a less obvious place for AI communities to gather and build, and it will thin out domestic competition, weaken startups, and narrow downstream experimentation. Whichever country’s models become the building blocks for global AI architecture will collect the commercial returns and exercise enduring influence over security, standards, interoperability, and trust.
Beijing appears to have grasped this before Washington. Chinese firms have moved aggressively into open releases, and the effects are visible in downloads, derivatives, customization, and geographic reach rather than in benchmark scores alone. The spread of Chinese models is a distribution story and a platform story, and, uncontested, it will eventually become a geopolitical story about whose tools the world uses.
What is happening in AI is happening elsewhere. Chinese firms lead in drones and battery chemistry, have reordered the electric vehicle industry, and set the pace of commercialization in fields even where American labs still hold the research edge. Innovation now emerges simultaneously from Silicon Valley, Shenzhen, Seoul, Taipei, Bangalore, and Munich, with “borrowing” and reverse-engineering accounting for less and less of Chinese industry with each passing year, and as China’s own scientific, research and tech innovation ecosystem is blossoming. A strategy premised on a single frontier, with a manageable set of gates behind it, was reasonable when American technological leadership was the unquestioned order of things, with the rest of the world following with adoption. That world produced the chokepoint approach to technological power, and that reality is fast receding.
Attracting – or Repelling – the Innovators
The measure of success, then, is no longer whether China can acquire a particular chip or replicate a particular model. It is whether the world’s best engineers still want to build companies in the United States, whether allied governments adopt American technology standards, whether American universities remain magnets for global scientific talent, and whether venture capital, cloud infrastructure, trusted supply chains, and research partnerships reinforce one another across a coalition of advanced democracies. Those questions are not just about technology strategy or trade policy, narrowly construed, but also touch on immigration, federal research funding, higher education, industrial policy, and alliance management. Yet Washington still debates each in isolation.
Under the Trump administration, the United States has become adept at wielding coercive economic tools but much less successful at articulating the affirmative strategy those tools are meant to serve. Export controls delay competitors. By themselves, however, they generate no scientific breakthroughs, attract no entrepreneurs, strengthen no alliances, and build none of the industries that will define the next generation of technological leadership. The proposal to restrict American AI developers’ open-weight releases applies the administration’s instinct for coercion to a problem that calls for an affirmative approach.
Enduring technological advantage has never rested primarily on exclusion. America’s deepest strengths, openness and dynamism, world-class universities, deep capital markets, an unmatched capacity to draw talent from everywhere, are ecosystem advantages, and far harder for competitors to replicate than any single technology.
The answer to open-weight competition is not to stop competing in the open. America needs a broader theory of technological power, one that recognizes that, in the age of open-source AI, leadership will belong to the country whose innovation ecosystem the world most wants to join.
