America’s lead in artificial intelligence will not be secured by building the world’s best model and locking it behind an API. It will depend on how quickly American innovation spreads across companies, developers and industries.
That is why open-source and open-weight AI must be treated as a strategic priority. Chinese models are getting better quickly. They may not yet match the best American closed models, but increasingly they can do the job at a fraction of the price.
Research by Georgia Tech found that open models achieve roughly 90 percent of closed-model performance when released and can close the remaining gap within about 13 weeks. A year earlier, they had taken 27 weeks. Inference can also cost dramatically less, about $0.23 per million tokens compared with $1.86 for closed models. Researchers estimated that reallocating demand between models could save the global AI economy roughly $25 billion annually. The implication is difficult to ignore: an eightfold price premium is hard to defend when the quality advantage lasts only three months.
OpenRouter data shows Chinese open models rising from virtually no share in late 2024 to nearly 30 percent. If American companies cannot provide competitive open alternatives, users around the world will not wait. They will use models developed elsewhere.
That makes open AI a question of American competitiveness.
Critics often raise security as a reason to restrict powerful open models. But openness can also be a security advantage. Open models can be inspected, modified and deployed locally, allowing organizations to keep sensitive information inside their own infrastructure.
Hugging Face recently demonstrated this. After a cyberattack, its security team needed to process more than 17,000 attacker events. A closed model’s safety guardrails refused to perform the task. The team instead used an open model, reconstructing the intrusion within hours while keeping sensitive forensic data inside its own environment.
The open model was not the vulnerability. It was part of the defense.
Sweeping restrictions could produce the opposite of what policymakers intend: weakening defensive capabilities while concentrating frontier AI in a handful of closed-model companies. A recent open-weights letter signed by major technology companies warned that blanket restrictions could “weaken competition” while leaving "critical technology" in only a few providers.
There is a broader principle at stake: how artificial intelligence learns.
Some policymakers are interested in restricting model distillation, the process by which one model uses another’s outputs to improve its performance. But distillation is part of how knowledge advances. Scientists study previous discoveries. Engineers build on existing technologies. AI development is no different.
That does not mean anything goes. Companies that illegally extract proprietary information, violate contracts or abuse another firm’s services can and should face targeted remedies. Restrictions should punish misconduct, not learning itself.
AI leadership is increasingly measured in months, not decades. Open models that once needed more than half a year to approach leading closed systems are now doing so in roughly three months.
Foreign competitors face no obligation to slow down because Washington does.
If the United States restricts American developers from learning from models, distributing open systems or building on observable information while competitors abroad continue doing all three, it will only handicap its own ecosystem.
America’s greatest technological advantage has long been an environment in which discoveries spread, entrepreneurs experiment, competitors challenge incumbents and ideas move quickly from laboratories into the economy. AI should be no different.
America should protect national security, enforce contracts and punish genuine theft. But it should resist turning those concerns into a broader policy of technological containment.
The central question in the AI race is not simply who produces the most powerful model first. It is whose technology becomes the foundation on which everyone else builds.
If Chinese open models become cheap, capable and ubiquitous while America’s most advanced technology remains concentrated behind a few companies, technical superiority will not necessarily translate into lasting leadership.
The country that wins the AI race may ultimately be the one that diffuses intelligence the most.
Mario Ottero is the Emergent Technology Policy Analyst at Americans For Prosperity. Previously, he worked as a policy fellow at the Information Technology and Innovation Foundation (ITIF), where he worked on the intersection of innovation, trade, deregulation, and tech. Also, he worked at Cato Institute with the Human Progress team. He holds a Master of Science in Public Policy and Management from Carnegie Mellon University and is a Fulbright-CFI scholar. He is passionate about advancing economic freedom, supporting entrepreneurship, and promoting evidence-based policymaking.