Diffusion as Strategy? Export Controls, Open Weights, and the Global Reach of Chinese AI Models, 2018–2026
This study analyzes 3,034 AI models from 2018 to 2026 to find that while US export controls did not immediately force Chinese developers to open their model weights, they did catalyze a strategic shift toward global diffusion, evidenced by a dramatic rise in derivative models built on Chinese bases and a widening frontier compute gap after late 2024.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the race to build smarter machines, a quiet debate has taken hold among policymakers and technologists. For years, the United States has restricted the sale of the most powerful computer chips to China, hoping to slow the country's progress in artificial intelligence. The logic was straightforward: without these specialized tools, Chinese developers would struggle to train the largest and most capable AI systems. Yet, a different story began to emerge from the other side of the Pacific. Instead of falling behind, Chinese developers started releasing their AI models in a way that allowed anyone to download and study the underlying code. This shift sparked a theory: perhaps the restrictions forced China to change its strategy, trading raw power for widespread adoption. If true, this would mean that a policy designed to block a rival's progress might have accidentally helped that rival's technology spread across the globe.
To test this idea, a researcher named Alfredo Merlet examined the actual history of thousands of AI models released between 2018 and 2026. He gathered data on over 3,000 models, tracking who built them, whether their internal "weights" (the core settings that make the AI work) were kept secret or shared, and what kind of computer chips were used to train them. By comparing the behavior of American and Chinese developers before and after three major rounds of US restrictions in 2022, 2023, and 2024, Merlet looked for a clear turning point. He wanted to see if the moment the chips were cut off, Chinese developers immediately started opening their doors, or if something else was driving the change.
The results tell a more complicated story than the simple narrative of cause and effect. It is true that by the end of the period, Chinese models were significantly more likely to be shared openly than American ones. However, this shift did not happen immediately after the first two rounds of restrictions. The data shows that the gap between the two nations did not widen until late 2024 and throughout 2025. Furthermore, the change was not just about Chinese developers opening their models; it was also about American developers closing theirs. The timing suggests that the move toward openness was a strategic choice that solidified in 2025, likely influenced by the success of specific new models and changing business decisions in the US, rather than an instant reaction to the chip bans.
While the link between the bans and the decision to share code is not as direct as some had hoped, the story of how these models spread is undeniable. Merlet found that Chinese AI foundations became the building blocks for a massive portion of new AI work worldwide. Before 2023, only a tiny fraction of new models were built on Chinese bases. By 2025, that number had surged to 51 percent. This means that more than half of the derivative models documented in the database—those created by developers in the US, Europe, and elsewhere—were built upon Chinese foundations. This diffusion happened even among American developers, with nearly one in five new models in 2025 relying on a Chinese starting point.
The study also looked at the hardware and computing power behind these systems. The data indicates that the gap in raw computing power between the largest American and Chinese training runs, which had been narrowing for several years, began to widen again after the final round of restrictions in late 2024. This suggests that the bans may have successfully slowed the pace of the very largest, most advanced models. At the same time, Chinese developers increasingly reported using chips designed to comply with the restrictions, showing an adaptation to the new rules. Yet, despite these constraints on the cutting edge, the reach of Chinese AI grew stronger.
Ultimately, the research confirms that Chinese AI models have become a dominant foundation for the rest of the world, but it challenges the idea that export controls were the sole driver of this outcome. The openness of Chinese models appears to be the result of a strategic shift that happened later than the bans and was reinforced by the success of specific open models. The restrictions may have kept China from training the single biggest models, but they did not stop their technology from becoming the standard upon which the rest of the world builds. The lesson for the future is that in the age of artificial intelligence, controlling the supply of chips might limit the size of a model, but it does not necessarily control where that model ends up or how widely it is used.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.