Policymakers Overlook How Open Source AI Is Reshaping Global Power

For the past several years, policymakers have been preoccupied with an AI competition narrative defined by closed “frontier model” development in the United States and China. Regulation, strategy documents, and public debate revolve around which country's corporations can train the most capable proprietary system, who controls the most compute, and how national champions stack up against one another.

Yet when we examine empirical evidence of how AI is actually being adopted and integrated into products and services, such as in our study on longitudinal data on global model downloads from the Hugging Face ecosystem, a different reality quickly becomes apparent. This open layer of development captures a far wider range of applications and often anticipates where the broader AI market is headed.

Work from the nonprofit research institute Epoch AI has shown that the performance gap between open and closed models is narrowing, even as open models remain dramatically cheaper to build and deploy. The real contest for influence is not unfolding behind closed doors, but in the open-source ecosystem. Open source and open-weight models form the substrate on which millions of developers, startups, and public institutions now build. They shape the languages supported in downstream applications, the assumptions embedded in everyday tools, and the way AI capabilities diffuse across borders. Because they are widely accessible, they also provide a rare empirical window into the global dynamics of AI adoption. By examining model download data, we can observe, rather than merely speculate, how influence in AI is shifting across nations, companies, and communities.

What emerges from this view is a world that is already rapidly shifting. In the early 2020s, American companies effectively owned the open source landscape. At one point, well over half of all open-weight downloads were tied to US industry models such as BERT, CLIP, T5. But that dominance has steadily eroded. Subsequently, the revolution of highly-performant early diffusion and speech systems led to a proliferation of unaffiliated developers, hobbyists, and artists, taking over much of the development from American industry. Since early 2025, China’s presence has expanded at remarkable speed, driven by the rapid ascent of companies such as DeepSeek and Alibaba, whose models have become global defaults almost overnight. Their rise is not simply a matter of national representation; it reflects a broader shift in the type of models that are gaining traction, such as large-scale reasoning architectures, mixture-of-experts systems, multimodal and video-generation models, and aggressively quantized networks optimized for real-world deployment. These organizations also release model variants and inference code at a staggering cadence, to support a wide variety of users.

Europe plays a quieter role, but an important one. Its footprint in open-weight downloads is smaller, yet more pluralistic: universities, nonprofits, and distributed research groups make up a far greater share of European contributions than they do in the United States or China. This is an under-appreciated strength. While Europe does not currently compete on raw scale, it contributes heavily to the public-interest infrastructure of the ecosystem: the tools, adapters, training resources, and scientific work that sustain open innovation. In a multipolar AI environment, this form of participation matters greatly.

Perhaps the most striking trend, however, is the rise of unaffiliated developers and loosely organized online communities. The open source ecosystem is no longer shaped primarily by large companies but by hobbyists, independent researchers, small collectives, and new intermediary groups that specialize in repackaging, quantizing, and adapting models for widespread use. These intermediaries increasingly determine which models become practical options for ordinary developers. A single repackaged or quantized release can influence adoption patterns as much as a major base model. This is a new center of gravity in the AI economy, one that barely features in policy conversations.

Underlying these shifts is a deeper structural tension. Even as open-weight models proliferate and diversify, transparency is collapsing. In 2022, the majority of downloaded models disclosed something meaningful about their training data. By 2025, that fraction had fallen below 40 percent, and for the first time, downloads of opaque open-weight models outnumbered downloads of models that meet basic open source criteria. Licensing is becoming more restrictive, model gating more common, and the gap between the rhetoric of openness and the reality of access is widening. This decline in transparency coincides with rising geopolitical stakes; yet the policy response often assumes that “open source” is a stable, inherently accessible category rather than a space undergoing rapid commercialization and fracturing.

The result is an AI ecosystem that is evolving in ways current governance frameworks do not yet recognize. If policymakers continue to focus narrowly on frontier training runs or compute expenditures, they will miss the more dynamic, diffuse, and globally distributed layer where real adoption occurs. The open source ecosystem is where influence is being negotiated: not just which models exist, but which are actually used; not just who can train a trillion-parameter network, but who can make it deployable, modifiable, and relevant across languages, cultures, and industries. It is also where power can shift fastest. The rapid rise of Chinese open-weight models shows that leadership in this space is not fixed and can be reshaped within a single model generation.

Recognizing this requires a conceptual shift. Open source AI is not the opposite of geopolitical competition; it is one of its primary arenas. Nations that ignore this layer risk misunderstanding both their own dependencies and their own leverage. The models that dominate open-weight adoption will shape global norms, software ecosystems, and economic opportunities just as profoundly as proprietary frontier models, and perhaps more so. They will determine which languages are well-supported, which cultural assumptions are encoded, which countries’ companies set technical defaults, and which developers can meaningfully participate in the AI economy.

If we want to understand the distribution of power in AI, and if we want to shape it, the open source ecosystem is where we must look. Our work is focused on developing technologies that detect and monitor emerging trends and concentrations of power before they become widely recognized by the community.

This post is based on findings from Economies of Open Intelligence: Tracing Power & Participation in the Model Ecosystem by Longpre, Kaffee, Akiki, Lund, Kulkanari, Chen et al. (2025).