The U.S. May Need More Open Models, Not Fewer

“Is the future of AI open?” asks a headline in The Washington Post. The question is becoming increasingly relevant as open-source models—which allow anyone to download, inspect, run, study, and modify the model—become more capable and widely adopted.

While closed models, accessed primarily through apps and application programming interfaces (APIs), have long dominated the artificial intelligence (AI) market, open models have historically fallen behind in capability and adoption. That dynamic, however, is beginning to change. The launch of DeepSeek and Kimi K3—Chinese open models that outperform several U.S. models on key benchmarks—has turned open models into a serious competitiveness issue and changed the nature of the open-versus-closed AI debate.

Many used to prefer closed models largely on the assumption that they were safer; but as Chinese open models become increasingly embedded in global businesses and approach frontier-level capabilities, they are introducing a new set of tradeoffs. U.S. policymakers are also paying closer attention to the claimed risks posed by frontier AI, while major industry players, including Microsoft and Nvidia, have publicly defended the importance of preserving open models.

These developments complicate the old open-versus-closed framing. The real question isn’t just whether AI should be open or closed, but how the growing capabilities and global reach of these models are reshaping the balance between innovation, competitiveness, and national security.

Notably, open models offer many benefits, including lowering the barriers to entry and costs for businesses to adopt AI, promoting AI experimentation and innovation and fostering competition at the model-development level. Given their lower costs and improving capabilities, open model adoption is growing, and the market reflects this trend. On OpenRouter, an LLM platform providing a single interface to switch between open and closed models, weekly usage of open models has grown from 1 trillion tokens in September 2025 to about 80 trillion today. Open model usage is also growing in high-value industries, including in legal and telecommunications. Latham & Watkins, for example, has engineers tailoring existing open models to its needs, while AT&T reported that open models currently account for 25 percent of its overall AI usage, and expects them to reach 80 percent over time. These adoption decisions are driven in part by lower costs and greater control over data.

However, one potential drawback of open models is security. Once a new user downloads, changes, and uses an open model, the original developer loses control over its applications, which can lower the barriers to misuse. A second commonly cited concern is dependence on Chinese AI technology. As Chinese models are increasingly adopted by U.S. and foreign businesses—studies show that 80 percent of developers building with open models are using Chinese models—policymakers are questioning whether their benefits justify the potential national security risks. While Congress remains divided on how to address open models, two House committees have already opened an investigation into the security implications of U.S. companies integrating Chinese open models, and the Trump administration is considering measures to ban U.S. access to Chinese open models. Any ban, however, must be narrowly tailored to address demonstrated risks, rather than a broad-based boycott of foreign products.

The concerns go beyond safety. Open models are critical tools for businesses and researchers innovating with AI, and their global adoption has the potential to determine which countries’ technologies shape the global AI ecosystem—and who has greater influence over its development. China’s growing leadership in open models raises the question of whether restricting Chinese open models makes sense if those restrictions also affect U.S. companies and researchers who use them. The broader challenge is how to address the security risks of increasingly capable AI without undermining the innovation and competitiveness that the United States needs to maintain its lead.

With rising tensions, one point is becoming clearer: neither open nor closed models are inherently better or worse. Their benefits and risks depend largely on how they are developed and deployed. Closed models, after all, can also be misused in ways that are difficult to detect.

Ultimately, if open models are going to shape the diffusion of AI, the path forward should not simply be to restrict them, but to build a stronger U.S. open-model ecosystem through a light-touch regulatory approach. This would give American companies more model alternatives and reduce reliance on models produced by geopolitical adversaries without losing the benefits of AI innovation.