The US is losing ground to China in open source AI development, creating potential supply chain vulnerabilities as Chinese companies like DeepSeek and Alibaba dominate the open-weight model space. While America leads in proprietary AI systems, experts warn this dependency on foreign open models could become a critical weakness if access is suddenly cut off or restricted.
America's AI dominance story just got complicated. While OpenAI, Google, and Meta race toward artificial general intelligence behind closed doors, Chinese companies are quietly winning the open source battle that could determine who really controls AI's future.
The wake-up call comes from Nathan Lambert, founder of the ATOM (American Truly Open Models) Project, who's watching Chinese models like DeepSeek-R1 and offerings from Alibaba rapidly gain ground among researchers worldwide. "The US needs open models to cement its lead at every level of the AI stack," Lambert told WIRED.
The irony runs deep here. Meta essentially kicked off the open source AI revolution when it released Llama in July 2023, giving developers their first taste of a truly capable model they could download, modify, and run locally. But as the AGI race heated up, American companies pulled back from openness, viewing their most advanced models as too valuable to share.
Meanwhile, China pivoted toward greater transparency. When DeepSeek dropped its R1 model in January 2025, it didn't just match US capabilities - it did so at a fraction of the training cost, sending shockwaves through Silicon Valley. Chinese firms quickly realized that open-sourcing lets them crowdsource improvements from the global research community.
The supply chain implications are starting to worry national security experts. Unlike proprietary models accessed through APIs, open-weight models can be downloaded and run on local hardware - making them essential for companies handling sensitive data or operating in air-gapped environments. If Chinese companies suddenly restricted access or stopped development, American businesses and researchers could find themselves cut off from critical infrastructure.
"Open models are a fundamental piece of AI research, diffusion, and innovation, and the US should play an active role leading rather than following other contributors," Lambert argues. His ATOM Project, launched symbolically on July 4th, is pushing for American intervention in what's becoming a geopolitical chess match.
Stanford's Percy Liang, who signed ATOM's open letter, thinks the US needs to go even further. He's leading the Marin project, funded by Google, Open Athena, and Schmidt Sciences, to create truly transparent models where even the training data is publicly available. "The view that we would get one company to build AGI and then bestow it on everyone is a little bit misguided," Liang said.
The economics make the case even more compelling. ATOM estimates that building and maintaining a competitive open source frontier model would cost around $100 million annually - roughly what Meta CEO Mark Zuckerberg offered individual AI researchers to join his new superintelligence lab.
Andrew Trask from OpenMined is proposing something more radical: a government-backed effort to help companies access nonpublic training data, similar to how ARPANET led to the internet. With potentially 180 zettabytes of untapped data available globally, such an initiative could leapfrog current limitations. China might have an edge here if its government can compel data sharing between companies - something much harder to coordinate in the US.
The competitive dynamics are shifting fast. While OpenAI and Google focus on their next-generation closed models, Chinese companies are building developer ecosystems around their open offerings. Each iteration incorporates feedback from thousands of researchers, creating a development cycle that closed systems can't match.
Zuckerberg's recent comments suggest Meta might not open-source its most advanced future models - a strategic retreat from the very approach that helped establish its AI credibility. Other US companies seem even less inclined toward openness as they chase the AGI prize.
The research community is taking notice. Lambert reports growing corporate interest in backing open-weight frontier model development, though he hasn't named specific companies. The relatively modest investment required makes it attractive compared to the billions being poured into proprietary model training.
The US finds itself in an unusual position - leading in proprietary AI development while falling behind in the open source ecosystem that might ultimately prove more important. With Chinese companies building global developer communities around their open models and $100 million potentially enough to fund competitive alternatives, the question isn't whether America can afford to act, but whether it can afford not to. The next few months will likely determine whether the US treats open source AI as a national security priority or continues letting others shape this critical infrastructure.
