A potential US ban on Chinese open-weight artificial intelligence models could cost American businesses up to $12 billion a year, according to a new study reported by the South China Morning Post that highlights the growing reliance of US developers on low-cost Chinese AI infrastructure. The findings come as Washington reportedly considers new restrictions on foreign open-source AI models, citing national security concerns.
How the $12 Billion Figure Was Calculated
The study, conducted by Daniel Yue, an assistant professor at the Georgia Institute of Technology’s Scheller College of Business, analyzed usage data from OpenRouter. This New York-based platform aggregates access to various large language models. Yue examined data from July 21 to July 27, 2026, to estimate the financial impact of forcing users to migrate from Chinese open-weight models to top-tier proprietary alternatives.
According to Yue’s analysis, if OpenRouter users were forced to switch to comparable proprietary models, the platform’s users alone would face an additional $2 billion in annual inference costs. When extrapolated to the broader US economy, the total cost could range from $3 billion to $12 billion per year, depending on the extent to which American businesses rely on Chinese open-weight models.
Yue emphasized that the figures are an “order of magnitude” approximation rather than a definitive projection, noting that OpenRouter captures only a fraction of the global market for LLM inference. However, the study underscores the significant cost advantage offered by Chinese models. Companies like DeepSeek, Moonshot AI, and Alibaba have aggressively priced their open-weight models, undercutting US rivals like OpenAI and Anthropic by substantial margins.
The Cost Advantage That Drove the Dependency
The cost disparity has driven widespread adoption of Chinese models among developers and startups seeking to deploy AI at scale. For many businesses, the cost of running proprietary US models is prohibitive, making open-weight alternatives a critical component of their AI strategy. A ban on Chinese models would force these companies to either absorb higher inference costs or scale back their AI deployments.
The debate over open-source AI has divided the US tech industry and policymakers. Proponents of a ban argue that allowing Chinese models to proliferate poses a national security risk, potentially giving Beijing access to sensitive data or enabling the development of malicious AI applications. Opponents counter that open-source development accelerates innovation and that banning Chinese models would harm US competitiveness by raising costs for domestic businesses.
A Cost-Benefit Problem With No Easy Answer
As Washington weighs its options, the study serves as a stark reminder of the significant economic stakes involved. The US-China AI race is increasingly defined not just by raw model performance, but by the cost of deployment. If American businesses are cut off from low-cost Chinese models, the financial burden could slow the adoption of AI across the US economy, potentially undermining the very technological leadership Washington seeks to protect.
The timing of the study is significant. Washington is simultaneously considering a range of measures to restrict Chinese AI, from targeted bans on specific models to broader prohibitions on open-weight AI from adversarial nations. Each of these measures carries a different economic cost and a different level of enforcement complexity. Open-weight models, by definition, can be downloaded, modified, and redistributed, making them far harder to ban than proprietary API-based services.
For US policymakers, the challenge is to design restrictions that meaningfully reduce national security risks without imposing prohibitive costs on the domestic tech sector. The $3 billion to $12 billion annual cost estimate provides a concrete benchmark for that cost-benefit analysis. Whether Washington decides that the security benefits of a ban outweigh its economic costs will depend on how seriously officials assess the threat posed by Chinese AI models, a question that remains deeply contested within the intelligence and policy communities.
Yue’s study also raises a broader point about the structure of the global AI market. EastFrontier has previously reported on the US-China AI sanctions dispute that threatened to derail September safety talks. The cost advantage of Chinese open-weight models is not merely a function of lower labor costs or state subsidies. It reflects genuine engineering innovation in model architecture and inference optimization. Addressing that advantage through policy restrictions alone, without investing in making US alternatives more cost-competitive, risks treating the symptom rather than the cause.
The study is a timely reminder that in the AI race, economic competitiveness and national security are two dimensions of the same challenge, and a policy that ignores one while pursuing the other is unlikely to succeed. As the debate over a potential ban intensifies in Washington, Yue’s $12 billion figure is likely to become a reference point that both sides invoke, supporters of a ban to argue the cost is manageable and opponents to argue it is not.