The United States has signaled potential sanctions against Chinese developers of open-source artificial intelligence (AI) models amid allegations that these platforms are built using stolen American intellectual property. U.S. Treasury Secretary Scott Bessent warned on Tuesday that while the current administration supports open-source models, it will act against unauthorized use of U.S. AI technology. He cited the presence of watermarks from U.S.-based large language models in Chinese AI systems as evidence of intellectual property theft.

This issue highlights ongoing tensions between the U.S. and China in the rapidly evolving AI sector. OpenAI and Anthropic have accused Chinese firms of leveraging American AI developments without authorization, employing a process known as distillation to reduce costs. These accusations come as Beijing pushes forward with its own AI advancements, exemplified by the recent release of Moonshot AI’s Kimi K3 model. Industry experts note that Kimi K3 has significantly narrowed the technology gap, with development lag measured in weeks rather than years.

Reports indicate that China is also considering tighter export controls on AI technologies and semiconductors, aiming to prevent the transfer of sensitive technologies to Western countries. China’s Ministry of Commerce has engaged with leading AI and chip companies to explore regulatory measures, reflecting Beijing’s focus on maintaining control over its technological assets.

Chinese AI developers typically release open-weight models that allow users to download and modify system parameters, in contrast to the predominantly closed and proprietary U.S. models. This approach enables greater user customization but raises concerns over security and intellectual property protections. Moonshot AI, which is planning a fundraising round potentially valuing the company as high as $50 billion, has drawn attention for technology that reportedly offers cybersecurity capabilities close to OpenAI’s GPT-5.6 Sol model at a fraction of the cost.

Separately, Zhipu AI, operating under the name Z.ai, has developed a large-scale AI data center powered entirely by Chinese-made chips. This move is part of a broader strategy to reduce reliance on foreign processors, including those produced by Nvidia. Z.ai’s recent acquisition of infrastructure software firm XCore Sigma, a spin-off from the Chinese Academy of Sciences, further signals an expansion of domestic capabilities. Despite these advances, Z.ai shares have declined recently amid competition from Moonshot’s Kimi K3.

Shifts in AI commercial models also are anticipated. Xu Li, head of Chinese AI company SenseTime, suggested that AI usage will move from a "token economy"—where users pay based on computing power consumed—to a "task economy," where users pay for specific services such as architectural reviews or video generation. This evolution is expected to be driven by declining computing costs and the commoditization of foundational AI models and infrastructure.

Concerns over AI-driven cybersecurity risks have prompted U.S. agencies to proceed cautiously. The Mythos model, developed by Anthropic and restricted initially to select U.S.-based firms and government agencies, demonstrated the ability to autonomously identify and exploit vulnerabilities in cyber architecture. This capability led to an emergency briefing involving the U.S. Treasury and Federal Reserve earlier this year, and triggered shifting policy discussions on AI oversight.

Following a period of regulatory reticence, the U.S. has taken initial steps toward AI governance. President Donald Trump reversed a previous executive order on AI regulation upon his return to office and later issued his own AI executive order in June, calling for voluntary testing of frontier models by security agencies.

China maintains strict regulations over AI development, including security reviews and adherence to “core socialist values.” However, once open-source models are publicly available, users with technical expertise can modify them, raising potential security concerns domestically and internationally.

Experts note that both the U.S. and China could soon face similar challenges as AI capabilities advance, potentially necessitating stronger regulation. According to the Institute for AI Policy and Strategy, China may develop AI models comparable to the U.S. Mythos platform as early as February 2027, underscoring the rapid narrowing of the AI technology gap.

Amid these competitive pressures, the United States and China have agreed to initiate discussions on AI safety. Treasury Secretary Bessent is slated to lead the U.S. delegation, with the first round of talks planned ahead of Chinese President Xi Jinping’s visit to Washington in September. Observers suggest that early dialogue could include agreements on human control over nuclear weapons, reflecting growing concerns about the intersection of AI and global security.