The United States should strengthen its AI position by expanding computing infrastructure, retaining technical talent and supporting open-weight development rather than relying primarily on restrictions targeting China, a new policy commentary from the Center for European Policy Analysis argues.
Elly Rostoum, a senior resident fellow at CEPA, points to the emergence of capable Chinese systems from Moonshot AI and DeepSeek as evidence that export controls may delay China’s progress but cannot permanently prevent it. The commentary contends that Washington should use that time to increase domestic semiconductor production, electricity generation and grid capacity while maintaining a competitive research ecosystem.
The argument centers partly on Moonshot AI’s Kimi K3, an open-weight model described as having 2.8 trillion parameters and performing strongly against proprietary systems on a prominent coding benchmark. Moonshot founder Yang Zhilin studied at Tsinghua University, earned a computer science doctorate from Carnegie Mellon University and previously worked as an intern at Google Brain and Meta. His decision to establish an AI company in Beijing illustrates the international competition for researchers who have experience in both Chinese and US institutions.
DeepSeek offers a related example. Its founder, Liang Wenfeng, studied at Zhejiang University and built a research operation around talent developed in China. The release of DeepSeek-R1 showed how Chinese laboratories can narrow performance gaps through engineering, model architecture and training techniques even while facing tighter access to advanced chips.
Rostoum’s commentary challenges efforts to treat model distillation and intellectual-property theft as interchangeable. Distillation generally uses the outputs of one system to help train or improve another, although disputes can arise over authorization, terms of service and the provenance of training material. The policy paper argues that suspected misappropriation should be handled through focused legal and commercial measures rather than broad limits on a training method used across the industry.
That position echoes a joint industry letter signed by Nvidia, Meta, Microsoft, Dell, IBM, Hugging Face and more than 20 other companies, with OpenAI adding its name later. The coalition opposed a general ban on releasing model weights, arguing that closed systems are not automatically more secure and that independent scrutiny can help researchers identify weaknesses.
Open-weight releases give developers access to the numerical parameters learned during training, allowing them to run, modify and study models without depending entirely on a provider’s hosted service. They can broaden research and commercial participation, but they also complicate efforts to withdraw a capable system or centrally enforce safeguards after release. That tension has made openness a central issue in AI governance, with policymakers weighing transparency and competition against the risk of misuse.
The CEPA commentary does not dismiss national-security concerns about advanced Chinese AI. Instead, it argues that both democratic and authoritarian governments are likely to obtain powerful systems, making shared risk controls necessary even amid strategic rivalry. Proposed measures include common testing practices, incident-reporting channels, crisis communications among governments and laboratories, verification mechanisms and agreed boundaries around especially dangerous capabilities.
Such arrangements would not require broad political trust, Rostoum argues. They would rest on the narrower premise that neither country benefits from an uncontrolled AI incident with effects that cross borders or disrupt global supply chains.
The recommendation leaves export controls as a tool for preserving a temporary hardware advantage, not as a complete strategy. The United States and its partners continue to hold strengths in advanced chip design and fabrication equipment, while access to power and data-center capacity is becoming an increasingly important constraint on model development. Under the proposed approach, Washington would pair carefully targeted controls with larger investments in fabrication, energy infrastructure, computing capacity and scientific talent, while avoiding domestic rules that could prevent US laboratories from releasing competitive open models.
Sources: AI safety policy