Washington urged to rethink AI strategy as Chinese open-source models advance

The United States’ artificial intelligence advantage, argues a new analysis from the Center for European Policy Analysis, depends less on restricting Chinese access to technology and more on a sustained commitment to competition, compute capacity, and an open innovation ecosystem, a conclusion underscored by an unusual display of unity among Silicon Valley’s fiercest rivals. The debate has been sharpened by the recent release of Kimi K3, a 2.8-trillion-parameter open-weight model developed by Beijing-based Moonshot AI, which reportedly became the first open model to surpass proprietary rivals on a major coding benchmark, a feat that prompted Washington to accuse the company of building its platform on stolen American intellectual property. The founder, Yang Zhilin, a Tsinghua-trained engineer with a Carnegie Mellon doctorate and internships at Google Brain and Meta, had turned down an offer from Apple to return to China, where he established Moonshot AI, a trajectory that echoes that of Liang Wenfeng, whose DeepSeek-R1 model demonstrated comparable reasoning ability to American counterparts and was also made freely available.

The strategy of imposing export controls and restrictions on Chinese semiconductor access, the analysis suggests, has shown its limits, buying American companies only a handful of years of lead time rather than securing permanent technological supremacy, given China’s substantial resources and deep pool of engineering talent. DeepSeek’s Liang, who earned both his degrees at Zhejiang University and deliberately built his research team from domestic graduates, exemplifies the indigenous capability that Washington cannot indefinitely suppress, and the success of two Chinese frontier AI labs indicates that the country’s researchers are simply highly capable, irrespective of the methods they employ. The CEPA piece, authored by Senior Resident Fellow Elly Rostoum, contends that policymakers must confront the uncomfortable reality that advanced AI will be a tool possessed by both democratic and authoritarian governments, and that attempting to silo access will merely accelerate Beijing’s determination to build frontier models without American input.

A more constructive approach, according to the analysis, would involve establishing international rules of the road that apply universally, encompassing verification regimes, testing standards, crisis communication channels between labs and governments, pre-agreed red lines on catastrophic capabilities, and incident reporting mechanisms for cases where both sides share an interest in immediate knowledge. Such an approach, it is argued, does not require trusting Beijing, but rather recognising that a catastrophic AI failure would not respect national borders, creating a mutual interest in prevention that transcends geopolitical rivalry. This perspective has found an unlikely echo in Silicon Valley, where NVIDIA, Meta, Microsoft, Dell, IBM, Hugging Face, and more than twenty other companies, many of them fierce competitors, published a joint letter rejecting a proposed ban on open-weight models, a position to which OpenAI added its signature days later, representing a rare consensus on a matter of profound strategic importance.

The tech coalition explicitly rejected the assertion that Kimi K3 was built on stolen material, drawing a crucial distinction between distillation—defined in the letter as the practice of using one model’s outputs to train or improve another—and unlawful misappropriation, which they argued should be addressed through targeted legal frameworks rather than sweeping prohibitions on a technique employed by every major laboratory. With Chinese and American models now approaching parity, the decisive factor, the analysis contends, is compute, an arena where the United States still holds a significant edge, as Huawei remains constrained by US chip restrictions, and Washington should press its advantage in fabrication, power generation, and grid capacity, as the entity that commands the most computing power will likely determine the winner of the AI race. The coalition further argued that closed AI models are not inherently safer than open ones, as they can be breached, misused, or fail in ways that outsiders cannot detect, and that concentrating advanced capability in a handful of closed systems creates single points of failure rather than eliminating risk.

Open models, by contrast, allow a broad community of researchers to examine behaviour, identify vulnerabilities, and develop safeguards publicly, with the letter asserting that just as open-source software demonstrated that transparency can enhance security, AI safety may depend on enabling more people to test and strengthen the systems on which society relies. Three distinct positions have emerged: the White House’s reported proposal to ban Chinese open-weight models outright; Anthropic’s opposition to the ban but support for curtailing industrial-scale training and mandatory pre-release safety testing for all sufficiently capable models; and the broader industry stance that distillation should remain unregulated and that openness itself constitutes a safety solution. Of these, the analysis concludes, only the ban is genuinely indefensible, as it would cede the open-source ecosystem to China while doing little to prevent the development of advanced capabilities that Washington cannot indefinitely forestall. The other two positions, it suggests, are not far apart, both steering policymakers away from a knee-jerk prohibition, with Anthropic insisting on nuance and the rest of Silicon Valley prioritising openness while treating jailbreak mitigation as a secondary concern.

Beijing, the piece observes, has as much at stake in stable supply chains and in the survival of its own political system, making it an unlikely candidate for deliberately unleashing a catastrophic AI pathogen it could not contain, and if a worst-case scenario is sufficiently grave to warrant caution, it is also serious enough that the United States and China cannot responsibly address it in isolation. The United States, the analysis concludes, cannot continue to restrict its own laboratories from releasing competitive open models while attributing a collapse in American competitiveness to distillation, nor can it out-compete China by containment alone; it must instead out-build its rival through aggressive investment in talent, compute, and infrastructure, a strategy that recognises the global nature of technological progress and the enduring value of openness as a cornerstone of innovation and safety.