Supermodel race narrows between Washington and Beijing: industrial distillation accusations reveal limits of chip restrictions

The pace of the technology competition between Washington and Beijing is accelerating as the knowledge gap between American models and their Chinese counterparts narrows, marking a strategic shift that puts Western export controls in front of a real test. Despite the strict ban imposed by the U.S. administration on China's access to the latest chips and semiconductor manufacturing equipment, specialized international reports, including a report from the Stanford Institute for AI, indicate that China has succeeded in establishing itself as a genuine balancing weight to the United States, and the gap between the capabilities of the two sides' models is now described as narrow and fragile.

Chinese presence is no longer limited to theoretical research; startups and giants have turned it into products that have put Silicon Valley on alert. The company “MonShot” launched its “Kemi K3” model, which quickly rose to third place globally in model intelligence according to a specialized analytics platform, before successive updates pushed it down to ninth. At the same time, “DeepSec” announced a preview of its “V4” model with advanced capabilities in knowledge, reasoning, and independent agent planning, after its previous model sparked a cost-efficiency revolution compared with American models. The competition did not stop there, as “Z.AI” released “GLM 5.2” and “GLM 5.3”, which improved programming and agent work capabilities, while “Alibaba” introduced the “Qian 3.8 Max” model to compete with the latest generations of “Claude” and “GPT”.

The most notable shift in Washington is the realization that a chip ban alone is no longer sufficient to protect technological superiority.Faced with this, U.S. security agencies, led by the Federal Bureau of Investigation, the National Security Agency, the Cybersecurity Agency, and the Infrastructure Security Agency, have leveled explicit accusations at Chinese developers of engaging in what they described as “industrial-scale distillation,” alleging the extraction of advanced American model capabilities through organized and systematic targeting. The company “Anthropic” joined these accusations, blaming Chinese startups for rerouting user queries to their models. While Beijing rejected these claims, asserting that distillation is a training technique employed across the entire AI sector and accusing Washington of seeking to monopolize the industry, technology policy analysts see U.S. research centers shouldering costly computing bills to push development boundaries, whereas Chinese models absorb those capabilities and later release them as open-weight models that spread worldwide.

Conversely, the Chinese arena is not free of deep security concerns reflected in warnings from the Ministry of State Security; Minister Chen Yixin warned in an official article that AI carries risks that affect political, institutional, and ideological security when exploited in wide-scale disinformation campaigns through deep-fake technology, explicitly cautioning about the ability of Western generative models to uncover cyber vulnerabilities and use them in complex intrusion operations, which drives Beijing’s leadership to combine accelerating technological self-sufficiency with a demand for an international framework to govern and regulate risks.

This fragmentation of supremacy in high-end models imposes new engineering and economic calculations on technology companies and institutions in the Gulf, Egypt, and the Levant.Regional institutions that are building their systems are no longer forced to limit their options to high-cost American APIs; instead, they now have a practical alternative in highly efficient open-weight models coming from the East, which can be hosted directly within local data centers to ensure full data sovereignty and substantially lower inference expenses. At the same time, the shift in U.S. focus toward monitoring model outputs and pursuing distillation operations requires development teams in the region to conduct strict vetting of synthetic data sources and specialized model training pipelines, avoiding any technical restrictions or licensing pursuits that closed platforms might impose on query flows.

This development shifts the AI equation from a mere race to arm data centers with silicon to a strategic battle over the transfer of programming knowledge through inference outputs. While major powers prepare for new rounds of political and regulatory negotiation, the need to diversify model infrastructure within the region and avoid single-source dependence on any provider, regardless of its computing capabilities, becomes clear.