ANALYSIS The US-China AI rivalry has entered a structurally new phase: for the first time, both governments are simultaneously constructing export-control regimes aimed at restricting the other's access to AI technology, even as Chinese hardware and model breakthroughs erode the assumptions underpinning Washington's containment strategy.
Why it matters
The decoupling is no longer one-directional. Treasury Secretary Scott Bessent on Tuesday threatened sanctions against Chinese AI companies if their open-weight models are found to have been distilled from American frontier models1,4. Within the same news cycle, reports emerged that China's Ministry of Commerce is consulting domestic AI and semiconductor companies — including Alibaba, ByteDance, and Zhipu — on measures to restrict overseas downloads of core AI model weights and training data, block foreign acquisitions of Chinese AI startups, and limit contract manufacturing of Chinese-designed chips at TSMC and other foreign foundries8,9. ANALYSIS The simultaneity is the signal: both capitals now treat frontier AI as a strategic asset to be hoarded, not a technology to be traded.
The big picture
Bessent's language was precise and pointed. "We've seen a lot of talk about open source models coming and threatening the large language models in the US," he said on Fox Business. "This administration supports open source models, but what we do not support is IP theft. If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft". The statement frames the competitive threat from Chinese open-weight models — most recently Moonshot AI's Kimi K3, which was assessed to have outperformed Anthropic's latest model Opus 4.8 on most benchmarks — as potentially rooted in intellectual-property theft rather than independent innovation.
Beijing's parallel moves carry a different strategic logic. China, which had focused on securing advanced foreign technology, is now transitioning to a stage where it views its own AI models, semiconductors, and robotics as key national assets and seeks to manage even their exports5. The proposed restrictions could eventually be incorporated into the next revision of China's catalogue of technologies prohibited or restricted from export. ANALYSIS Where Washington's controls aim to slow a rival, Beijing's aim to lock in domestic advantages it now believes it possesses — a posture that would have been implausible even a year ago.
ANALYSIS Three developments explain Beijing's newfound confidence. First, Chinese open-weight models have reached competitive parity on key benchmarks: Kimi K3's performance against Anthropic's latest offering is the most visible proof point. Second, China's open-weight AI ecosystem has expanded rapidly through companies including Moonshot AI, MiniMax, and Alibaba. Third, a Chinese AI chip combining software-defined computing and 3D near-memory architecture was unveiled in Shanghai on July 13, achieving 520 TFLOPS on a 14nm manufacturing process — demonstrating an alternative path to high-performance AI computing that does not rely on leading-edge process technology3.
That chip's 6.4 TB per second of memory bandwidth, achieved through vertical stacking of compute and memory, directly addresses the memory-wall bottleneck that constrains conventional architectures. ANALYSIS The architectural approach sidesteps the very chokepoint — access to sub-7nm fabrication — that US export controls were designed to exploit.
Meanwhile, the discussions on China's side reportedly reflect lessons learned from Meta's failed acquisition of Manus, the Chinese AI startup. ANALYSIS Beijing appears to be treating that episode as a warning that Western firms could acquire Chinese AI capabilities outright, reversing the usual narrative of technology flowing from West to East.
Hugging Face CEO Clem Delangue offered a counterpoint to the IP-theft framing, arguing that distillation "is a practice that everyone is doing, including companies in the U.S.," and calling it "a very small factor in the ability to create good models". ANALYSIS His comments underscore the difficulty Washington will face in drawing a legally defensible line between legitimate knowledge transfer and sanctionable theft — a distinction that will matter enormously if Treasury moves from rhetoric to enforcement.
ANALYSIS The immediate question is whether either side's proposals harden into binding rules. On the US side, Bessent's comments signal an investigation, not yet an action; the Trump administration is also reportedly considering a wholesale ban on Chinese open-source models, though that claim has been disputed. On China's side, the Ministry of Commerce is still gathering industry feedback, and the proposed measures are likely to be incorporated into catalogue revisions rather than enacted immediately. In April, the White House said it would work closely with AI firms to combat IP theft. ANALYSIS The policy machinery on both sides is now in motion, and the window for any negotiated framework governing cross-border AI technology flows is narrowing with each announcement.