China’s AI progress despite US chip restrictions
Recent Chinese AI models are closing the gap with US leaders, but experts argue this does not mean export controls have failed.
Recent releases from Chinese AI developers, including Kimi K3 and DeepSeek V4 Pro, have sparked debate about the effectiveness of US export controls. These models are approaching the performance of leading American systems on public benchmarks, often at lower costs and with open-weight architectures that allow for greater user modification. Consequently, some observers have concluded that the US strategy to restrict access to advanced semiconductors has failed.
However, according to analysis reported by GN technics/ai (en-US), this conclusion is premature. While the progress is real and significant, it does not prove that the restrictions have had no impact. To understand the true effect of the controls, one must compare current Chinese models against a hypothetical scenario where no restrictions existed. The gap between these two scenarios remains a key indicator of the controls' influence on China's AI trajectory.
Benchmarks mask practical limitations
Public benchmarks provide a snapshot of capability but often fail to capture how models perform in complex, real-world tasks. A joint analysis by US and UK standards bodies found that while Kimi K3 performed well, it still trailed leading US models in difficult, multi-step evaluations. For instance, in a complex cyberattack simulation, US models reached an average of 28.5 steps, whereas Kimi K3 reached 17. The developer itself acknowledges that its model still trails the best in overall performance.
The discrepancy between benchmark scores and practical utility highlights a critical trade-off. High scores in controlled environments do not necessarily translate to consistent reliability in open-ended problems. This suggests that while Chinese models are competitive, they may lack the robustness and depth of their US counterparts when faced with unstructured challenges.
Compute scarcity limits scale
A major factor in this gap is access to computational power. Research on scaling laws indicates that AI models improve with more training compute, which requires advanced chips. Chinese firms have repeatedly stated that US bans are hindering their progress. One estimate suggests Chinese companies owned only about 5 percent of the world’s AI compute through official channels at the end of 2025, less than any single US hyperscaler.
This scarcity has tangible consequences for users. Chinese developers may release open weights because they lack the infrastructure to deploy models at scale. Moonshot AI, for example, suspended new subscriptions for Kimi K3 after demand overwhelmed its capacity. This indicates that while the model architecture is strong, the ability to serve hundreds of millions of users consistently remains constrained by hardware limitations.
Enforcement gaps persist
Despite these constraints, the US control regime is far from perfect. Enforcement challenges allow significant amounts of hardware to bypass restrictions. In March 2026, US prosecutors charged a co-founder of Super Micro with conspiring to divert billions of dollars worth of NVIDIA servers to China through a front company. Estimates suggest that as much as a third of China’s total compute may have been smuggled in, representing a non-trivial fraction of global AI processing power.
These loopholes complicate the narrative of total success or failure for export controls. While the restrictions clearly limit the scale and speed of Chinese AI development, they have not halted it. The outcome is a complex landscape where Chinese models are competitive but constrained, and US controls are effective but imperfect. The ongoing arms race in AI continues to be shaped by both technological innovation and the geopolitical struggle over semiconductor access.
