US chip export controls slowing China's AI despite model advances

Recent Chinese AI models from companies like Moonshot AI and DeepSeek have approached or matched leading US systems on public benchmarks, prompting questions about whether US semiconductor export controls have failed. But new analysis from the Institute for AI Policy and Strategy argues that benchmark scores tell only part of the story.

As of September 2026, Chinese model Kimi K3 scored 158 on the Epoch Capabilities Index compared to 169 for GPT-6 Astra — a gap representing roughly four to ten months of progress at recent rates. While these results demonstrate real capability, they don't capture performance on extended, open-ended tasks where compute advantages matter most.

Compute access remains constrained

Chinese AI companies owned approximately 5 percent of global AI compute through official channels at the end of 2025, according to estimates cited in the analysis — less than any single US hyperscaler. This figure excludes smuggled chips and remote cloud access, but the remaining gap still limits experimental scale and model deployment capacity.

The compute constraint shows up in practical ways. Moonshot AI suspended new subscriptions for Kimi K3 after demand overwhelmed its serving capacity. Chinese companies often release model weights publicly, likely because they lack the infrastructure to deploy systems at the scale of US competitors.

A July 2026 evaluation by US and UK standards bodies found Kimi K3 reached step 17 on average in a 32-step cyberattack scenario, while leading US models reached step 28.5. Moonshot AI itself acknowledged that Kimi K3's overall performance "still trails" the best models.

Smuggling and cloud loopholes persist

Export controls face enforcement challenges. US prosecutors charged a Super Micro co-founder in March 2026 with conspiring to divert up to $2.5 billion worth of NVIDIA servers to China through Southeast Asian intermediaries. One estimate suggests smuggling may have provided as much as a third of China's total compute through 2025 — equivalent to roughly 3 percent of global AI compute.

Rented cloud compute represents another access channel. Estimates suggest China accesses at least 3 percent of global compute remotely, though this figure carries substantial uncertainty. Such access is generally legal when chip owners aren't headquartered in China. The Remote Access Security Act, which passed the House of Representatives and awaits Senate action, would authorize the executive branch to apply export controls to remote access.

Why it matters

The debate over export control effectiveness often conflates two different questions: whether China can build capable AI models, and whether controls have slowed Chinese progress relative to an unrestricted scenario. Strong benchmark performance demonstrates the first but doesn't answer the second. The distinction matters for policymakers weighing whether to maintain, tighten, or relax semiconductor restrictions as China's domestic chip industry improves and US leverage potentially weakens over time.

Long-term leverage questions

China's domestic semiconductor industry continues to lag but is advancing. Export controls may accelerate Beijing's push for self-reliance by creating stronger incentives to adopt domestic chips and develop supporting software. However, the controls also restrict access to advanced chipmaking equipment, potentially offsetting faster indigenization.

The analysis concludes that US chip controls appear to have succeeded at their stated objective: slowing Chinese AI progress and limiting deployment scale rather than preventing capable models entirely. This preserves frontier advantages and buys time for defensive applications like hardening software against autonomous cyberattacks.

The findings were published by Erich Grunewald, Senior Researcher at the Institute for AI Policy and Strategy, and first reported by East Asia Forum.