The export ban was supposed to slow them down. Instead, China's AI labs just posted their best month ever relative to Silicon Valley.
The Summary
- Chinese AI models closed the performance gap with US counterparts to just 6% in June, down from 9% in May — the tightest margin on record
- Bloomberg Intelligence flags this as evidence that Zhipu's breakthrough wasn't a fluke, and that Moonshot's recent gains confirm a systemic acceleration
- The compression questions whether chip restrictions and export controls can actually preserve US AI dominance long-term
The Signal
The gap narrowed by three percentage points in one month. That's not incremental progress, that's a different gear. Bloomberg Intelligence points to Moonshot and Zhipu as proof points that Chinese labs have figured out how to train competitive models despite being cut off from Nvidia's best chips.
The export restrictions were supposed to create a moat. No access to H100s meant no frontier models. That thesis is breaking down in real time. Chinese labs are optimizing around the constraint, using older chips more efficiently, building custom silicon, and apparently getting results that close the gap faster than US policymakers expected.
"The sustainability of US technological supremacy in the sector" is now an open question, according to Bloomberg Intelligence.
This isn't about raw benchmark scores. It's about acceleration. Here's what matters:
- The gap was 9% in May. It's 6% in June. At this rate of compression, parity arrives before year-end.
- Zhipu's success looked like an outlier in Q1. Moonshot proving it wasn't a one-off changes the narrative from exception to trend.
- Chinese labs are doing more with less silicon, which means they're solving the hard problem: algorithmic efficiency under constraint.
The US strategy has been containment through hardware denial. But if Chinese models can close a 9% gap to 6% in thirty days without access to cutting-edge chips, the containment strategy has a shelf life. Either the models plateau soon or the export controls weren't the firewall Washington thought they were.
What's driving the compression likely includes better training techniques, more efficient architectures, and massive investment in homegrown chip production. China's been pouring capital into semiconductor independence since the first round of restrictions hit. Those bets are starting to pay off in model performance, not just capacity.
The Implication
If this trajectory holds, the AI race stops being about who has the best chips and starts being about who can do more with constrained resources. That's a different game, and China's been training for it since 2022.
For anyone building in the agent economy, this matters because model diversity accelerates. More competitive options from Chinese labs means more choice, more price pressure, and potentially more innovation in efficiency over raw scale. The monoculture risk of everyone depending on the same three US model providers starts to crack.
Watch what happens in Q3. If the gap keeps compressing at this pace, the geopolitical assumptions underlying the entire AI export regime need revision.