As China continues to hit AI development milestones, the United States has pushed to slow their innovation — but research shows their efforts are not always successful.
In October 2022, the U.S. banned exports of advanced GPUs to China, betting that cutting off access to frontier chips would slow the PRC's AI ambitions. To test that bet, this new NSDPI paper uses a model that estimates what Chinese AI publication output would have looked like absent the embargo. The results are striking: Chinese AI publication outputs have actually outpaced projections every year since 2022, and the gap has widened annually. Rather than curbing China's trajectory, the embargo coincided with acceleration.
Why did China's AI ecosystem prove so resilient to the United States embargo? What does China's progress in algorithmic efficiency — like DeepSeek's Mixture-of-Experts architecture — reveal about how an adversary innovates around hardware restrictions? And how should U.S. competition strategy adapt in response?
Key Takeaways:
- Chinese LLMs have achieved near-parity with Western frontier models despite strategic efforts from the United States. We suggest that the embargo constraints likely energized an already fertile and productive research environment, inspiring creative new approaches to AI — such as DeepSeek’s Mixture-of-Experts architecture — despite hardware restrictions.
- The U.S. government needs to adapt their approach to AI governance and policy. As agentic usage increases and AI becomes more costly, algorithmic efficiency will play a decisive role. The United States must diversify its approach to competition with China — considerations include utilizing an open model, encouraging technological cooperating with allies, and taking threats of adversarial distillation seriously.