Margin of Safety #34: Looking at China’s AI Playbook

China’s AI strategy is a high-speed experiment in central planning, blending top-down regulation with massive infrastructure and human capital

Recently, Author Dan Wang’s new book Breakneck: China’s Quest to Engineer the Future has been widely discussed. It’s a fascinating exploration of how China is using industrial policy, infrastructure, and state direction to accelerate their ambitions. In that spirit, we wanted to jot down a few highlights from recent developments that caught our eyes:

Cover of “Breakneck: China’s Quest to Engineer the Future” by Dan Wang

• AI Regulation
  • State Council issued guidelines to integrate AI across industries, society, and government. By 2027, penetration of AI & agents in six key domains >70%. By 2030, 90%+ penetration (Link to source)
  • Model alignment. China’s Interim Measures for the Management of Generative AI Services (2023) require outputs to reflect “core socialist values” (Link to source)
  • It’s not clear how much you can regulate your way to AI success. The State Council guidelines, in many respects, mirror the top-down mandates American companies have tried to use to spur AI adoption. In the US, those mandates have resulted in failed efforts and bottoms-up creativity has been where we found success. Do we expect the Chinese model to have inverted results and, if so, why? And as for model alignment, it’s not clear to us how one accomplishes that. Fine-tuning only goes so far, and there simply isn’t enough curated training data in existence to fully constrain models to ideology.
• Self-reliance
  • Beijing is targeting self-reliance by 2027, heavily subsidizing purchases of domestic chips like Huawei’s Ascend to replace NVIDIA dependencies (Link to source)
  • Over $20B in subsidies last year alone, plus a new ¥1 trillion ($137B) national fund to back automation startups (Link to source)
  • Robotics and factory automation may be the U.S.’s biggest non-LLM weakness. This is where the lag is most visible and potentially most painful.
• Infrastructure
  • Electricity generation. (Link to source)

• • This is emblematic of a broader capability: the ability to centrally push massive infrastructure projects that would be mired in zoning and permitting elsewhere. It’s not free (bad forecasts or misallocations have costs) but it’s an advantage when national speed matters
• Security & Export Controls
  • On October 9th, China added five heavy rare earths to its export-control list and requires licensing for processing equipment that will go into effect on December 1st. This reflects use of materials (rare earths, gallium, germanium) as strategic chokepoints to counter Western tech embargoes (Link to source)
• Human Capital
  • (alleged) Deepseek’s top AI engineers had their passports taken away and issued travel ban (Link to source)
  • Nvidia CEO Jensen Huang Sounds says “50% Of the world’s AI Researchers Are Chinese” (Link to source)
  • Even in US companies, over half are Chinese. See the recent recruits for Meta’s Super Intelligence Lab (Link to source).
  • China benefits from a huge population base with strong math programs. The US has a smaller population base here, but has classically benefited from tons of immigration (especially of highly qualified PhDs) and a best in class university system. In an area like AI, where there’s a lot of speculative research, it’s hard to compete without a lot of very smart researchers. Other areas trying to compete need to think about how to hit that critical mass.
• Data & IP
  • Chinese AI firms benefit from a permissive data/IP environment, where scraping and using copyrighted content for training is largely tolerated. Temporary data copies may not constitute “reproduction,” effectively shielding model training (Link to source).

Reach out to us if you are building in the space of AI & security. We have some thoughts!
Kathryn Shih – kshih@forgepointcap.com
Jimmy Park – jpark@forgepointcap.com

This blog is also published on Margin of Safety, Jimmy and Kathryn’s Substack, as they research the practical sides of security + AI so you don’t have to.