China Can’t Buy Nvidia’s Best Chips. It Doesn’t Need To.

Geopolitical and Security Incidents

• Four days after filming, NVIDIA CEO Jensen Huang admitted to an affair with a woman U.S. intelligence officials suspect is a Chinese spy recruited to gather information and push for the sale of NVIDIA's most advanced chips to China.
• In December 2024, DeepSeek, a Chinese company, released an AI model rivaling top U.S. models for approximately $6 million, compared to the tens of millions spent by American counterparts.
• Following the DeepSeek release, NVIDIA lost nearly $600 billion in value in a single day, marking the largest one-day loss in U.S. stock market history.

Current State of AI Competition

• While China's best models trail the top U.S. models by roughly six months, the U.S. possesses approximately 16 times more AI computing power.
• China's six leading AI companies collectively earn about 10% of the revenue generated by OpenAI and Anthropic combined.
• Alibaba's chatbot has 380 million monthly users, whereas ChatGPT has more than double that weekly.
• Experts estimate China has a less than 20% chance of overtaking top American AI models within the next three to five years.
• Despite the lag, China's AI industry is inventive, with over half of the world's top AI researchers holding Chinese undergraduate degrees and staying locally.

Constraints on Chinese AI Development

• Export Controls: Since 2019, U.S. and Dutch regulations have blocked China from accessing ASML's most advanced chip-making machines.
• Memory Restrictions: U.S. rules banning the sale of specialized high-speed memory to China, effective December 2024, have forced Huawei to rely on Chinese-made memory for only 16% of its AI chips this year.
• Computing Power Gap: Due to these restrictions, the U.S. generates roughly 50 times more AI computing power than China.

Methods of Chinese Access to U.S. Technology

• Legal Purchases: In January, the U.S. approved sales of NVIDIA's H200 chips to China, though Beijing advised companies to delay using them to protect domestic industries.
• Gaming Card Circumvention: Chinese firms purchased over one million NVIDIA graphics cards designed for gaming, which fall below strict export limits but are used for AI training.
• Smuggling: Smugglers have moved at least $160 million worth of chips, including instances where top-tier AI chips had their labels peeled off and replaced with fake brands.
• Cloud Renting: A significant portion of Chinese computing power comes from renting U.S. chips abroad; for example, ByteDance runs a cluster of NVIDIA chips in Norway, while others rent in Malaysia and Singapore.
• Distillation: Chinese labs use "distillation" to train their models by imitating the outputs and reasoning of stronger American models, a technique estimated to close roughly half the performance gap.
• Data Harvesting: Chinese entities have used roughly 24,000 fake accounts to scrape data from American models; Anthropic successfully reduced harvestable data from 90% to 1% by improving defenses.

Strategic Divergence: "Smartest AI" vs. "AI Integration"

• While the U.S. race focuses on building the most capable frontier models, China is prioritizing the integration of AI across all factories, hospitals, and government offices.
• Energy Advantage: The U.S. faces a projected 44-gigawatt energy shortfall for data centers by 2028, whereas China is expected to add nearly six times more generating capacity over the next five years.
• Manufacturing Dominance: China installed 59% of the world's industrial robots last year and 97% of all humanoid robots sold globally in the first half of this year.

Security Vulnerabilities

• Top U.S. AI companies are rated only 2 or 3 on a 5-point security scale, far below the "nuclear-level" security (Level 5) required to thwart major state actors.
• In January, a former Google engineer was convicted of stealing AI training data and supercomputer blueprints for Chinese companies.
• Stolen AI models could bypass safety safeguards, potentially providing access to dangerous capabilities like automated cyber-attacks or bioweapon design.
• Unlike physical models, the "breakthroughs" in training techniques are harder to obsolete and could provide China with a lasting technological boost even if they lack current computing power.

Policy Debates and Forward-Looking Statements

• Incentive Conflicts: NVIDIA CEO Jensen Huang and former AI czar David Sacks have advocated for maintaining chip sales to China to prevent total decoupling, despite holding financial stakes in AI companies.
• Proposed Countermeasures: A House-passed bill aims to restrict cloud renting of U.S. chips; analysts suggest this, combined with tighter export controls on ASML maintenance, could double the U.S. lead by the end of the year.
• The "Wake Surfer" Analogy: Policy experts argue China relies on "distilling" U.S. innovation; if the U.S. slows down or closes access channels, China's progress could stall as it lacks the independent "engine" to maintain the gap.
• Security Recommendations: Experts argue the U.S. should prioritize voluntary security standards to reach Level 5 protection to prevent the theft of models and training techniques.
• Strategic Conclusion: The U.S. has a potential "breathing room" of six months to slow down and implement safety measures, provided it closes loopholes regarding chip renting, smuggling, and model security.