China's AI didn't overtake through homegrown innovation—it runs on rented American compute and copied American models
Chinese models lag American ones by about six months, but this gap is largely sustained by overseas rentals of Nvidia chips and knowledge distillation from American models—once America closes both routes, the real gap would be much wider than it appears.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
DeepSeek's shock was misread as catching up
After DeepSeek's release, Marc Andreessen called it AI's "satellite moment," and Nvidia's market value dropped nearly 600 billion USD in a single day—the largest single-day decline in US stock history. But nearly two years later, China's best models still lag America's best models by roughly six months, while US AI computing capacity is about 16 times that of China. The real question isn't ‘has China caught up’ but how China can be only six months behind given such a vast disparity in computing power.
— Luisa RodriguezTwo physical bottlenecks choke China's chip development
The most advanced AI chips depend on two things: ASML, the only company that makes cutting-edge lithography equipment, and specialized high-speed memory. Since 2019, the US has worked with the Netherlands to ban China from importing the most advanced lithography machines; in December 2024 it also banned the required memory, leaving Huawei with only about 16% of its AI chips this year using Chinese-made memory, the rest relying on stockpiles, smuggling, or inferior substitutes. Combined, US computing capacity reaches about 50 times China's—export controls are working.
— Luisa RodriguezThe biggest loophole is renting, not selling
America banned Nvidia from selling its most advanced chips to Chinese companies, but didn't ban Chinese companies from renting the same chips at overseas data centers. Tencent rents top-tier chips from Oracle; ByteDance runs Nvidia clusters in Norway; but the largest concentration of rentals is in Malaysia and Singapore. In 2025, four Chinese engineers flew to Malaysia, rented about 300 servers packed with Nvidia chips, and when network speeds proved too slow, they loaded training data into 15 hard drives, packed them in luggage, and flew back to train models at home.
— Luisa RodriguezThose shouting 'every month matters' push faster chip sales
Huang Renxun (黄仁勋) warned that ‘China is only a few nanoseconds behind America,’ while simultaneously advocating for selling more Nvidia chips to China. Trump's former AI czar David Sacks said ‘every month matters,’ yet while in office pushed to loosen restrictions on Nvidia overseas sales. The New York Times found he holds stakes in 449 AI-related companies; the exemption granted to him came from White House counsel itself—signed just one month after Trump fired the head of the independent ethics office. Those shouting loudest about the ‘race’ are often those most hoping chips keep flowing to China.
— Luisa RodriguezDistillation copies most capabilities with one-tenth the effort
Chinese labs massively query top American models, recording answers and reasoning, then train their own models to mimic them—this is distillation. Anthropic reported that in just three months, Alibaba exchanged over 151 million messages with Claude. The Institute for Progress estimates distillation erased roughly half the gap between US and Chinese models, equivalent to freely giving Chinese AI companies four times the computing power. To block distillation requires identifying vast numbers of fake accounts; in one operation, DeepSeek, Moonshot AI (月之暗面), and MiniMax collectively deployed about 24,000 accounts.
— Luisa RodriguezChina's AI leaders admit they're far behind
At a Beijing conference in January, Lin Junyang (林俊旸), then technical lead for Alibaba's Qwen, put the odds of Chinese AI companies overtaking top American ones within three to five years at under 20%, calling this ‘quite optimistic’—the reason being computing power gaps. Tang Jie (唐杰), chief AI scientist at Zhipu (智谱), was more pessimistic, arguing that since many of America's best models aren't public, the gap may actually be widening. People sitting at the center of the table don't think their odds are good.
— Luisa RodriguezChina plays a different race—deployment, not frontier models
China's goal isn't to build the smartest chatbot but to embed AI in every factory, hospital, and government office, and in this race China has real advantages: US data centers may face a 44-gigawatt power shortfall by 2028, equivalent to needing the grid to absorb California's worth of capacity in just over a year, while China plans to add nearly six times as much generating capacity over the next five years. Last year China installed about 354,000 industrial robots, 59% of the global total; in the first half of this year, 97% of all humanoid robots globally came from Chinese companies.
— Luisa RodriguezTop AI models defend only at level 2 or 3
The industry measures model theft protection on a scale of 1 to 5, with 5 meaning it can withstand the world's top intelligence agencies, but current best estimates show top American AI companies only reach 2 to 3, with plans to hit 5 by 2028 or 2029, and entirely voluntary—unlike nuclear power plants with mandatory federal safety standards. Former Google engineer Linwei Ding was convicted of stealing Google's AI supercomputing architecture and simultaneously working secretly for two Chinese tech companies. If a top model or its training methods are stolen, the window America wins through computing power and export controls could be erased overnight.
— Luisa RodriguezIn their own words · checked verbatim
NVIDIA lost almost $600 billion in value in one day — the biggest one-day loss in US stock market history.
Luisa Rodriguez0:00
Because China is keeping up. But it couldn’t do that without American help.
Luisa Rodriguez1:40
China is just nanoseconds behind the US.
Jensen Huang9:23
put China’s chance of overtaking top American AI companies at less than 20% in the next three to five years. And he called that “highly optimistic.”
Lin Junyang12:07
Matt Sheehan, who studies Chinese AI policy, compares China to a wakesurfer riding the wave behind a boat. It keeps up partly because the boat is doing a lot of the work.
Luisa Rodriguez15:10
If China stole a model like that… they’d basically have a cyberweapon with unprecedented capabilities.
Luisa Rodriguez23:10
They don’t seem to be fully in control of their own models, but they’re using them to build the next generation of AIs anyway.
Luisa Rodriguez26:00
Figures
| US to China AI computing capacity ratio (overall) | about 16x | 0:00 |
| Nvidia single-day market value drop | nearly 600 billion USD | 0:00 |
| US to China computing capacity after export controls | about 50x | 4:57 |
| ByteDance chatbot monthly active users | about 380 million | 3:03 |
| China's six largest AI companies' revenue as share of OpenAI plus Anthropic combined | about 10% | 3:03 |
| Chip smuggling network transaction volume | at least 160 million USD | 6:24 |
| Malaysia and Singapore server export decline to China (after licensing) | about 80% | 6:24 |
| Messages exchanged between Alibaba and Claude in three months | over 151 million | 10:51 |
| Global humanoid robots from Chinese manufacturers (first half of year) | 97% | 13:47 |
Glossary
- distillation
- Training a smaller model to mimic a larger model's abilities using recorded question-answer and reasoning pairs.
- wakesurfer
- Riding waves created by another's efforts; metaphor for catching up by leveraging others' work.
How to listen
Practitioners wondering whether US chip export controls on China actually work, investors and policy observers who care about AI safety and geopolitical tradeoffs.