The case for export controls has retreated all the way down to a single premise that still holds
First it was about preventing military use, then it retreated to stopping frontier models, then to limiting how many AI agents China can run; the only self-consistent premise left for this policy is a belief in a short sprint to godlike superintelligence, winner takes all.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
What the China discourse lacks is a target state
Nathan says the missing piece in America's discussion of China is a ‘target state’: everyone jumps straight to ‘we want to win’, but nobody answers what winning looks like, what happens if we lose, or whether there is an achievable condition for victory. His proposed goal is Pax Machina, or better, Pax Robotica — the two countries and the world sharing the prosperity and peace of the AI era, robots into factories, into fields, into homes, while avoiding the AI version of the sword of Damocles. He is explicit that this is not about winning, it is about arriving together. A competition with no target state leaves only escalation itself.
— Nathan LabenzChina is not rising, it is returning to historical normal
To the claim that ‘China can't innovate’, Nathan applies one word: cope, and says it is isomorphic to cope in the AI debate — conceding that AI has come this far while insisting it won't reason, won't discover new knowledge, won't change paradigms. His judgment is that China's return is the long-run historical norm, and that the period he and Dario grew up in — China down, America the sole superpower — was the anomaly. Any effort to slow China's progress can therefore work at most in the short term; China will be a first-tier player in nearly every major human domain. This is not a prediction, it is, in his view, something already locked in by historical structure.
— Nathan LabenzOn purchasing power, China is half again as large as the US
In dollar-denominated GDP, the US is about 30 trillion and China about 20 trillion, so the US is 50% larger; on purchasing power parity, China is about 45 trillion against the US's 30 trillion, so China is 50% larger. Nathan says a good meal for two in Shanghai can come in under $10, which is impossible in New York, San Francisco, or even Detroit, so the PPP measure is closer to reality. A more concrete picture: a roughly 45-minute pedicure costs about $12, and the young employee in her twenties at the shop pulls out an iPhone 17 — the iPhone 17 is not cheaper in China than in the US, and high-end Xiaomi and Huawei phones also approach a thousand dollars. His explanation is the one-child structure, with the wealth of four grandparents plus two parents concentrated on one person.
— Nathan LabenzThe chip ban hits a century of humiliation
China sees itself as a top-tier civilisation that spent most of history at the technological frontier; it missed once and was invaded, forced to accept opium, with the European powers and Japan as the perpetrators — and someone who was 20 when the Cultural Revolution ended in 1976 is still alive today. The core lesson is that you cannot fall behind at the technological frontier: ‘we are never going back’. Huawei is the tech champion under this framework: the first to genuinely reach the top of a global frontier industry, only to be blocked by the US pressuring allies over espionage risk, when Snowden has already proven that the US listens to everyone. The chip ban is the same script again: AI is barely off the ground, there isn't even an agent that can reliably book a flight, and the ban comes first. Nathan says this is not Xi Jinping's view alone; it has seeped into educated society. He also concedes that American companies' grievance over stolen IP is entirely legitimate.
— NathanChina doesn't talk about an AI race, it talks about public goods
Nathan did not hear a single person the whole trip say China is ahead, or that China will win; almost nobody even framed AI as a race. What he heard was prosperity, material abundance, how to care for an aging population. In a mall he asked a car company's staff about autonomous driving, and they volunteered that ‘but Tesla is better’. At WAIC, Xi's speech opened by noting that the field of AI began at Dartmouth College in the US — he could have left it out, and careful observers read it as an olive branch; another line was that countries should jointly oppose ‘over-extending the concept of national security in the AI domain’. China's hardest signal so far is open weights: take them, run them for free, fine-tune them however you like, and once they're out you can't take them back. Nathan says this is both a sincere wish and a strategic position, quite credible now, but entirely capable of changing later.
— NathanThe rationale for export controls keeps retreating
Nathan thinks export controls do have an effect, but mainly in the future — the large-scale expansion of chip capacity hasn't landed yet. What he really objects to is the drift in the policy's rationale: initially preventing military applications, then retreating to stopping frontier models, now retreating to limiting how many AI agents China can run; that sounds like an attempt to suppress the Chinese public, to keep Chinese companies and individuals from using the tools Americans use themselves. The only self-consistent premise for this policy is a belief that we are in a short race toward godlike superintelligence and that whoever gets there first takes all — which is basically Anthropic's belief, so Dario advocating these policies is not self-contradictory: he genuinely believes recursive self-improvement kicks off in '27 or '28. Nathan concedes he can follow the logic, but still argues for slowing the frontier together instead.
— NathanThe most dangerous part of decoupling is not understanding each other
He asked his Chinese interviewees: if decoupling continues and the two sides end up on different branches of the technology tree, no longer understanding what the other has, would that worry them? The answer was ‘not yet’; when would it? ‘When China's chip ecosystem is good enough that we can do our own thing entirely.’ From this he derives the core risk: a hardware fork drags along a hardware-algorithm co-design fork, and add to that China stopping open weights and stopping publishing in English, and it may not take long to reach a point where ‘we don't know the shape, architecture or capabilities of each other's AI’. At that point both sides would genuinely be right to panic, and genuinely right to race.
— Nathan LabenzReplace the arms race with a disease medal table
He starts with a rule: don't train LLMs to maximise money-making on the open internet, and don't train AI to accomplish military objective x by any means necessary — that is training the very thing we fear most to seize power. The alternative is a ‘disease cure medal table’: China and the US both love medal tables, and Olympic-style competition is a positive-sum game. The table could be most cancer types cured, most diseases cured, highest energy-density battery, lowest annual traffic deaths. He says this needs very little capital to catalyse and could be offered unilaterally as a global public good. The landing point: he doesn't care whether the cancer cure comes from China or the US, he just wants cancer cured.
— Nathan LabenzIn their own words · checked verbatim
Maybe we could even call it a Pax Robotica.
Nathan Labenz8:56
China has arrived. China is a world power. And any notion that we might have that we can contain China's rise is frankly ridiculous at this point. That ship sailed a long time ago.
Nathan Labenz18:56
In China, I think they are living their eternal 1991 right now.
Nathan33:51
We will distill off of your models if we need to, and we won't feel too bad about it because we're never going back there again.
Nathan1:25:10
The only way that makes sense is if you believe that we are in a short term race to godlike superintelligence and that the only thing that matters is who gets there first.
Nathan2:11:27
This looks an awful lot like something that the CCP would do. I don't want us to take actions that are exactly the kind of actions that we think are the ones that make the CCP bad. This is not the American way.
Nathan2:21:46
And the response I got was not yet. And so I asked, well, when would you be worried about that? And the answer was, at the point where the Chinese chip ecosystem gets good enough that we're doing entirely our own thing, that's when that could happen.
Nathan Labenz2:47:07
I don't care in the end if the cure for cancer comes from China or comes from The United States. I just want cancer cured.
Nathan Labenz3:07:25
Figures
| A pedicure in Shanghai (about 45 minutes) | about $12 | 22:51 |
| Beijing–Shanghai high-speed rail, second class | about $60, at roughly 200 mph | 28:44 |
| Number of frontier or near-frontier LLM companies | 11 in the US, 12 in China | 44:11 |
| Year China's first five-year plan mentioning AI was adopted | 2016 | 44:11 |
| Xi's age | 73 | 59:30 |
| English-speaking population of China (estimating 10% speak English) | about half the English-speaking population of the US | 2:52:18 |
Glossary
- Pax Robotica
- Nathan's proposed target state: China, the US and the world sharing the prosperity of the AI era, robots into factories and homes, rather than suppressing one another.
- eternal 1991
- Dario's phrase for American complacency after winning the Cold War; Nathan flips it to say the Chinese government is living in its own 1991.
- cope
- A psychological defence that concedes the fact but refuses its implications; Nathan uses it for the claim that ‘China can't innovate’.
- confidential computing
- Data is processed inside a protected environment and only the conclusion is returned; even the operator cannot see the raw data.
- formal methods
- Proving mathematically that software behaviour matches its specification; Nathan suggests using it as a condition for lifting the Huawei ban.
- system three thinking
- Nathan's coinage: transcending your situation, dreaming, having insight, breaking out of an existing paradigm — the residual advantage leaders hold over AI.
How to listen
Founders, investors and teams going overseas who need to read where US-China AI policy is heading, especially anyone assessing export controls, decoupling and open-weight model risk.
1:46:03–2:06:00, the section on America's own advantages, can be fast-forwarded.