I keep coming back to something Jensen Huang said last week on a podcast. He was talking about DeepSeek's V4 model, which dropped Friday with Huawei chip support. "The day that DeepSeek comes out on Huawei first," he said, "that is a horrible outcome for the US." And I thought: for whom, exactly? For Nvidia's shareholders, clearly. For the US strategic position in AI, arguably. But for American workers whose jobs are being automated away by AI systems, does it matter at all whether the AI doing it runs on American or Chinese chips?
This week the Trump administration announced a crackdown on "AI distillation," accusing China of running industrial-scale operations to extract learned capabilities from US frontier models. The memo from Michael Kratsios described "tens of thousands of proxy accounts" and systematic jailbreaking. Congress passed bipartisan legislation targeting foreign actors who copy US AI capabilities. And the whole apparatus of Washington's AI policy is now pointed at a single question: how do we make sure America wins the AI race?
Nobody in Washington seems to want to ask the follow-up question. What does winning look like?
In the 1990s, the US government tried to control encryption as a munition. Phil Zimmermann, who released PGP, was investigated for arms export violations. The NSA pushed the Clipper Chip, a hardware backdoor proposal that went nowhere. The result of all that effort? Encryption proliferated globally regardless, US tech companies were put at a competitive disadvantage during the restriction period, and the mathematics involved, being mathematics, turned out not to be controllable by export law. The technology spread. The world encrypted. The attempted controls achieved nothing except slowing down the companies they were supposed to protect.
Now substitute "distillation of AI model capabilities" for "public key cryptography." The underlying logic is identical. DeepSeek's advances do not depend on stealing OpenAI's outputs. They depend on the same publicly available research literature that everyone in the field reads. The mathematics of transformers, mixture-of-experts architectures, and efficient inference are not state secrets. They are published papers. You cannot export-control a paper.
But here is the thing that is harder to say, and that the crackdown framing actively prevents anyone from saying: even if the US somehow wins this race by a clear margin, who benefits?
When the US AI industry builds models that can replace knowledge workers, those workers lose income. When they build agentic systems that coordinate and execute business processes, the humans who did that coordination lose their roles. When AI inference costs approach zero, which DeepSeek is actively engineering with its $3.48-per-million-token pricing, the economic barrier to deploying AI at scale disappears. Every individual company that deploys AI instead of workers makes a rational choice. The collective result of all those rational choices is a demand collapse in slow motion: fewer employed people means fewer customers means a loop that feeds on itself.
This is not an American problem or a Chinese problem. It is a structural problem that applies equally to both countries. Chinese workers are being displaced by AI too. Chinese companies are racing to deploy the same labour-replacing technology. The AI race Washington is so determined to win is a race that both economies are running toward the same destination, which is an economy where capital owns intelligent machines that produce everything, and the question of who owns the wages those machines would have paid is answered by silence.
WarGames got it right in 1983. The computer, WOPR, runs every scenario for thermonuclear war and reaches the same conclusion every time: "A strange game. The only winning move is not to play." This is an arms race, and arms races are hard to stop even when the destination is obvious. The framing of it as a competition to be won, by restricting what China learns, prevents the more important question from being asked: what are both countries going to do about the workers who are going to be displaced regardless of which side's model is marginally better?
The White House memo is worried about China "exploiting American expertise and innovation." That is a reasonable concern in a narrow IP sense. But the innovation that is being exported, whether through distillation attacks or through the ordinary spread of published research, is innovation that will displace workers on both sides. Winning the race to build the most powerful displacement engine is not a prize worth the effort. It is just arriving at the cliff faster than the other car.
I do not know what the right policy response looks like. Neither does anyone in Washington, or Beijing. But I am quite sure the wrong response is to optimise the race without asking where it ends.