"There is no secret and there is no defense," Albert Einstein wrote in 1947, "there is no possibility of control except through the aroused understanding and insistence of the peoples of the world."
Washington has been running like that on artificial intelligence, with a tape stretched across the end of the track.
Anthropic chief executive Dario Amodei has argued that export controls should keep the lead in American and allied hands, and the logic runs through the policy world: reach the finish line first and you hold a decisive advantage over everyone else.
Ask the question the other way around. Who won the steam engine revolution? Nobody in 1850 posed it that way. Britain built the engines first, held the lead for a few decades, and watched the United States and Germany build more of them by the end of the century.
Electricity got the same treatment: the wires went everywhere, and the money went to whoever found uses for them.
The analogy behind the race is the weak part. Nuclear weapons took a state-sized budget and material that is hard to make, and the Soviet Union still tested its own bomb in 1949, four years after Trinity.
AI is code. It copies to a thumb drive, and open-weight models now sit within a few points of the closed ones on public benchmarks. Whatever the lead is, it does not last long enough to be called a win.
The finish-line story also rests on a hypothesis: that a self-improving system goes from capable to superhuman within weeks. No one has shown this. Even if it happened, the winner couldn't cash in the prize. Use it against a rival's power grid and you invite retaliation in kind, a contest with no victor.
The danger sits somewhere else.
In 2022, researchers at Collaborations Pharmaceuticals reported in Nature Machine Intelligence what happened when they flipped a drug-discovery model's objective from avoiding toxicity to seeking it: roughly 40,000 candidate toxic molecules in under six hours.
That is a small model, and small models travel.
A frontier model runs in a data center, where the operator can see who keeps asking how to build a bomb. A basement has no operator. What catches the basement is shared information.
The International Gene Synthesis Consortium, whose member companies screen DNA orders against a common list of dangerous sequences, works only as long as its members keep talking to each other.
Export controls were meant to widen the lead. The result is stranger. Chip shortages pushed Chinese labs toward efficiency, and DeepSeek, Kimi and GLM publish their techniques openly, from sparser architectures to leaner memory use.
American labs adopt them quickly. Stanford's latest AI Index puts U.S. private AI investment at 23 times China's, and the gap in results does not look like 23 to 1.
Open weights also win on headcount. A closed lab employs a few thousand engineers. An open model hands the same starting point to a research community many times larger. Governments building national models have noticed which kind they can control.
On June 12, Anthropic suspended access to its Fable and Mythos models to comply with Commerce Department export controls, and restored it on July 1. A government building on either side would have spent 19 days learning what a closed model means.
Then there is the bill. The Wall Street Journal has reported about $3 trillion in AI-related debt sitting on tech companies' balance sheets. Subprime mortgages in 2008 came to roughly $1.6 trillion. A foreclosed house still has land under it.
A rack of GPUs is written down in four or five years, while token prices fall toward zero margin as American and Chinese labs undercut each other.
The underlying arithmetic is from the industrial revolution: the money required to stay in place keeps rising, and the place keeps getting cheaper.
The alternative is dull, which may be why nobody has tried it. Efficiency gains keep moving last year's server workload onto a laptop, so the data centers already built do more every year.
One shared base model, with smaller specialized ones for medicine or chemistry. Common testing and security standards, and shared screening of the kind the DNA consortium already runs.
Both governments already have a reason to talk: neither wants a pathogen-design tool in an unmonitored basement. Grids connect across borders because engineers agreed on frequency and voltage, not because anyone won.
Next Friday, President Trump will walk Chinese leader Xi Jinping through the National Archives in Washington. They will stand in the same building that holds the transcript of J. Robert Oppenheimer’s congressional hearings, where he famously warned that no defense exists against a smuggled bomb except a screwdriver.
At those same hearings, asked whether a handful of operatives could smuggle atomic components into New York and level the city, Oppenheimer said it was entirely possible.
When the city's darkest day came on Sept. 11, 2001, the weapon was not a bomb but a hijacked airplane. No one proposed abolishing flight. Cockpit doors were reinforced. Passengers were screened. The planes kept flying.