The AI Race Is a Race to Make Stopping Too Expensive

It doesn’t take a plan. It takes a few hundred billion dollars, a friendly president and a public that thinks this is all science fiction.

Sometimes I wonder what we are racing. Sometimes. Who am I kidding. All the time. And the longer I look at who is actually running, the less it looks like a race towards anything.

The official answer is China. Donald Trump gave it in July last year, at an event called, with no irony anyone noticed, Winning the AI Race: “It’ll be a policy of the United States to do whatever it takes to lead the world in artificial intelligence.” Then, in case the tone wasn’t clear: “We dominate the future. We conquer new frontiers and we control our own fate.” Whatever it takes is what people say once they’ve decided the goal excuses the methods.

China is a real answer, and a very convenient one for anyone who wants American AI companies left alone. Somebody is paying to keep it in circulation. In May, Wired reported that a dark-money group linked to the super PAC funded by Andreessen Horowitz and OpenAI’s president was pitching influencers up to US$5,000 a TikTok to deliver lines like “China is trying really hard to beat the US in AI. If they do, it could mean that China gets personal data from me and my kids.” A staffer at the agency running it said the aim was to “subtly shift public debate”. The rivalry is real enough as geopolitics. It’s also a product, with a marketing budget, and what it sells is the idea that regulation is surrender.

So look at who is running, and at what the running does. The United States is locked in now, and the people running don’t need to want the same future for that to hold. They only need to keep making decisions that raise the price of changing course, and every one of them is doing that. Changing course means anything they wouldn’t choose: a release delayed, a founder removed, a limit that sticks.

Start with the man who says whatever it takes. He has scrapped his predecessor’s AI order and set about overriding the states’ AI laws because “you can’t have a state with standards that are so high that it’s going to hold you up”, and promised that the technology “will make America richer, stronger, greater, freer, and more powerful than ever before.” You can read that as a leader who believes this will make him all-powerful. I do. Whatever he believes, the record shows a leader clearing every obstacle in front of the companies, and the states were the only place in America that had written binding rules for these models at all.

The economy then supplies the excuse, because AI has become its life preserver. Jason Furman, who chaired Obama’s Council of Economic Advisers, worked out that in the first half of 2025 spending on information-processing equipment and software was four per cent of American GDP and 92 per cent of its growth. He was counting a category of spending rather than AI itself, and a fair slice of it went straight back out as imported chips, so the AI share is smaller. It doesn’t matter. The US grew at an annualised 1.5 per cent in the second quarter of this year, a pace the Center for Economic and Policy Research said would be fine if it “were not so dependent on AI-related spending”, and four companies plan, on their own guidance, to spend upwards of US$700 billion on capital expenditure this year. Once that much growth is booked to a thing, any limit on it can be sold as a recession, and nobody votes for one of those on purpose.

Then the billionaires and the would-be billionaires, who can smell power before being supplanted by whatever they create. Each has raised the price of stopping in his own way. Elon Musk asked Tesla shareholders for a trillion-dollar package and explained that the money was beside the point; what he wanted was enough voting control that he wouldn’t be “building a robot army here” and “then being ousted”. That is protection against removal, said out loud. Mark Zuckerberg published an essay last month promising to deliver superintelligence “to everyone” and warning that “there is no such thing as a singular benevolent superintelligence”; the only authority over Meta’s releases it mentions is Meta’s own board, which is how his products have always worked: everyone gets access, nobody gets a say. Sam Altman says OpenAI is “looking at commitments of about $1.4 trillion over the next 8 years”, a sum that turns any future stop into everyone’s problem.

The political protection is already under construction. Leading the Future, the super PAC, and its affiliated nonprofit plan to spend about US$140 million on the midterms, according to Axios. In June it spent roughly $8 million to help defeat a New York state assemblyman, Alex Bores, in a Democratic congressional primary. His offence was sponsoring a state law, already signed, that requires frontier AI companies to publish how they handle catastrophic risk. The man who beat him had backed the same law and had the retiring congressman, the governor and Michael Bloomberg behind him, and the rival faction, with Anthropic’s money prominent in it, spent about $19 million trying to save Bores. So the money didn’t decide it. The money is still the point. This year, both sides of the argument about whether to regulate these companies have been paid for from inside the industry.

And then the rest of us, whose job it is to say no. Too many of the people who could be doing that job still treat the whole enterprise as some nerd’s wet dream, and the moment it’s in that folder the money, the land and the law go in with it. You can refuse to take the dream seriously and still count the power stations. The governments have done no better. When OpenAI released GPT-6 Astra last week, the first model to cross the company’s own “critical” threshold for cyber capability, Washington had already seen it, under an executive order signed in June that lets a company show the government a new model for up to 30 days before release, voluntarily, and says in terms that nobody needs a licence or a permit from it. Access, not permission. Greg Brockman said the government asked for no changes. My own government’s new AI office, announced in July, so far concerns data-centre power and water; nothing in it could delay a model.

The data centres, the override of state law, the voting control, the 30-day look, the $140 million: none of it is waiting on the technology to deliver what its owners promise. You don’t have to believe a word of the singularity talk to notice that a very singular group of people is acquiring, quickly and on the public record, the kind of position you don’t hand back.

So, what are we racing? The moment it becomes too expensive to stop. Nobody has to want that moment for it to arrive. They only have to keep making the decisions that bring it closer, and they’re making them faster than the rest of us are paying attention. The contest that counts, over who can still say no, they’ve been winning all year while the rest of us, me included, looked at the machine instead of at them.