Not a startup story. An empire story.
Karen Hao covered artificial intelligence for MIT Technology Review and was the first journalist to profile OpenAI at length from the inside. Empire of AI takes the company at the centre of the boom and refuses the usual frame. Instead of visionary founders and inevitable progress, she asks the questions you would ask of any empire: what is being claimed, who pays for it, and what story makes the taking sound like a gift.
Every empire needs a civilising mission. This one calls it benefiting humanity. The Mdrn Urban
Data workers in Kenya.
The question is not whether the technology works. It is what had to be taken to make it work at this speed.
Four moves, and every empire makes them.
Hao’s organising idea is that the leading AI companies behave in a pattern historians already have a name for. They claim resources that are not theirs — text, images and code scraped from the open internet, then land, power and water. They rely on labour that is cheap because it is far away and out of frame. They justify all of it with a civilising mission, in this case building artificial general intelligence for the benefit of all humanity. And they insist on a race against a darker rival, which converts caution into betrayal and makes the pace itself unquestionable. Whether or not you accept the analogy, it does real work: it gives you four things to look for in any announcement.
The bet that bigger is the whole strategy.
The technical heart of the story is a wager: that capability rises predictably with more compute, more data and more parameters, so the winning move is to build bigger than anyone else can afford. That bet paid off spectacularly, and it also decided the shape of the industry. Once progress is bought rather than invented, the moat is capital, the dependency is on whoever owns the chips and the power, and a research lab founded to keep this technology out of the hands of a few is obliged to raise from exactly the few who can fund it. Hao is at her strongest tracing that drift, because nobody in the story had to be cynical for it to happen.
Somebody has to look at the worst of it.
The reporting that stays with you is from Kenya. Models do not learn to refuse harmful material on their own: the material has to be labelled first, which means people sitting shift after shift reading and classifying the most degrading text and imagery on the internet, for a few dollars an hour, on outsourced contracts, with the psychological cost carried entirely by them. It is the least visible layer of the supply chain and the most load-bearing. Hao’s framing is precise — the safety a user experiences as a property of the model is in fact a service that specific, underpaid, distant people performed.
Nothing in this industry is automatic. Every layer that looks automated is a place where somebody was paid very little to stand.
The Mdrn UrbanA cloud that needs land, power and water.
The other half of the reporting follows the buildings. Data centres are sited where land is cheap, power is available and local objection is manageable, which in practice means communities with little leverage negotiating against companies with a great deal. Hao follows that story to Chile, among other places, where water use in a dry region became the point of conflict. One caveat belongs here and is worth stating plainly: in late 2025 she issued a correction, having overstated one Chilean facility’s water consumption by a factor of a thousand because of an error in a government document. The correction does not undo the chapter’s argument, and noting it is the right way to read a book that is asking for accountability from others.
A non-profit shell around a commercial engine.
A long thread of the book concerns structure: an organisation founded as a non-profit with a mission to keep this technology safe, wrapped around a business that needed tens of billions of dollars to continue. Hao reconstructs how that contradiction was managed, and how it broke into the open during the board crisis of November 2023, when the same governance designed to be able to stop the company discovered what happens when it tries. Read it as the general case rather than one company’s drama: safety structures that depend on the goodwill of the people they exist to constrain are a familiar failure mode, and not only in technology.
Read it as a supply chain, not as a verdict.
Be fair to both sides of this. The capability gains are real, the tools are genuinely useful, and plenty of serious researchers dispute the empire framing as too neat for an industry with several competing labs, open models and a great deal of published science. What survives the disagreement is the accounting. If you build with these systems — and increasingly everyone does — the honest position is not to refuse them but to know the supply chain behind them: whose text trained it, who labelled the harmful material, where the electricity came from, and who in the chain has the least ability to say no.
Ask of any technology what you would ask of an empire.
Who is paying for this, who is being told it is for them, and who would have to agree before it could be stopped. Those three questions survive any argument about whether Hao’s analogy is exact, and they are the most portable thing in the book — as useful for a procurement decision in Kuala Lumpur as for a hearing in Washington.
Summarised in my own words from Empire of AI by Karen Hao (Allen Lane, 2025; US edition subtitled Dreams and Nightmares in Sam Altman’s OpenAI, Penguin Press). Winner of the National Book Critics Circle Award for Nonfiction. A correction the author issued in December 2025 regarding water use at a Chilean data centre is noted in section 04.