Empire of AI

Karen Hao · 2025

Thesis

The leading artificial intelligence companies, OpenAI foremost, have become a new kind of empire: in pursuit of ever-larger models, they lay claim to data, land, energy and water that are not theirs, depend on cheap and hidden labor, and concentrate knowledge and power in a few hands, all justified by a quasi-religious belief in the coming of artificial general intelligence.

The main ideas

AI companies resemble empires in structure, not just by metaphor: they seize resources, exploit labor and monopolize knowledge.
OpenAI was founded as an open, nonprofit counterweight to big technology companies and became one of them.
The push for scale, ever more data and computing power, drives the costs the book documents.
Belief in artificial general intelligence divides insiders into those who want to build fast and those who fear it, but both conclude they must control it.
The 2023 firing and return of Sam Altman exposed how little the company's governance could check its leader.
The costs fall on workers and communities far from Silicon Valley, from data labelers in Kenya to towns in Chile facing data centers.

The empire

Karen Hao studied mechanical engineering at MIT before turning to journalism. As a senior artificial intelligence reporter at MIT Technology Review, she embedded at OpenAI for three days in 2019 and wrote what she describes as the first profile of the company, published in early 2020. She found it controlling about what she could see and whom she could speak to, despite its claims of openness. She later covered technology for the Wall Street Journal.

Empire of AI, published in 2025, draws on more than 300 interviews with about 260 people, along with correspondence and internal documents. OpenAI declined to cooperate. The book won the National Book Critics Circle Award for nonfiction.

Hao argues that the comparison to empire is structural. Like the colonial empires, AI companies lay claim to resources that are not theirs, including data, land, energy and water; they exploit large amounts of labor; and they monopolize the production of knowledge by drawing in talent and research.

Implication: judge an AI company by what it takes and from whom, not only by what its products can do.

From nonprofit to empire

OpenAI was founded in December 2015 as a nonprofit research laboratory by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever and others, promising to develop artificial intelligence for the benefit of humanity and to share its research openly.

Hao traces how that mission changed. Building more capable models required ever more data and computing power, which required money on a scale no nonprofit could raise. In 2019 OpenAI created a capped-profit company under the nonprofit, and Microsoft invested $1 billion. The company became more secretive, more commercial and more focused on scale.

The demand for scale, Hao argues, drove many of the harms she documents, pushing the company toward lower-quality data scraped from the internet, toward more labor to clean it, and toward vast data centers.

The mechanism is that a strategy built on scale turns every constraint, from data to electricity, into something to be acquired, and the costs of acquiring it fall on others.

Implication: when a company's strategy depends on scale, ask who pays for each resource it needs to grow.

The firing

The book opens with the board's firing of Sam Altman in November 2023, an episode the publisher says is told in full for the first time. Within days, after most of the staff threatened to leave, Altman was reinstated and the board largely replaced.

Hao's account gives a central role to Sutskever, OpenAI's chief scientist, and Mira Murati, its chief technology officer, both of whom had raised concerns about Altman's leadership. In an excerpt published in The Atlantic, Hao reports that Sutskever told colleagues, "I don't think Sam is the guy who should have the finger on the button for AGI." She also reports him saying at a 2023 meeting, "We're definitely going to build a bunker before we release AGI." Sutskever later reversed his position.

The mechanism, in her account, is that a governance structure designed to restrain the company's leader failed the first time it was tested, because the leader commanded the loyalty of staff and investors.

Implication: a company's safeguards are only as strong as their ability to survive a confrontation with the person who holds the most power.

Boomers and doomers

Hao describes belief in artificial general intelligence, machines that match or exceed human abilities, as quasi-religious. Within OpenAI and the wider industry, she identifies two factions. Boomers want to build and release quickly to bring the benefits sooner. Doomers fear the technology could be catastrophic and want more time for safety.

The two camps argue fiercely, but Hao finds that they share an assumption: that artificial general intelligence is possible and coming soon. From that premise, both conclude that the technology must be controlled by people like themselves rather than shaped democratically. In her account, the boomers have largely won.

The mechanism is that an argument conducted entirely within one belief leaves that belief unquestioned, so the debate over how to build the technology crowds out the question of whether, and on whose terms.

Implication: when insiders debate the speed of a technology, ask what assumptions both sides share and who is left out of the decision.

The hidden labor

Systems like ChatGPT depend on human workers who label data and filter out harmful material. Hao follows these workers to the margins of the global economy.

In Kenya, OpenAI contracted the outsourcing firm Sama in late 2021 to label toxic text, descriptions of sexual abuse, hate speech and violence, so that its models could learn to avoid producing it. Workers were paid less than $2 an hour, and some described lasting psychological harm. Hao tells the story of Mophat Okinyi, one of these workers.

Her earlier reporting followed data labelers in Venezuela, where economic collapse had created a pool of educated workers willing to accept very low pay on labeling platforms.

The mechanism is that companies whose products appear automatic rely on hidden human labor, and they buy it where workers have the fewest alternatives.

Implication: when a technology seems to work by itself, look for the people doing the work behind it and the terms they work on.

Water and land

The last cost Hao documents falls on communities near data centers, which need land, electricity and, for cooling, large amounts of water.

In Uruguay, during a severe drought in 2023, when Montevideo's tap water became saline, Google's planned data center was disclosed to need about 2 million gallons of water a day from the drinking supply. Google later scaled back and switched to air cooling. In Cerrillos, near Santiago in Chile, local activists challenged a Google data center; in 2024 an environmental court partly revoked its permit over its water use, and Google paused the project to redesign it.

The mechanism is that the physical demands of large AI models are met in places with cheap land and power and weak bargaining positions, often in the Global South.

Implication: ask where a technology's physical infrastructure is built and what local communities give up for it.

Limits

One of its water figures was badly wrong. The researcher Andy Masley showed that the book overstated a planned data center's water use relative to the town of Cerrillos by a factor of about 1,000, and Hao acknowledged the error. Masley and other critics argue that the book's case on water is weaker than it appears.
The empire framing has been questioned. Critics argue that the analogy to colonial empires is sometimes forced, since an AI company is not a state, and that the labor chapters rely on moving individual stories selected to fit it.
It is built without OpenAI's cooperation. Its portrait of Altman is darker than that of other accounts published at the same time, and much of it rests on sources hostile to him.

Careless People — an insider's account of another technology company's power and indifference to its effects.
Zero to One — the case for building monopolies, the strategy Hao sees AI companies pursuing.