Empire of AI; Review

The Moment That Stopped Me

The moment came about a third of the way in. Hao is sitting with a man named Mophat Okinyi in Nairobi. A vendor called Sama hired him on OpenAI's behalf to read the worst text the model could be pushed to produce, hour after hour, so ChatGPT would learn to refuse it. The pay was under two dollars an hour. TIME's Billy Perrigo put that number on the record first, and Hao follows it home, into what the work did to the man. Okinyi came out the other side of the contract changed. I put the book down right there.

I spent years around intelligence systems in rooms where the data was not allowed to leave the building, and that kind of work cures the reflex of awe. You quit asking what the machine can do and start asking who's doing the work. Hao went and found out. She flew to Nairobi, to Caracas, to the dry valleys of central Chile, to the places where the cost of the miracle actually lands, and she wrote down what she found.

The Thesis

Her title is the argument. She means empire in the old sense. A power claims a frontier and calls the taking a civilizing mission, pulls cheap labor and raw material off the land, and routes the wealth back to a small group that calls itself humanitarian. She spends four hundred pages showing OpenAI running that pattern.

The civilizing mission is AGI, a general machine intelligence that stays a few years away no matter what year it is. Hao is sharp on how the company runs that promise as a faith. Some inside believe the machine will hand us paradise, and some believe it could end the species, and both camps land on the same working doctrine: the stakes are so large that OpenAI alone must build the thing, and the normal rules should not apply on the way. The company was founded as a nonprofit with a safety board seated above the chief executive, in case they really were building the most dangerous technology in history. By the end of the book that structure has been cashed in for growth capital, and Microsoft's thirteen billion dollars bought the kind of leverage that makes a board's firing power evaporate inside a week. The book opens on those five days in Nov 23 when the board fired Sam Altman for not being "consistently candid," then watched him walk back through the door.

Where It Lands

Start with the round trip, because once you see it you can't unsee it. The model was trained on billions of answers human beings posted on the open internet for free, people helping other people for nothing. OpenAI scraped that. Then it shipped the raw output to Kenya and Venezuela, where more human beings cleaned and labeled it for a dollar or two an hour. Then it put the result behind an API, priced it, and marketed it as the arrival of machine intelligence. Humans wrote the knowledge, and humans scrubbed it. The company's original contribution was the fence and the sales pitch, and the valuation on that contribution now runs toward a trillion dollars.

Hao makes the labor end of that pipeline specific, and this is the reporting nobody else went to get. Okinyi in Nairobi, screening trauma so the assistant sounds calm to a subscriber in California who pays more each month than Okinyi made in an hour. In Venezuela, an economic collapse produced a pool of educated, desperate people who labeled data through gig platforms until the work dried up overnight and stranded them. When Scale AI's Remotasks platform pulled out of Kenya in Mar 24, thousands of workers lost the income with no notice at all.

She goes to the physical plant too. Data centers are rising across central Chile and Uruguay, pulling power and water out of ground that has neither to spare, while Chile sits in a drought it has been measuring in decades. The land and the tax breaks get offered, the water gets negotiated away, and the value lives offshore, the way it always has. That chapter is where the word empire quits being a metaphor.

There's an old idea sitting under all of this, and Hao circles it the whole book without naming it. Alan Watts liked to say people mistake the menu for the meal, the description of a thing for the thing itself. A language model is a description of human knowledge. The industry sells the description as if it were the knowing, and keeping that sale alive requires the humans who did the knowing to disappear from the story. The people who wrote the training data and the people who cleaned it get cropped out. Hao's whole method is walking them back in, one name at a time.

She's fair to the believers, and the book is stronger for it. She lets the AGI faithful talk at length, and the faith comes through sincere. These are engineers who think they're delivering salvation, and she never sneers at them. She just keeps reporting until you see what the belief does in practice. A god under construction never has to explain the wage in Nairobi or the water bill in Chile. The cause is too big, and the clock is too short.

Where It Misses

Two honest problems, and both read as gaps.

The first is arithmetic. Hao sometimes lets the indictment outrun the numbers. Her most-quoted figure for data center water use in Chile looks high, and Andy Masley picked the specific number apart in print. The worry underneath it is real, one facility asking for more water than a town in the middle of a fifteen-year drought. But when the moral case is this strong the math has to be airtight, and one soft number hands the industry a rebuttal it never earned.

The second is the one I feel from where I work. When she turns from diagnosis to cure, she points to Te Hiku Media, a Māori-led project building language tools owned and governed by the community whose speech trained them. Right model, and a real one. She leaves it at community scale. The same logic holds one layer up, in the small shop and the integrator who could run their own right-sized intelligence on their own terms, and that chapter never comes. In fairness, it was never hers to write.

The Trades View

Here's the complement I'd add from the field. Small intelligence, held close, trained on one real domain, and run by somebody who knows that domain beats the giant central oracle run by somebody who doesn't. I've worked systems where the data could not leave the building, and the lesson from that work is plain. The model is a tool, and the operator decides what the tool is worth. The empire's pitch erases the operator on purpose, because an erased operator buys the miracle forever.

The market has started confessing on its own. These companies spent two years preaching the end of the coder. In spring 25 the head of Anthropic said AI would write ninety percent of code within six months, and the real figure came in near thirty. IBM's chief executive put it at twenty to thirty. This past May, Altman said he was "delighted to be wrong" about the jobs apocalypse he'd spent years forecasting; the entry-level desks he expected to sit empty were still full, and his counterpart at Anthropic reversed in the same window.

Now put those reversals on a calendar. Anthropic filed confidentially for a public offering on 1 Jun 26, fresh off a raise at a $965 billion valuation. OpenAI filed on 8 Jun 26, carrying its own forecast of a fourteen-billion-dollar loss for the year and no profit expected before 2029. The prophecy got walked back the same season the offering paperwork went in, and both companies are hiring software developers as fast as they can. Hao's empire thesis doesn't need my help, but that calendar backs her up.

Verdict

This is the most honest account I've read of what this industry is and what it costs. Hao earned it by going to the two rooms everybody else skips, the boardroom for the belief and the far end of the supply chain for the bill. I finished it admiring the reporter and cold toward her subject, and I think that's the reaction she was working for.

Read it for the reporting. Then hold on to the oldest thing in it: the work was human the whole way through, and the people who did it belong in the story. The chapter about working people running this technology on their own terms still needs writing. That one's on us.