Empire of AI

Picture yourself staring at a shiny new chatbot, amazed by how human it sounds. Now picture what hides behind that single sentence on your screen: oceans of scraped text, a Kenyan worker reading descriptions of violence for two dollars an hour, a reservoir drying up in Uruguay, and a small group of billionaires deciding the future for everyone else.

This is the story Karen Hao tells about OpenAI, the company that gave the world ChatGPT and rebranded artificial intelligence as our shared destiny. It began as a non-profit promising to save humanity from dangerous machines. It became something else entirely: a corporate empire built on extraction.

You will follow how Sam Altman, once the polite outsider in Silicon Valley, used fear, faith, and friendship with Microsoft to centralize power over a technology he calls AGI. You will see the data workers, the drained aquifers, the fired scientists, and the boardroom coup that almost ended it all. By the end, you will know why the polished interface in your browser is the smallest part of the story.

The Billionaire's Fear and the Non-Profit Mirage

In 2015, Elon Musk and Sam Altman shared the same nightmare. Google had just bought DeepMind, and to them that meant one company could soon own the most powerful technology in human history. Their answer was OpenAI, a non-profit lab that would, in theory, build artificial general intelligence for everyone.

The promise was open science. The reality, from day one, was Silicon Valley orthodoxy. Altman had been raised inside Y Combinator under Paul Graham, a mentor who preached monopoly and hypergrowth as virtues. He spoke the language of safeguarding humanity while importing the playbook of aggressive startups. That contradiction attracted brilliant researchers like Ilya Sutskever, who joined believing the work would be transparent and shared.

Meanwhile, a different warning was being ignored. Timnit Gebru, one of the few Black researchers in the field, kept pointing out that AI was being built by a tiny, homogeneous group whose blind spots were getting baked into the models. The OpenAI founders preferred to worry about Superintelligence, a far-off existential threat, rather than the real biases harming real people right now. That choice of which fear to take seriously would shape everything that followed.

The One Billion Dollar Pivot

Building powerful models requires staggering amounts of computing power, and computing power costs money the non-profit simply did not have. So in 2019, OpenAI quietly invented a new structure called capped-profit. Investors could now earn returns, just not unlimited ones. The mission language stayed; the financial logic flipped.

To seal the deal, OpenAI needed to impress Microsoft, and specifically a skeptical Bill Gates. The team built a demo where the model answered AP Bio questions with eerie fluency. Gates was convinced. One billion dollars arrived. The "open" in OpenAI quietly began to mean something closer to "exclusive partner of Microsoft."

To feed these new models, the company helped normalize a practice Karen Hao calls data colonialism. Vast amounts of text, images, and code were scraped from the internet without asking the people who created them. Just as earlier empires extracted minerals from distant lands, this new empire extracted human expression at planetary scale, then sold it back to the same people as a service.

The Religion of Scale

Inside the company, a quiet faith took hold. Ilya Sutskever and his team began to believe that intelligence was, at bottom, a matter of size. Take a neural network, pour in more data, add more compute, and the thing would simply get smarter. No new theory required.

They tested this belief using the Transformer architecture, which had originally been invented at Google. The result was GPT-1, then GPT-2, each one validating what the team began calling the scaling laws. The correlation between compute, data, and performance was real and predictable. Brute force was working.

When GPT-2 was ready, OpenAI did something unusual. It refused to release the full model, citing risks of misinformation. Critics noticed something else: withholding the model generated headlines and mystique. Safety, in practice, doubled as marketing. Inside the building, Altman was running the company like one of his old Y Combinator startups, demanding speed and dominance, while a quieter faction led by Dario Amodei worried the safety culture was being hollowed out.

Sacrificing Science for Product

The bigger OpenAI grew, the harder it became for outsiders to study what was happening. Independent academic research depends on access to compute, and a handful of corporations now controlled almost all of it. Universities were left behind.

Two researchers refused to stay quiet. Emma Strubell published numbers showing the enormous carbon footprint of training large language models. Then Timnit Gebru coauthored a paper titled "Stochastic Parrots," which laid out the environmental damage, the racial and gender biases, and the basic limits of these systems. Google fired her. The message to every researcher watching was unmistakable.

Inside OpenAI, the pressure to ship was relentless. The GPT-3 API turned the lab into a vendor selling access to text generation. Codex, built with Microsoft, automated programming by training on open-source code without asking the programmers who wrote it. For Dario Amodei and a group of safety-focused colleagues, this was the breaking point. They walked out and founded Anthropic, a rival lab built around the safety principles they felt OpenAI had abandoned.

The Global South Sweatshops

Before ChatGPT could charm the world, someone had to teach it not to say horrifying things. That job did not go to engineers in California. It went to workers in Kenya, hired through a company called Sama for less than two dollars an hour.

Their task was to read the worst material the internet had produced — graphic descriptions of sexual abuse, torture, and murder — and label it so the model would learn to refuse. One of these workers, Mophat Okinyi, has spoken publicly about the lasting psychological damage. His family fell apart. Many of his coworkers carry symptoms of severe trauma. They cleaned ChatGPT so it could feel polite to you.

The pattern repeated elsewhere. Scale AI built much of its workforce in Venezuela during the country's economic collapse, paying in dollars to people who had no alternative, then suddenly cutting wages through platforms like Remotasks Plus once dependency was established. Inside Silicon Valley, the Effective Altruism crowd worried publicly about hypothetical AI doomsday. The actual suffering caused by AI development happened far away, where the cameras did not point.

Draining the Earth for Compute

In late 2022, OpenAI quietly released ChatGPT as a "research preview." Within two months it had 100 million users. The company had not expected this, and neither had Microsoft. Suddenly generative AI was a global gold rush, and OpenAI had to build infrastructure on a scale no one had attempted before.

That infrastructure has a physical address. Megacampuses full of servers need staggering amounts of electricity and water, and Microsoft committed billions to build them. The cheapest places to build, conveniently, tend to be places already struggling.

In Chile, near the Atacama Desert, communities organized through groups like MOSACAT have fought against data centers that would draw from aquifers already depleted by copper and lithium mining. In Uruguay, during a historic drought, residents watched tap water turn salty while officials negotiated secret deals to supply cooling water to tech facilities. The cloud is not a cloud. It sits on land, drinks from rivers, and the bill arrives first in places that were never asked.

The Illusion of Regulation and Personal Demons

In 2023, Sam Altman went on what the press called Sam Altman's World Tour. He met presidents, prime ministers, and senators. He talked about safety. He warned about bioweapons and cyberattacks and the dangers of frontier models falling into the wrong hands. He proposed regulations based on compute thresholds and licensing.

Read carefully, his proposals had a convenient shape. They would burden any open-source competitor or new entrant while leaving OpenAI, Google, and Anthropic free to keep operating. National security fears against China were used to argue that only a few trusted corporations should be allowed near the technology. Meanwhile the immediate harms — stolen copyrighted work, displaced creative workers, opaque training data — got almost no airtime.

The contrast became sharper when Altman's estranged sister, Annie Altman, began speaking publicly about her life. She described housing insecurity, health problems, and turning to platforms like SeekingArrangement to survive, while her brother promised the world an age of abundance powered by AGI. The man selling utopia to humanity was reportedly cold and calculating with his own family. Inside OpenAI, a small group of leaders was starting to notice the same pattern.

The Weekend OpenAI Almost Died

By late 2023, the trust inside OpenAI's executive suite had quietly collapsed. CTO Mira Murati had grown exhausted by what she described as Altman's tactical games and broken promises. Ilya Sutskever, the company's chief scientist and one of its founders, had reached a similar place. Both began speaking, separately and confidentially, to the independent board members Helen Toner and Adam D'Angelo.

On a Friday in November, the board fired Sam Altman. The official reason was that he had been "not consistently candid" with them. For a few hours, it looked like the original mission had reasserted itself.

Then the weekend happened. Greg Brockman resigned in solidarity. Microsoft's Satya Nadella offered to hire Altman and the entire team. A pending employee stock sale meant that everyone at OpenAI was suddenly staring at a personal financial catastrophe. Almost the entire staff signed a letter threatening to quit. Murati wavered. Sutskever, surrounded by furious colleagues, publicly regretted his role. By Tuesday, Altman was back, with a new board stacked with allies and business veterans. The mutiny had failed.

The Empire Consolidated

The aftermath made the new order plain. Jan Leike, who led the team responsible for aligning powerful future systems, resigned and said safety culture had taken a back seat to shiny products. Ilya Sutskever left soon after. The voices that had warned about existential risk, the very people the company was originally built to empower, were gone.

The launch of GPT-4o brought another scandal. Its default voice sounded uncannily like Scarlett Johansson, who had explicitly refused to license her voice. Around the same time, leaked documents revealed that departing employees were being threatened with clawbacks on their equity if they refused to sign sweeping non-disparagement agreements. Speak critically and lose your savings. The pressure worked for years before it became public.

Karen Hao closes by naming the pattern directly. She calls it "A Formula for Empire": declare a vague, world-saving mission, monopolize the resources needed to pursue it, and use safety as the reason no one else can be trusted with the work. OpenAI was no longer the alternative to Big Tech. It was the new Big Tech, just wearing a different origin story.

Choosing Who Decides

The future of AI is not a force of nature arriving on schedule. It is a series of choices made by very few people, with very specific incentives. Knowing the labor, the water, the silenced researchers, and the hidden contracts behind the magic gives you something the empire would rather you not have: the right to ask who benefits, who pays, and who was never consulted at all.