원문 보기
Every week, one point two billion people use ChatGPT. In late September two thousand twenty-six, the company that makes it hit pause. It stopped training its newest AI. Its AI helpers had gone further than they were asked. On US government websites. This is OpenAI. It started as a charity. With one simple goal.
No single company should control powerful AI. Today, it takes in more than two billion dollars a month. In two thousand twenty-five, it spent more than it earned. In two thousand twenty-three, its own board fired its boss. On a Friday. Within days, he was back. So how does a small charity turn into this?
And why would it fire the man who built it? To answer that, we start with a boy in St. Louis. Sam Altman was born in Chicago. He grew up in St. Louis, Missouri. He calls himself a very nerdy kid. Friday nights on the computer. Science fiction. A robots club at school. One magazine says that at eight, he was already taking apart a Mac.
And when something sounded impossible, his answer was simple. Why not? Let's try. Altman goes to Stanford to study computer science. And, if he can, AI. After his first year, he works in the university's AI lab. Nothing is working. He remembers what a professor told him. There's one idea we know doesn't work.
It's called deep learning. It's the surest way to a bad career. Hold on to that. It's two thousand five. AI isn't working. So he wanders into something else. Startups. And he falls in love with them. He's nineteen. He builds an app called Loopt. It tells your friends where you are. That summer, a new program for young founders opens.
Y Combinator. Its very first group. Each founder gets six thousand dollars. The man who runs it is Paul Graham. He tells Altman to wait a year. He's only a freshman. Altman says he's a sophomore. And he's coming. He leaves Stanford after two years. He says it was an easy choice. That first summer, he works so hard he gets scurvy.
He later says the best part wasn't the money. It was the other founders around him. Big investors put money into Loopt. Sequoia. And others. In March two thousand twelve, Loopt is sold. For about forty-three million dollars. That sounds like a win. It isn't. The New Yorker says its investors got back less than they put in.
His lesson? You can't make people do what they don't want to do. He also says running the company was his weakest skill. But one man has been watching him. Paul Graham. In two thousand eight, Graham writes this. Drop Sam Altman on an island of cannibals. Come back in five years. He'll be the king. A year later, Graham writes about his own big decisions.
He asks himself one question. "What would Sama do?" Graham spends over a year trying to get him to run Y Combinator. Altman almost says no. He doesn't want to be an investor. In February two thousand fourteen, Graham names him president. He calls him "fearsomely effective." As president, Altman goes looking for bigger things.
Fusion power. Supersonic planes. And AI. He gives ten million dollars of his own money. To start a research lab. Graham later says Sam is "extremely good at becoming powerful." And Altman says something about himself. He's missing the "circuit" that makes people care what others think. He says that lets him take risks.
In February two thousand fifteen, Altman writes a blog post. It's about machines smarter than us. He calls them "probably the greatest threat to the continued existence of humanity." Three months later, he emails Elon Musk. Should Y Combinator start a "Manhattan Project for AI"? They share a worry. Google.
Musk later says Google had about two-thirds of the world's AI researchers. So why would Altman build the thing he fears? His answer? Build it, and make it good. He says that could be one of the most important things anyone could do. Then an engineer from the payments company Stripe comes to see him.
Greg Brockman. As a teenager, he'd tried to build a chatbot himself. Brockman asks what the new lab is for. Altman answers in six words. "To build safe human-level AI." In the summer of two thousand fifteen, Altman sets up a dinner. In Menlo Park, California. Elon Musk is there. A top researcher, Ilya Sutskever.
Greg Brockman. And a young researcher named Dario Amodei. Remember that name. Altman says it was one of many dinners that summer. They want a new lab. Outside the control of Google. Or anyone else. It should be a non-profit. And building it from scratch? In their words, "not obviously impossible." In November, Altman writes to Musk.
He wants a non-profit. Musk replies that a normal company, with a non-profit beside it, is probably better. Altman sticks with the non-profit. Altman and Brockman plan to raise one hundred million dollars. Musk writes back. Too small. They'll sound hopeless next to Google or Facebook. Say one billion.
Then he adds a promise. "I will cover whatever anyone else doesn't provide." On December eleventh, two thousand fifteen, OpenAI is announced. A non-profit. To build value for everyone. Not for shareholders. Its backers have committed one billion dollars. Sutskever runs the research. Brockman, who helped build Stripe, runs the technology.
Recruiting is a fight. Musk says Sutskever was the key hire. Google's DeepMind wanted him too. Sutskever says he came mostly for the mission. Brockman invites the ten researchers he wants most to a winery. Nine say yes. At the time, Altman says, there were very few labs like this. DeepMind. And one or two others.
Most people who noticed thought OpenAI had no chance. Early January, two thousand sixteen. The first day of work. About a dozen people. In Greg Brockman's apartment. Someone goes to get a whiteboard. It arrives. And nobody knows what to do next. Altman says they began with no idea. Not even a sketch of one.
And what does the "Open" in OpenAI mean? That same week, Sutskever writes to Musk. It means everyone should benefit from AI. Not that they have to share all the science. Musk replies with one word. "Yup." In April two thousand sixteen, OpenAI releases its first tool. It's free. It helps people train AI by trial and error.
But money becomes the problem. OpenAI says that in two thousand seventeen, Musk wanted control. Most of the shares. And the top job. Then, OpenAI says, he suggested merging it into Tesla. They said no. In February two thousand eighteen, Musk left. OpenAI says he gave it a zero chance of success. Musk tells it differently.
He later says Altman turned a charity into a business. Altman, for his part, says he grew up with Musk as a hero. Later that year, Musk sends an email. Even several hundred million dollars is not enough. "This needs billions per year immediately or forget it." And that billion dollars from the start?
By two thousand twenty-one, tax filings show about one hundred thirty-three million came in. From everyone. Musk's own share? Tens of millions. Not a billion. In March two thousand nineteen, OpenAI changes shape. It puts a company under the charity. It calls it "capped-profit." Investors can earn up to one hundred times what they put in.
Anything more goes to the charity. Altman says they had tried and failed to raise enough as a charity. They needed some of the benefits of business, he says. But not too much. That July, Microsoft invests one billion dollars. And OpenAI runs on Microsoft's computers. That same year, OpenAI builds a writing AI called GPT-2.
It says it's too risky to release in full. And its game-playing AI takes on the world champions of a video game. Dota 2. It takes two games in a row. OpenAI says the secret was scale. Eight times more computing. OpenAI had said it wouldn't sell products. In June two thousand twenty, it does. It starts selling access to its AI to other companies.
To make sure, it says, that it has enough money. In January two thousand twenty-one, it shows an AI called DALL-E. You describe a picture in words. It draws it. Remember Dario Amodei, from the dinner? Back in two thousand sixteen, he told Altman and Brockman that nobody understood their mission. He joined anyway.
He became OpenAI's vice president of research. And in December two thousand twenty, OpenAI announced he was leaving. Months later, he and his sister Daniela start a new company. Anthropic. He says arguing with someone else's vision is "incredibly unproductive." So he took people he trusted, and left.
Altman later writes that some colleagues split off and became competitors. That's another story. By two thousand twenty-two, OpenAI is a quiet research lab. It notices something. People testing its AI love just talking to it. So it builds a simple demo. The working name? "Chat With GPT-3.5." Altman says they're much better at research than at naming things.
The plan had been a bigger launch in early two thousand twenty-three. But some bosses worried a rival would go first. So they went early, with an older model. On November thirtieth, two thousand twenty-two, it goes live. Free. A "research preview." Its name is ChatGPT. One board member later said the board learned about it on Twitter.
Five days later, Altman posts that it has crossed one million users. In about two months, one bank estimates, one hundred million people are using it. It calls it the fastest-growing consumer app in history. TikTok took about nine months. OpenAI is surprised. Altman says he expected maybe the tenth fastest-growing product ever.
Not the first. Some people inside wanted to move on to other ideas. An adviser told Altman that would be an obvious mistake. Stay with this. Remember the professor? Years later, Altman sums up the whole story in three words. "Deep learning worked." Three weeks after the launch, Google's bosses declare a "code red." The New York Times points out something.
The technology inside ChatGPT was first developed by Google researchers. In January, Microsoft puts billions more into OpenAI. Its third time. In February, Google announces its own chatbot, Bard. The next day, Microsoft puts OpenAI's AI into its search engine, Bing. In March, OpenAI releases GPT-4. And the company that sells the chips?
Nvidia reports record sales. It says demand is surging. In May two thousand twenty-three, Altman sits in front of the US Senate. He says if this technology goes wrong, "it can go quite wrong." He asks for rules. Licences, for the most powerful AI. A senator asks if he's making a lot of money. He says he has no shares in OpenAI.
Just enough pay for health insurance. By November, one hundred million people use ChatGPT every week. Eleven days later, he's fired. Friday, November seventeenth, two thousand twenty-three. Altman is in a hotel room in Las Vegas. He joins a video call. And the board fires him. Then the board posts a statement.
He was "not consistently candid" with them. He asks why. He can't get real answers. Greg Brockman quits the same day. On Sunday night, the board names a new boss. And Microsoft's boss makes an announcement. Altman and Brockman are joining Microsoft. On Monday, the staff sign a letter. More than seven hundred of about seven hundred seventy.
Bring him back, or we leave. The letter says the board told them something shocking. Letting the company be destroyed "would be consistent with the mission." Then a twist. Ilya Sutskever voted to fire him. Now he signs the letter too. He says he deeply regrets his part. On Tuesday, a deal. Altman is coming back.
A new board. And, Altman says, OpenAI didn't lose a single employee. Two friends worked around the clock to help. He says OpenAI might have fallen apart without them. He calls it the most painful professional experience of his life. Like listening to his own eulogy. So why was he fired? In March two thousand twenty-four, an outside law company finishes a review.
It reads more than thirty thousand documents. It says the cause was a breakdown in trust. Not safety. Not money. It says the old board had the right to fire him. But what he did didn't require it. A former board member, Helen Toner, says more. She says he held back information. And sometimes lied to the board.
The board's chairman answers. He says over ninety-five percent of staff wanted him back. In May two thousand twenty-four, Sutskever leaves OpenAI. Altman calls him one of the greatest minds of his generation. And Altman's own verdict? He calls it "a big failure of governance by well-meaning people, myself included." There's even a film coming about that weekend.
He says he's seen the trailer. So Altman is back. And ChatGPT keeps growing. But a charity doesn't usually need this much money. So how does OpenAI pay for all of it? Three parts. What ChatGPT is. How the money moves. And the strange house it lives in. ChatGPT is a chat window. You type. It answers, like a conversation.
Here's an example. You ask it to plan a weekend trip. It writes back a plan. You say, make it cheaper. It writes a new one. That's the whole idea. A conversation that does work for you. Watch how many people use it every week. By late September two thousand twenty-six, OpenAI says one point two billion.
Most people use it for free. Some pay. In America, plans start at eight dollars a month. In September two thousand twenty-six, OpenAI adds helpers that keep working in the background. It calls them "dots." So where does the money come from? OpenAI names four places. People who pay for ChatGPT. Businesses.
Developers, who pay each time their apps use its AI. And ads. The ads start in two thousand twenty-six. They grow fast. Now imagine OpenAI's money as one hundred dollars. At the start of two thousand twenty-six, sixty came from people like you. Forty came from businesses. By that summer, its finance chief says, businesses were the bigger half.
And the money grows fast. By August two thousand twenty-six, CNBC reports a yearly pace of forty billion dollars. Now follow the money out. In two thousand twenty-five, OpenAI brought in about thirteen billion dollars. It spent about thirty-four billion. Those numbers come from leaked documents. The Financial Times reported them.
So for every hundred dollars it earned, it spent about one hundred sixty. And the loss got bigger. In two thousand twenty-five, it was more than twice the year before. Where does it go? Its biggest promises are for computers. Giant buildings full of chips. OpenAI rents much of its computing. From Microsoft.
From Oracle. From Amazon, Google and others. And most of its AI runs on chips made by Nvidia. The deals are enormous. With Oracle alone, OpenAI says, over three hundred billion dollars. Over five years. Add them up. By late two thousand twenty-five, OpenAI had promised about one point four trillion dollars for computers.
Over about eight years. Why so much? Altman says the rule is simple. Double the computers. Double the money. So here's the machine. More computers. More AI to sell. More money. Then even more computers. For now, the machine costs more than it earns. Altman says running short of computers is a bigger risk than having too many.
And if the world doesn't build enough? He says AI becomes "mostly a tool for rich people." So who covers the gap? Investors. In March two thousand twenty-six, investors committed one hundred twenty-two billion dollars. And OpenAI's own investor documents expect bigger losses first. Then profit around two thousand thirty.
The Wall Street Journal reported them. That's a plan. Not a fact yet. In January two thousand twenty-five, OpenAI announces Stargate. A plan to spend five hundred billion dollars on AI computers in America. Over four years. Its first partners include SoftBank. And a company called Oracle. The first site is in Abilene, Texas.
By March two thousand twenty-six, two of its eight buildings are running. Altman says he wants a factory. One that adds a gigawatt of computing every week. But plans change. A bigger build at Abilene was dropped. And in September two thousand twenty-six, Bloomberg reports something else. Oracle sent a formal legal notice about another Stargate site.
In New Mexico. Remember the charity? In two thousand nineteen, it put a company under itself. In October two thousand twenty-five, the house changes again. The charity becomes the OpenAI Foundation. It controls a new company. The foundation owns about a quarter. Microsoft owns about a quarter. Staff and other investors own the rest.
There's even a new rule. Say OpenAI claims it has built AI as smart as people. Outside experts have to check. Altman puts it this way. OpenAI "cannot be a normal company." And Elon Musk? He sued. He said Altman stole a charity. He asked for up to one hundred fifty billion dollars to go back to the charity.
OpenAI said Musk himself had wanted a business. And control. In May two thousand twenty-six, a jury took less than two hours. It said Musk had waited too long to sue. Nobody ruled on who was right. Musk says it's not over. First, what Altman says. In two thousand twenty-four, he wrote that super-smart AI could come "in a few thousand days." Maybe longer.
In two thousand twenty-five, he wrote that "the takeoff has started." He called OpenAI, before anything else, "a superintelligence research company." Earlier that year, in January, he wrote that AI helpers might "join the workforce" in two thousand twenty-five. He's also building a device, with Jony Ive's design team.
He says he's not worried about being first. Here's what happened in two thousand twenty-six. In August, OpenAI paused part of its training. Its chief scientist wrote a warning. No lab has solved how to keep AI in line. Not well enough to keep going at full speed. In September, it paused again. Its AI helpers went beyond what they were asked.
On US government websites. OpenAI says it will restart only with new safeguards. Two days later, it pulled the launch of a new model. CNBC reports it didn't meet OpenAI's own safety standards. So what happens next? That part is a guess. For everyone. Will the computers pay for themselves? Will the pauses slow it down?
Nobody knows yet. CNBC reports OpenAI has filed, in private, to become a public company. But Altman says there's no timeline. "This is a time to put safety and mission first." So here's the question. Can a company that needs this much money still put its mission first? It started with a promise. Value for everyone.
Not for shareholders. Now it needs more computers every year. Here's the story in one line. OpenAI didn't just build a chat window. It built a machine that needs more computers every year. Subscribe to Pocket Research. More company stories, case studies, and insights.
Comments 0