원문 보기
Sam Olman runs the most famous AI company in the world. And a little over a year ago at a dinner with a group of reporters, he asked out loud whether investors had gotten over excited about AI. Then he answered his own question. My opinion is yes. So the man at the very center of the AI boom thinks it's a bubble.
And then several months later, his company went out and raised $120 billion. Now, a bubble is what happens when prices run far ahead of the money actually coming in, held up by a story everyone wants to believe. So, let's look at the money actually coming in. Last year, OpenAI brought in about $13 billion and it spent about $34 billion.
For every dollar in, more than 2.5 went out. Most businesses that did that would be getting a worried phone call from their bank. Open AAI got valued at more than $800 billion. And here's what makes this one stranger than the bubbles you've heard of. A lot of the money keeping Open AI going comes from the very companies that sell it chips and computer power.
Nvidia puts money into OpenAI, and a lot of that money comes back to Nvidia as sales. This summer, Nvidia even disclosed in an official filing that it had promised to cover up to $105 billion of the rent and power bills on a data center campus for OpenAI in Ohio if OpenAI can't pay. When the seller is funding its own customer, the boom looks bigger than it really is.
And the last time sellers did this in a big way during the internet boom of the late '90s, it ended in a crash. This week, we got a small taste of how nervous everyone is. One news report said OpenAI was bringing in less money than people had thought, and within hours, Nvidia lost about $170 billion of its value.
So, in this video, we're going to take Alman's question seriously. Is this a bubble? We'll open the books, follow the money, and find out who's actually holding the bill. Because if it does pop, the people who lose the most may not be the ones who made the bet. Some of them might be you. Welcome to Money Forensics.
We take the money stories everyone's talking about and open the books so you can see how they actually work. If that's your kind of thing, hit subscribe. Now, let's get into it. At that same dinner, Olman also said that someone is going to lose a phenomenal amount of money. He was talking about small AI startups raising money at sky-high prices.
It's natural to assume he meant other people. So, let's start with his own company and with the obvious question. Does OpenAI make money? It does bring money in and a lot of it. Most of it comes from people and businesses paying to use Chat GPT and from companies that pay to build OpenAI's AI into their own products.
But it spends far more than it brings in. Those $13 billion came in last year and about 34 billion went out. And this year, the gap is getting wider, not narrower. In the spring, in just 3 months, OpenAI lost about $12 billion on less than $7 billion of sales. Now, it's worth being fair about where that money goes because it isn't what most people assume.
It isn't that answering your questions is a money loser. Once you take out the cost of running the computers that answer everyone's questions, OpenAI keeps about a third of every dollar of sales. The big losses come from somewhere else. They come from building each new bigger AI model, which takes enormous amounts of computer power, and from research.
Last year, that alone cost about $19 billion, more than OpenAI's entire sales. And most of the people using Chat GPT aren't paying at all. More than 900 million people use it every week. About 1 in 18 of them pays. And Altman has never pretended otherwise. Here he is a little over a year ago. I think the rational thing to do is to just be willing to run at a loss
uh for quite a while and continue to do that, which is also part of why I think it it it's nice not to be public. By OpenAI's own plans, that while lasts until about 2030. That's when it expects to finally start making money. Until then, it needs other people to keep covering the difference. And that raises the question that matters most here.
Who's paying? In March, OpenAI raised about $120 billion in a single round. In return, the investors got a slice of ownership at a price that valued the whole company at about $850 billion. And almost all of that money, about 110 billion of it, was promised by just three companies. SoftBank, the giant Japanese investment firm, Amazon, and Nvidia.
Now, look at the last two. Amazon runs one of the world's biggest cloud businesses, which means it rents out enormous warehouses full of computers called data centers to companies that would rather not build their own. And Nvidia makes the chips that AI runs on. Chips and rented computer power are exactly what OpenAI spends most of its money on.
So when Amazon and Nvidia put money into Open AI, a lot of that money finds its way back to them as sales. If that sounds odd, it shouldn't be completely new. This is how OpenAI started. Its first big backer was Microsoft. Starting in 2019, Microsoft promised about $13 billion, and much of that money reportedly came back to Microsoft as OpenAI paid to use Microsoft's computers.
Microsoft now owns about a quarter of Open AAI, and the money still flows that way. According to OpenAI's own accounts for last year, which were audited and later leaked, OpenAI paid Microsoft about $17 billion in 2025. Remember, it only brought in about 13. So last year, OpenAI paid one supplier more than its entire sales, and investors money made up the difference.
And the suppliers have gotten creative. Last year, the chipmaker AMD struck a deal with OpenAI that lets OpenAI buy up to 160 million AMD shares for a penny each. The catch is that OpenAI earns the right to them as it buys and puts AMD's chips to work and as AMD's share price hits set targets. In other words, the more chips OpenAI buys, the more of AMD it gets to own practically for free.
And AMD is happy to do it because it gets a giant customer. It expects the deal to bring in tens of billions of dollars of sales. Then there's Nvidia. A year ago, its chief executive, Jensen Huang, announced a partnership with OpenAI to build a staggering amount of computer power with Nvidia planning to invest in OpenAI as it went.
This partnership, I mean, this is, you know, monumental in size. There's never been an engineering project, a technical project of this complexity and this scale, um, ever. So, how much of this is going on? One way to measure it comes from a venture investor named Tom Tungus. Last year, he added up all the money Nvidia had committed to its own customers.
It came to about 2/3 of a whole year of Nvidia's sales. That number on its own doesn't mean much. So, here's what it means. In the late '90s, during the internet boom, a company called Lucent made the equipment that telephone and internet networks ran on. and it was a giant of its day. To keep sales growing, Lucent lent money to its own customers, mostly young internet and telephone companies, so they could buy its equipment.
Those loans came to about a quarter of a year of its sales. Its sales looked wonderful. Then the boom ended. Many of those customers went bust and Lucent was left writing off billions of dollars of loans it would never get back. So by Tungus's measure, Nvidia's money in its own customers is proportionally nearly three times what Lucent was.
To be fair to Nvidia, there are real differences. Lucent made loans and Nvidia mostly buys ownership stakes. Nvidia is vastly richer than Lucent ever was. Its biggest customers are companies like Microsoft, Google, and Meta, which pay with real cash from businesses that make huge profits. Nobody credible accuses Nvidia of faking its sales.
And Nvidia itself has said the comparison is unfair. But the point isn't that Nvidia is loosen. The point is narrower. When a seller pays for part of its own demand, the boom looks bigger than it really is. And a boom that looks bigger than it really is is a pretty good description of a bubble. The money already spent is only half of it.
The other half is the money OpenAI has promised to spend. Last year, Alman said OpenAI had committed to about $1.4 trillion dollars of computer power over the following 8 years. Since then, the plans have been trimmed, but they still run to hundreds of billions of dollars by 2030 for a company that brought in 13 billion last year.
Not everyone found that reassuring. On a podcast last fall, the investor Brad Gersonner asked Alman how a company with $13 billion in sales could make $1.4 trillion of spending commitments. Alman said OpenAI was doing well more revenue than that. And then he said this. Second of all, Brad, if you want to sell your shares, I'll find you a buyer.
I I just enough. A little later in the same conversation, Gersonnner suggested that OpenAI might reach a hundred billion a year in sales by 2028 or 2029. Alman's answer was, "How about 27?" We'll come back to how that's going. So, who's actually on the hook for all those promises? It helps to think of the companies around OpenAI in three roles.
The first is the landlord. That's Microsoft. It rents OpenAI computer power. Open AAI uses it and Microsoft sends the bill. It gets paid as it goes and it has a giant profitable business of its own. If OpenAI stumbled, Microsoft would feel it, but it isn't betting the company. The second is the builder.
That's Oracle. For decades, Oracle was best known for business software, the programs big companies use to run their accounts and records. In the last few years, it has gone into building data centers. And in 2025, it signed a contract reportedly worth about $300 billion over 5 years to build computer power for Open AI.
And by the count of the rating agency S&P, that one customer is now roughly half of all the future business Oracle has lined up. Here's the strange part. In Oracle's official reports to investors, that customer is never named, not once. To build all this, Oracle has had to borrow and spend faster than it earns.
In its last quarter alone, it spent about $5 billion more on building than its whole business brought in. In July, S&P cut Oracle's credit rating to one step above what's called junk, which is the grade for borrowers considered at real risk of not paying their debts back. In September, the cost of insuring against Oracle not paying back its debts hit a record.
In other words, lenders have never been more worried about getting their money back from Oracle. And at one of the data centers it's renting for this buildout in New Mexico, Oracle has asked for the right to delay its rent by up to 3 years if construction runs late. At least one of the banks that lent money for that site has sold its loans for less than 90 cents on the dollar.
The third role is the co-signer, and that's Nvidia. If you've ever co-signed a loan for a relative, you know the deal. If they don't pay, you do. In August, in its official quarterly filing, Nvidia disclosed that it had guaranteed payments on 20-year leases for a giant data center campus in Pike County, Ohio, being built for OpenAI.
The guarantee covers parts of the rent and the power bills up to $105 billion. And if OpenAI doesn't pay, Nvidia says it may have to take over the lease itself. So the company that sells OpenAI its chips also invests in OpenAI and now also stands behind its rent. Investor, supplier, and co-signer allin one.
All of which raises an awkward question. If OpenAI can't pay and the builders and co-signers get into trouble, does anyone catch them? Last November, OpenAI's chief financial officer, Sarah Frier, was asked about financing at a Wall Street Journal conference, and she floated an answer. backs stop the guarantee that allows the financing to happen
for some federal backs stop for chip investment. Exactly. A backs stop from the federal government. In other words, taxpayers. There was an outcry and that same evening she took it back saying OpenAI was not seeking a government backs stop and that the word had muddied the point. The next day, Alman wrote that taxpayers should not bail out companies that make bad business decisions.
and the White House's adviser on AI, David Saxs, put it bluntly. There will be no federal bailout for AI. So, officially, nobody's going to catch them. It's just worth noticing that the question came up at all and that it came from OpenAI's own finance chief. Now, you might be thinking that this is a fight between billionaires and giant companies and that if it goes wrong, it's their money.
Partly, yes. But it reaches ordinary people, too, mostly through retirement savings. If you have a retirement account, there's a good chance it's in an index fund. That's one of the most common kinds of retirement investment. Instead of picking companies, an index fund buys a little of every big company on a list, usually 500 of the largest companies in America, but it doesn't buy them equally.
It buys each one in proportion to how much it's worth. And because the AI companies have become the most valuable companies in the world, they're now the biggest pieces of your fund automatically. You never chose them. Their size chose them. So here's the one number to keep. On the day of that news report, more than a fifth of a typical index fund sat in just five companies whose businesses are tied to OpenAI.
Nvidia, Microsoft, Amazon, and the chip makers Broadcom and A&MD. Add Oracle and it's more. In other words, if you own an index fund, you already own a piece of this bet. You just didn't place it. That doesn't mean a wipeout. Banks are far stronger than they were before the financial crisis of 2008, and the people who watch over them say so.
But in July, the Bank of England warned that if AI stocks fell, the damage might be amplified by high concentration. That's a polite way of saying that when this much of the market rides on one idea, a fall in that idea is a fall in everything. Which brings us to what happened this week and why it rattled everyone.
First, a word about how these companies describe their size because it explains the whole episode. When an AI company says how much it's making, it usually doesn't give you a year of sales. It takes its sales from the most recent month and multiplies by 12. It's called the run rate. If you had one great month at work and called that your salary, your accountant would have some questions.
But that's the number this industry runs on. So, here's what happened. At the end of September, the news site Axios reported, citing unnamed people, that OpenAI's run rate was nearing $70 billion. OpenAI never confirmed that number on the record, but it didn't correct it either. Not until this week, and in the meantime, the number traveled everywhere.
Then on Thursday, the Financial Times reported that OpenAI had actually told its own investors the number was about 50 billion. Most of the gap turned out to be a difference in counting. When a partner like Microsoft sells OpenAI's AI to its own customers, the bigger number counted the full price those customers paid, the way some of OpenAI's rivals count their sales.
Open AAI only counts the slice that comes back to it. So, no customers had vanished. 50 billion is still more than double where OpenAI was at the start of the year. Nothing about the business had changed between Wednesday and Thursday, but the selling sped up within hours of the story, and Nvidia lost about $170 billion of its value in a day.
Over a $70 billion number OpenAI had never actually confirmed. That's what a bubble feels like from the inside. Prices held up not by what a business is earning, but by what people believe it will earn. That same day, Alman was asked about it on camera. I don't know where these reported leaks or reported numbers come from, but I I am obviously not going to comment either way on that, but I will say like very pleased with the growth.
Not going to comment either way. A few hours later, Bloomberg reported that OpenAI had told investors it expects its run rate to reach $70 billion by the end of the year. So, the 70 billion is back. It's just in the future now. And remember Altman's how about 27? OpenAI's actual sales in the first half of this year came to about 12 billion and its own projection for the whole year as reported by the Financial Times is about $ 36 billion.
Getting to a h 100red billion next year would mean nearly tripling in 12 months. We'll only know for sure when OpenAI publishes its real audited books. It has quietly filed to sell its shares to the public, probably next year. Until then, the most important number in the AI boom is one that nobody outside the company is allowed to check.
Now, plenty of very smart people have put very large sums into Open AI, and they're not fools, so it's worth hearing their case properly. The growth is real. Open AAI's run rate went from about $20 billion at the end of last year to about 50 billion in September. Its sales to businesses more than doubled in a single quarter.
Very few products in history have ever reached 900 million weekly users, let alone this fast. Running chat GPT already pays for itself with money left over. So, the bet is that once the building slows down, the business underneath turns out to be enormously profitable. And there's history on their side, too.
When the internet bubble burst, most of those companies died. But a few like Amazon and Google survived the crash and became some of the most valuable companies ever. The investors are betting that OpenAI is the Amazon of this round. That's a real argument, but there are three things it has to get past.
The money left over after running the computers is shrinking, not growing. It fell from about 40 cents on the dollar the year before last to about 33 cents last year because running the computers cost more than OpenAI planned. Only one user in 18 pays. And the spending isn't just running ahead of the sales, it's running years ahead on promises that other companies have borrowed against and co-signed.
So, let's answer Altman's question. Is this a bubble? Well, by the definition we started with, yes. Prices are running far ahead of the money coming in. A company that this spring spent nearly $3 for every dollar it brought in is valued at more than $800 billion and is talking to investors about more than a trillion.
Part of what looks like demand is being paid for by the sellers. And a single report knocking down one number, a number the company had never confirmed, can wipe $170 billion off the market in a day. But it's not a fake bubble. At that same dinner, Alman asked a second question. Is AI the most important thing to happen in a very long time?
And again, his answer was, my opinion is also yes. Both can be true. The internet was real and the internet bubble was real. The question was never whether the technology mattered. It was who paid for the over excitement. And here's the part that surprised me most. Sam Alman, the man at the center of all this, has said he owns no shares of Open AI directly.
He told the United States Senate in 2023, "I have no equity in Open AI." And no change has been reported since. So when he says someone is going to lose a phenomenal amount of money, he's describing a game in which by his own account he holds almost none of the chips. The ones who do are easier to name.
The builder who borrowed to build largely for one customer. The co-signer who promised to cover the rent. The investors who are also its suppliers paying for some of their own sales. And a little further down the line, everyone whose retirement savings are sitting in the biggest companies on the market, which is a lot of people.
So, if you've got a retirement account, you're already in this story. Whether you meant to be or not, it's worth knowing what you own. Now, I'd like to hear from you. Is OpenAI the Amazon of this boom or the Lucent? Tell me in the comments and tell me why. We read everyone. And if this helped you see the story a little more clearly, a like and a subscribe really do help.
We're a small team and it tells us the work is worth doing. If you want another one, watch our film on Driscolls called In America, a fruit is an Idea about how a company can own a strawberry it never grows. Thanks for watching.
댓글 0