The Goose That Could Not Lay a Trillion Dollars
How a trillion-dollar industry started doing Econ 101 in public, and discovered that the math was never on its side.
Somewhere in a glass tower, a banker looked at OpenAI's actual audited numbers, did the arithmetic that everyone had spent three years politely declining to do, and concluded that the most important company of the future could not be sold for a trillion dollars. Then she went and got a coffee. This, more or less, is the whole story.
The rest is detail, and the detail is delicious.
The trillion-dollar limbo bar
OpenAI is reportedly thinking about pushing its IPO to 2027, which is a strange thing for a company to do two months after it was, by all accounts, sprinting toward the public markets with the enthusiasm of a man who has just remembered he left the oven on. Sam Altman pushed past his own CFO's caution. A confidential filing was teased. The leak was pre-leaked. And then everything stopped.
Why? Because the bankers did not think it would clear a trillion dollars on flotation. Sit with that. The pre-money valuation is around 765 billion dollars. Getting from there to a trillion is roughly a 30 percent bump, the kind of move that should be a formality for the self-described future of everything. For a normal hot company, that is a Tuesday. For this one, apparently, it is a bridge too far.
And here is the part that should make you put your drink down. The reporting suggests advisers told the company, in so many words, to find a way to reach a trillion-dollar valuation. Find a way. That is not the language of confidence. That is the language of a group project at 2 a.m. A company that is genuinely worth a trillion dollars does not need to be told to look like one. It simply is one, and the number shows up on its own.
The grim joke underneath is that waiting does not fix this. If your investors bought in at 765 billion, a flat round leaves them needing an exit well north of a trillion just to break even on their hopes, and venture capital has not exactly been showering people with 5x returns for timing things this badly. An exit at exactly a trillion is, in venture terms, a disappointment wearing a party hat. So delaying to 2027 does not buy a better price. It buys eighteen more months of the meter running, during which, as one observer put it, a single stick of RAM may end up costing roughly the GDP of a small nation.
The death spiral has a hardware problem
About that RAM. This is the part that turns a bad quarter into a structural vise, so it is worth getting right.
AI accelerators need a specific, premium kind of memory called high-bandwidth memory, and it competes for the same scarce manufacturing capacity as the ordinary stuff in your phone. High-bandwidth memory hogs that capacity, which means the world can suddenly make far less of everything else. The memory makers, who spent years selling at bargain-basement prices and eating losses, have abruptly discovered they hold all the cards. Prices are up. Long-term supply contracts have been signed in a panic. The memory companies are, for the first time in living memory, swimming in cash.
For everyone building data centers, this is a quiet catastrophe. The exact same data center, the identical amount of compute, now costs 20 to 40 percent more than it did, and it does not do 20 to 40 percent more in return. You are not buying more intelligence. You are buying the same intelligence at a markup, set by suppliers who have you over a barrel and know it.
Now lay that over the one rule these companies cannot break. They cannot slow down growth, because growth is the only thing that lets them afford the compute, and the compute is the only thing that produces the growth. It is a wheel that has to keep spinning faster just to stay upright, and someone just bolted a weight to it. That is what a death spiral looks like before anyone is willing to say the words out loud.
The carousel that only sells tickets to itself
Ask the obvious question. All these data centers, the hundreds of billions of dollars of concrete and silicon, who exactly are they for?
The honest answer is uncomfortable. A frightening share of the demand is the industry buying from itself. OpenAI builds for OpenAI. A cloud provider builds for OpenAI. Another cloud provider builds, once more, for OpenAI. Round and round it goes, each lap generating a press release about unprecedented commitment and not very much in the way of an outside customer happily paying a profitable price.
Build a hundred gigawatts of data center and you have conjured something like 400 billion dollars of annual compute demand that needs to exist. It does not exist. The only buyers spending at that scale are the two big labs, neither of which is profitable, and even together they do not come close to filling it. This is not a conspiracy and it is not a surveillance plot. It is something far more embarrassing, which is boring, bad business executed at a scale that would be funny if it were not wired into the pension funds of half the planet.
No new ideas, only bigger ones
Here is the question that hangs over all of it. With affordability as the central problem of the entire market, why is every frontier lab racing to build the most powerful, most expensive models possible, including ones so costly they cannot even afford to release them publicly?
The deflating answer is that they do not have any other ideas.
These labs are not scrappy startups that clawed their way up under constraint. They are carbon copies of the hyperscalers that birthed them. The infrastructure was built for them. The money was handed to them. They have never once had to run a real business or live inside a budget. Like a giant tech company, the only three things they truly know how to do are hire people, fire people, and spend money. They are extraordinarily good at going very fast in one direction and almost comically bad at adapting, because adaptation is a muscle you only build when someone forces you to.
Put plainly, they are rich kids living off their parents' credit cards. That arrangement works beautifully right up until the moment someone at the table asks, gently, whether the thing could possibly be made cheaper. At which point the rich kid blinks and says he does not know what you mean, because nobody has ever made him find out.
The truly cursed irony is that the one place real efficiency is happening is precisely where the constraints are tightest. China's labs are doing more with less because they have to. Whether you call it efficiency or, as one less charitable observer does, AI dumping - shipping cheap, capable models into the American market to drag everyone into a war of attrition - the effect is the same. Someone described it as financial Afghanistan, the act of luring a richer, prouder rival into spending itself hollow chasing an enemy who is spending almost nothing. It is an unkind phrase. It is also a good one.
The plagiarism machine, plagiarized
There is a deeper trap lurking inside the cheap-model story, and almost nobody is pricing it in.
The popular open-weight models that everyone is suddenly excited about are, to a large degree, distilled from the frontier labs. They learn by feeding on the outputs of the expensive models. There is a certain cosmic justice in watching the plagiarism machine get itself plagiarized, and I would not dream of spoiling that pleasure. But follow the logic. A model trained on another model's outputs can only ever chase the leader. It can close the gap. It cannot get out in front.
So if the frontier labs run out of money and stop training the next big thing, the cheap models do not keep improving on their own. They hit the ceiling and stay there, because the ceiling was someone else's roof. The entire thriving ecosystem of affordable, open alternatives is quietly parasitic on two hosts that are both bleeding money. Kill the hosts and the parasites stop growing too. Nobody wins except whoever is still willing to burn cash training frontier models in private, and that list is getting shorter by the quarter.
A goose, a factory, and the gospel according to Masa
Which brings us, as all things eventually must, to SoftBank.
At the company's 46th annual shareholder meeting, Masayoshi Son delivered a tight one hour and fifty-five minute presentation, at the heart of which sat a goose. Not a metaphorical goose mentioned in passing, but a load-bearing, thesis-defining goose. The true source of value, he explained, is not the golden eggs but the goose that lays them. The goose is also, somehow, a factory. You do not want the eggs, you see. You want the factory that is also a bird.
Translated out of poultry, this is Son pleading with the market to stop valuing his actual assets and instead value the magical premium of his judgment in selecting them. He has tried this before. A decade ago he asked investors to recognize a goose premium, a little extra sprinkled on top for the sheer privilege of his involvement. This is reportedly the same man who, during a rough patch in 2020, reached for a comparison to Jesus on the grounds that Jesus too was misunderstood, and then, sensing he had room to run, invoked the Beatles.
It would all be wonderful, harmless theater if SoftBank were not the largest investor in OpenAI, to the tune of around 64 billion dollars. If OpenAI cannot go public, SoftBank cannot get its money out. The firm reportedly tried to borrow a mere 10 billion dollars against its OpenAI stake, offering up the entire 64 billion as collateral, and the banks declined. They could not even agree on what the stock was worth well enough to lend a sixth of its claimed value against it. When the people whose entire job is assigning numbers to things refuse to assign a number to your headline asset, that is not a vibe. That is a verdict.
The time bombs nobody is defusing
The truly unnerving part is not any single company. It is the wiring underneath.
The third largest company on the Japanese stock market is SoftBank. The second is a megabank that has poured tens of billions into AI data center deals. The largest, fittingly, is a memory and storage company whose valuation has floated upward on the very data center boom in question. The fortunes of an entire national market are now braided together by a bet that may turn out to have been one long, expensive mistake. There is no bailout shaped like this one. Even if someone bought OpenAI for a trillion dollars tomorrow, SoftBank made a dozen other bets that would still need rescuing, and OpenAI would still be unprofitable the morning after.
Meanwhile the signals are getting impossible to misread. Someone at Goldman Sachs recently observed that the first hyperscaler to pull back on capital spending would be rewarded by the markets. That is not analysis. That is an engraved invitation. It only takes one large player to stand up and say it is reevaluating whether any of this has a point, and everyone else will exhale and admit they were thinking the same thing the whole time. Permission is the only thing the herd is waiting for.
What happens when the music stops
The bear case can be early, and being early is functionally the same as being wrong for as long as the party lasts. People have called this reckoning before and watched the line go up anyway, and it is worth saying plainly that the boosters are not hallucinating every bit of value. The tools are useful. Some of the demand is real. The technology is genuinely remarkable in places. None of that was ever the question. The question was always whether the price tag and the revenue would eventually shake hands, and the honest answer is that nobody building this has shown their work.
The most likely ending is not a spectacular explosion but a quiet deflation. Large language models could end up a bit like Oracle, a sturdy and unglamorous licensing business that throws off a couple billion a year as somebody's subsidiary, while the rest of us look back at this era and marvel at how confidently brilliant people lit a fortune on fire. The future, as ever, was real. It was just much further away and much more expensive than the pitch deck promised.
And the genuinely dangerous part, the part worth being angry about rather than amused by, is the simplest. For years the people with the money did not ask the questions. Not because the questions were hard, since profit is revenue minus costs and a child could chair the meeting, but because asking would have spoiled the fun. They are starting to ask now, all at once, at the worst possible moment, with the IPOs at the door and the RAM bill climbing. Whatever comes next, a lot of people who never got a vote are going to find out what happens when the cleverest folks in the room finally do the arithmetic and do not like the answer.
In the meantime, there is one phrase that covers most situations. Until someone explains, in plain numbers, how any of this gets cheaper, the only reasonable response to the next confident announcement is the one the moment deserves. Shut up, Sam.