The question is everywhere now. Is this 1999 again? The charts get lined up and they do line up, roughly — the euphoria, the capital, the valuations that no longer bother to justify themselves. Some reach further back, to 1929 and the recession that followed, to the long record of what happens when a market decides it has found the future and prices it all at once.
The comparison has one thing going for it: a bubble is what people know how to fear. And the dot-com record is real. Companies with essentially no revenue — pets.com went from an IPO to liquidation in under a year on the strength of a sock puppet. The market priced eyeballs instead of earnings, and when the eyeballs failed to become dollars, the unwind was fast.
But there is a number that breaks the analogy, and it is worth sitting with.
The one firm that currently sells the shovels for this entire buildout keeps roughly 70 cents of every dollar it takes in. That is not an exaggeration for effect; it is the margin structure of the company that supplies the chips everyone else is now trying to build for themselves.
Think about what a 70% margin means to a customer. If you are a company that wants intelligence, and you rent it from the one vendor who has it, you are handing over seven dollars for every ten you spend — forever. That is not a supplier. That is a tax. And a tax, unlike a supplier, is something every serious company eventually tries to avoid.
That single number explains more of the current behavior than any narrative about euphoria. It explains why Microsoft is designing its own silicon instead of renting it. It explains why Tesla is building a $16.8 billion chip factory in Texas. It explains why SK Hynix is pouring $38 billion into memory fabs. These are not acts of speculation. They are acts of tax avoidance — and tax avoidance is the most rational behavior in finance.
Now add up what happened in roughly a single week.
Intel filed to sell $15 billion of its own stock, after the shares had risen 400% in a year. Tesla broke ground, in effect, on a $16.8 billion chip plant. SK Hynix approved $38 billion of memory fabs. Celestica — a company most people have never heard of — closed a $3.45 billion equity raise for AI infrastructure. Berkshire Hathaway, in the same stretch, bought $10 billion of Alphabet.
That is more than $73 billion of announced buildout and repositioning in the space of a few days — and that is only the numbers that made headlines. The point is not that $73 billion is a lot. The point is what it all points at: capacity that is functionally identical, being built by people who all expect to be the winner.
When that capacity comes online, roughly together, the price of the thing they built will fall. That is not a prediction; it is the definition of supply. Hundreds of billions of identical chips, fabs, and datacenters will meet a market that only needs so much of them. And the value everyone was trying to capture by building will transfer quietly to the one party nobody is competing to own: the customer.
So the sellers and the buyers are not telling different stories about a bubble. They are doing the same arithmetic and drawing different conclusions about the timing. Intel is selling because it can — $15 billion of stock that cost almost nothing, at 400% markup. Berkshire is buying because it can wait. Neither is wrong. Both are rational. And that is the first clue that the word "bubble" is not the right word.
A bubble is a failure of judgment. People believe things that are not true, pay prices that make no sense, and when the belief breaks, the selling does the rest. The 1999 market was irrational in exactly this way — companies with no earnings, priced on hope.
What is happening now is the opposite failure, and it is a harder one to name. There is a mathematician who spent his life on it. John Nash showed that a group of rational actors, each making the best move available, can settle into a state none of them would have chosen — and then stay there, because leaving alone means losing alone. He called it an equilibrium.
And the equilibrium has a physics. When independent oscillators are coupled together, they synchronize. Each one, acting on its own, adjusts to the others — and the adjusting makes them lock step. The Millennium Bridge in London opened in 2000 and closed two days later: every walker, quite rationally, widened their stance to keep their balance, and the thousand rational adjustments pushed the bridge into a dangerous sway. Nobody walked badly. Everybody walked correctly. The engineers fixed it not by asking people to walk differently, but by adding dampers — something that absorbs the synchronized energy no single walker could see.
The AI buildout is that bridge. Every firm, acting alone, sees a rival building and rationally builds too. The adjustment feeds the swing, the swing feeds the adjustment, and the whole industry locks into the same frequency. And the dampers — the coordination, the specialization, the willingness to let someone else hold the capital risk — are precisely the thing nobody is building.
The distinction matters because the recovery paths are entirely different. A bubble pops — it clears, painfully but quickly. An equilibrium does not pop. It settles. It writes itself down, quarter by quarter, balance sheet by balance sheet. Slower, quieter, and harder to correct — because you cannot correct a rational equilibrium the way you correct a bout of irrationality. A mania burns itself through. An equilibrium just keeps being sensible, which is the thing keeping everyone in it.
Which brings us to the deadline everyone keeps naming but few are actually pricing.
The September IPO wave — the queue of companies whose entire premise is "we built our own stack" — is the first moment the equilibrium stops being an abstraction and starts being a balance sheet. Public markets, asked to value companies built on the same bet everyone else made, will sort. A few will justify the build. Most will not.
Intel selling while Berkshire buys is not a contradiction. It is two answers to the same question. The sellers think the answer is nobody. The buyers think the answer is everyone, eventually. September is the referee.
And the deeper answer is that they might both be right, just not at the same time. The buildout will finish. The compute will arrive. The customer will win. And somewhere between the selling and the buying, a generation of firms will learn what every over-built industry before them learned: that everyone being rational is not the same as everyone being right.
Content is for informational and analytical purposes only — not investment, financial, or legal advice.