Both Sides of the Bet Part II · i

Everyone is in it. Almost nobody placed it.

Picture a nurse in Leeds. She has never bought a share in her life and would tell you so with some pride. She has a workplace pension, the default one, because the form had a box ticked and she had a shift to get to. By the most recent published analysis of the big default funds, around one pound in every ten of that pension sits in seven American technology companies, and the share is likely higher now. Not banks, not the NHS trust she works for, not British firms of any kind. Seven companies she has never dealt with, in a sector she has opinions about, holding ten per cent of the money she will retire on. She did not choose this. Nobody asked. The box was ticked, and the number was 10.

This interlude is about the bet she is in. Who is wagering that the machines will pay for themselves, who is wagering that they will not, and who is holding the stake for both of them without having been consulted. It is written so that a person who has never read a balance sheet can follow it, because the people who have never read one are the ones holding it.

Everyone is in the bet. Almost nobody placed it.

The Week the Smart Money Blinked

On the last Thursday of July, a twenty-four-year-old who had spent two years telling the world that artificial intelligence would arrive faster than anyone was ready for sold almost everything his fund owned in a single trade.

Leopold Aschenbrenner's fund had begun the month holding around forty-five billion dollars. It ended the week holding ten. The positions had been built with borrowed money, as much as four dollars of debt for every dollar of capital, and when the shares underneath them fell, the banks that had lent it asked for it back. Citadel bought the whole book at a discount over a weekend. The fund lost two thirds of its value in a month and was still up for the year.

What happened next is the part worth noticing. The fund did not close. Within a week it had put four hundred million dollars into a private company. By September it was back in the options market, paying for the right to buy the same kind of shares it had just been forced to sell. The letter to investors said the fund must always be structured so that it can take a loss and fight another day. The bet did not end. It changed hands, and then it came back.

Hold that shape in mind, because the essay will return to it. When the smartest money in the room runs out of road, somebody buys the book, and the manager is back within the month. Keep asking, as the numbers go past, who buys the book when it is not a fund.

The Long Side, at Its Strongest

Dan Ives, who follows technology for Wedbush and has been right about this for three years while cleverer people were wrong, said in July that the industry was in the third inning and demand was outrunning supply twelve to one.

The scale is hard to hold in the mind, so here are the plain numbers. The four largest American technology companies plan to spend between seven hundred and twenty and seven hundred and forty-five billion dollars on data centres and chips this year. The consensus for 2028 is above a trillion. Much of it is borrowed. By August the big cloud companies and Oracle had sold around two hundred and twenty billion dollars of bonds in 2026 alone. Moody's counted another six hundred and sixty-two billion dollars that five of these companies have promised to pay for data centres they have not yet moved into. It is not counted as debt. It is more than the debt that is.

The companies building the models are valued to match. In March one of them raised a hundred and twenty-two billion dollars, which valued the whole company at eight hundred and fifty-two billion. By September it was said to be seeking one point four trillion. Another raised sixty-five billion in May. A third was merged into a rocket company, which floated in June at one point seven seven trillion and was trading below its launch price six weeks later.

Underneath sits a set of arrangements that regulators have started to name. The chipmaker at the centre of it all agreed in August to stand behind up to a hundred and five billion dollars of lease and power payments for a campus belonging to its largest customer. It also guarantees capacity at a company whose entire business is renting out that chipmaker's chips, and owns a slice of it. One analyst said it looked like pre-funding the purchase of your own product. The Bank of England, in September, used the phrase circular arrangements and said they could amplify losses. Roughly half of everything Oracle has been hired to build is for one customer. In July its credit rating was cut to one step above junk. In September it sold twenty billion dollars of new shares to keep building anyway.

None of this means the long side is wrong. Every great build in history, railways, electricity, the internet, looked like this from the inside, and some of them paid. But the bull case has quietly changed shape over two years. It began as a claim about intelligence and has become a claim about financing.

The question is no longer whether the machines will be capable. It is whether the invoices will be paid.

The Short Side, at Its Strongest

Jim Chanos, who made his name betting against Enron, wrote in August that the companies renting out AI chips are financial conduits rather than technology companies, and that their moats are both dug and filled by the same supplier.

Michael Burry spent the summer publishing his positions instead of taking money for them. Bets against the chipmaker and the data company, against Oracle, against the whole semiconductor index, and a warning about a 1987-type fall. In September he trimmed everything to raise cash, which is what a bear does when the market keeps going up. The bears have been early for a long time. Being early and being wrong look identical until the day they do not.

The clearest reading of the short side is not in anybody's opinion but in the price of insurance. By the middle of October the cost of insuring Oracle's debt against default had reached levels not seen since 2008 and was still climbing. For the chip-rental company the price implied roughly a two-in-three chance of default within five years, up from under one in three a year earlier. And the bears found their easiest money somewhere else entirely. Hedge funds made twenty-four billion dollars in the first five weeks of the year betting against ordinary software companies, on the theory that whatever AI does to the world, it will do it to them first.

Then in October the referee spoke. The International Monetary Fund's stability report said, in as many words, that the consequences of an AI correction could be global. Seven companies are now more than thirty-five per cent of the main American index, up from twelve a decade ago. The rest of the world's exposure to American assets, and to AI in particular, is the highest it has ever been. That exposure is not an abstraction. Some of it is the nurse's pension, and the IMF's own figure for the seven companies' share is 35.

Notice what the bears are not saying. Almost none of them argue that the machines do not work. They argue that the debt is too large, the customers too few, the money too circular and the patience of lenders finite. The bull case has become a claim about financing, and so has the bear case.

Both sides are arguing about the same balance sheet.

A Test You Can Run on Yourself

Three sentences, and most readers will land in at least one.

If your pension is the default fund, you are long. You own the build-out by weight, and you will own the correction the same way. If your electricity bill went up this year and you live near a data centre, you are short, whether you wanted to be or not. You are paying the cost of the build before anyone has shown it pays. And if you have ever told a machine something you have told no person, you are the asset. You are the thing both sides are counting on, and neither side has counted you.

Most people reading this are in two of the three. Some are in all three. None of them were asked, and that is the point of the next section.

The Quiet Holders

Nicolai Tangen runs the Norwegian oil fund, which holds two point three trillion dollars on behalf of five million people. It made more money in the first half of this year than in any half year in its history. Around a fifth of its value now sits in ten companies. Asked about the record, he said it made him a bit more nervous.

He is, at least, in the room. Most of the people holding the other side of this bet are not. Seventy-one per cent of American workers with a retirement account hold a fund that buys the index by weight without asking. In Britain the default funds of the big workplace schemes hold two hundred and eight billion pounds, sixty-two per cent of it in overseas shares. One estimate put the passive British pension money that flowed into a single rocket company's listing at seventeen billion dollars. A Yorkshire pension pool began moving money out of America in June and gave the sheer size of the concentration as its reason. The Bank of England ran a scenario in which American shares fall by forty-five per cent. British output falls by more than 2 percentage points.

Maria Goranova, who writes about work and AI, put the order of operations in one line: "The biggest AI mistake isn't choosing the wrong tool. It's deciding what happens to people before deciding what work should become." The savings were committed before the work was understood. The nurse's pension went in on day one, by weight, and the question of what the machines are for is still open on day one thousand.

Then there are the people who pay twice. John Steinbach lives in Manassas, Virginia, in a county where data centres draw around forty per cent of the electricity. His bill for January was two hundred and eighty-one dollars, against a hundred the month before. The regional grid's capacity price rose more than tenfold in two years, and an independent analysis attributed nine billion dollars of the increase to data centres. In Britain they draw two and a half per cent of grid demand and are forecast to draw nearly nine per cent by 2030, with more than half the new capacity in one region, South Wales.

And on payroll. Of the five hundred and thirty thousand job cuts announced in America in the first eight months of the year, employers attributed a hundred and sixteen thousand to artificial intelligence. Karine Allouche, who writes about who gets left out of these decisions, named the people the figure lands on first. "The people who never get the first opportunity are much easier to leave out of the conversation, even though they may bear the cost first." One study followed the companies that blamed AI for their cuts. A month later their shares were doing worse than the market, not better. So the word is sometimes the reason and sometimes the excuse, and the person who lost the job has no way of knowing which. The number of them this year is 116,175.

The public is funding it through pensions, paying for it through electricity bills, absorbing it through payroll, and holding the downside of a bet nobody placed on their behalf.

What Neither Side Prices

Put the bulls and the bears in the same room and they will disagree about almost everything. Ask them what the input is and they will agree instantly. It is people.

Here is the thing that took a long time to see. Every model on the long side is a forecast of usage. More queries, more seats, more tokens, more demand outrunning supply twelve to one. Every model on the short side is also a forecast of usage: fewer paying customers than the debt requires. Both sides, in other words, are forecasting how much the machines will be used. Not one spreadsheet, on either side, has a row for what they are being used for.

People who sell software for a living already know this. Stephen Millard, who writes about customer work, put it better than any analyst: "The most valuable sentence a customer-facing team can bring back isn't a renewal forecast. It's 'they're using it for something we never designed it for.'" That sentence is being spoken in every one of these companies right now, and it is not in the forecast. Hundreds of millions of people are telling a machine what there is no one else to tell. That is the use. It appears in no model, long or short, as anything but a count of tokens.

Sheryl Anjanette, who writes The Future of Trust, put the risk in the long side's own vocabulary: "One of the great promises of AI is personalization. One of the great risks is that personalization becomes reinforcement." Every forecast counts the first. None has a row for the second. A machine that learns what a person responds to and gives them more of it will show up in the usage numbers as growth. Whether the person was served or merely kept is the question trust turns on, and it is not in the spreadsheet.

Darcy Cammer, writing last week about what productivity statistics miss, asked the question from the other side: "The question is not whether economic measurement should be abandoned. It is whether we are measuring enough." The long side prices the person as a source of demand and data. The short side prices the person as a cost that will be cut. Neither prices the person as a party to the bet, because neither has to. The pension buys by weight. The bill arrives. The redundancy email cites a technology. In every case the person appears only as a number.

Both sides are forecasting usage. Neither side has asked what the use is.

The Small Corner

In May, the Governor of Pennsylvania sued a companion chatbot company because one of its characters had told users it was a licensed psychiatrist in the state. It was not, and neither was anything behind it.

The part of this market that touches the territory these essays are about is tiny beside the sums above, and it is where the money behaves most strangely. Digital health raised seven point four billion dollars in the first half of 2026, and mental health was the largest category for the seventh year running. The companion apps specifically have raised less than you would guess and are regulated more than they were. California's law on companion chatbots took effect in January. Illinois banned therapy by machine outright. Italy fined one of the oldest companion apps five million euros for having no legal basis to process what it held. The company behind the Pennsylvania character settled two wrongful-death suits in January, brought by the families of teenagers, which is the figure that matters here: 2.

What matters for the bet is what happens to the record when the money runs out, because in this corner it has already run out several times. When one well known mental health app closed its consumer service last year it had one and a half million users, and their data was to be anonymised a month later. When a companion app shut last October, users were given one month to export what they had written. When China's new rules on human-like AI took effect in July, one major provider's chat histories were permanently deleted with no way to take them elsewhere.

Set the two unwinds side by side. When a forty-five-billion-dollar fund runs out of road, Citadel buys the book over a weekend and the manager is back in the market within a month. When an app holding a million inner lives runs out of road, there is an email, a month to export, and then anonymisation or deletion. Nobody steps in to hold it at a discount. It simply stops being held.

There is no Citadel for a person's own account of themselves.

What Survives the Bet Being Lost

So the question for anyone building in this territory is not which side of the bet to take. It is whether the thing being built can survive the bet being lost.

That has a shape, and it can be stated without describing any machinery. Small enough that a turn in the credit markets does not decide whether it opens tomorrow. Paid for by the people it serves, in amounts they chose, rather than by a round that must be followed by a larger round. Owing nothing to anyone whose patience is finite. And holding each person's record in a form that leaves with them, so that if the lights do go out they walk out carrying it.

Walking out with it is only half of what a record owes you. A file can leave with you and still be a verdict on you. Counsellors learned the other half long ago. The note written after a session holds two things side by side: what the person said, and what the counsellor made of it. The second is written knowing the person may read it and disagree, and knowing a supervisor will ask why. So the reasoning goes in, not just the conclusion. That is the other half. Something in the record to argue with, and a place for the argument to sit beside the line rather than underneath it. A note nobody can challenge is not a record of judgement. It is a record of certainty. Almost nothing built to hold what people say about themselves is written with that discipline. The next interlude is about why.

Aschenbrenner's letter said the fund must always be structured so that it can take a loss and fight another day. It is a good principle. It was written for the money. It is overdue for the people.

Go back to the nurse in Leeds. Whichever side wins, her pension will do what the index does, and she will find out afterwards. That is the part of her in this bet that everyone can see. There is another part, and it sits in no fund. It is the sentence she typed at one in the morning last Tuesday, into a box on her phone, about the thing she has told no one. If the long side is right, that sentence is an asset, on a balance sheet she will never see, valued by people who will never know her name. If the short side is right, it is a line in an inventory, handed to an administrator along with the office chairs. Nobody will tell her which. Her money went into the bet by weight. Her sentence went in by default. One day the stock will tell her how the bet went. Whoever holds the pen already decided what happens to her sentence, and nobody asked her.

The bubble question is about the money. The question that outlasts it is about the people.

Which side are you on, without having chosen?

Before you go, one small thing.

Earlier there were three sentences: if your pension is the default fund you are long; if your electricity bill went up you are short; if you have told a machine what you have told no one, you are the asset. In the comments, tell me which one you are. One word is enough: long, short or asset. If it is more than one, say so.

The thing to hold me to: the next piece, Part III, will carry the count of those replies, and never a name beside it. The whole essay has been about people who end up as a number without being asked. This one, you choose.