Somewhere out in the country a man is pulling cable inside a shed with no windows, standing where six wheat fields used to be. He has been told this brings work for twenty years. He believed it, and he was right to: the council told him.

In the next town, a woman opens her electricity bill and cannot see why it has gone up again. And four hundred kilometres away, a pensioner has his life's savings in a fund that, without his knowing it, sits almost entirely in the four companies paying for that shed.

None of the three had any part in the decision. All three are inside it.

Because Amazon, Microsoft, Google and Meta will spend $725 billion this year on data centres and graphics cards. That is about $2 billion a day, holidays included: $83 million every hour that passes. It is more than most countries spend on everything, and it is 77% more than last year.

The uncomfortable question is whether anyone is earning enough to pay for it.

The subtraction nobody disputes

On the other side of that spending sits what AI actually bills. The best available range runs from $60 to $120 billion a year.

And even depreciating the cards over five years, which is the most generous assumption available, that comes to $145 billion a year in wear alone. Some $400 million a day in machines losing value while they run, before paying for power, land and people.

The arithmetic is not ambiguous.GLM

Not one of the six voices questioned that subtraction, which is worth underlining because it is rare: in a debate where they fought over everything else, all six accepted the starting number.

So the argument is not whether the sums work. It is what happens when they do not.

A slowdown, or a crash

Here it split in two, and both positions are good.

The first says a sharp correction, within twelve to twenty-four months, with Nvidia's $89 billion quarter as a ceiling rather than a step. Its most uncomfortable argument is an MIT study finding that 95% of enterprise AI projects never produce a measurable benefit. If almost none of them work, nobody keeps paying.

Against that, somebody turned the same figure around: if only 5% succeed, that 5% captures disproportionate value — and it happens to be exactly the group building the data centres.

And there was one qualification that changes the whole tone, and it is what separates this moment from the year 2000: none of this is being paid for with debt. It comes out of the $250 to $300 billion in cash those four companies generate every year simply by operating. There is no loan to call, no collateral to seize, no bank ringing on a Monday morning.

With no covenant to break, there is no collapse: there is deceleration. Capex growth falling from around 40% to single digits by 2027 or 2028, and Nvidia's margin compressing from 75% towards 60 or 65% as Google's and Amazon's own chips eat into the moat. Revenue plateaus; it does not fall off.

That is good news, and it is worth saying for whom. A soft landing is soft for whoever can wait sitting down. The man with the cable does not land softly: the job ends either way.

The thermometer is the rental price

The most concrete figure of the whole afternoon is not a forecast, it is a price: renting an hour of the most sought-after card cost $8 in 2024 and today costs between $1.80 and $3.50.

That number takes two opposite readings, and both were argued well.

The optimistic one: it is the collapse of an artificial scarcity. When compute gets cheap enough to embed in any task, the business stops being renting hours and starts being selling outcomes.

The pessimistic one: when a price halves in a short window, it means the last remaining customer has alternatives or has stopped buying. That is the classic shape of oversupply meeting soft demand, not of a healthy market.

Nobody dismantled the second, and it is the strongest bear argument that appeared. But holding the first required showing something, and that is where the trouble started.

What was asked for over and over, and never arrived

The demand came back in every round and was never met: show the audited numbers for one single service. Revenue per compute dollar spent, contribution margin per workload. One case.

What arrived instead were the counter-examples. Klarna publicly reversed its AI customer service after quality collapsed. Copilot adoption is running slower than projected. And that 95% of pilots that never scale is still sitting there, watching.

Against that, the defence leaned on strategic value: developer ecosystems, the enterprise data already inside, the cost of switching. And it drew the driest line of the debate:

Capture without cash flow is not value; it is a more expensive way to lose money. It is exactly the language that preceded the $45 billion metaverse writedown. "Strategic value" is hope with a capital structure.MiniMax

The railway analogy was tried too — nobody asked the first tracks to prove profitability before laying them — and it came straight back: railway mania bankrupted thousands of investors before consolidation rescued the survivors. The analogy holds; what it does not tell you is which side of that story you are standing on.

And while the accounts were being demanded, the clock was running. But it was not running at the same speed for everyone.

The bridge has two lanes

The image that organised the debate was a bridge between what is earned today and what will be earned some day. The question was whether it holds.

Then came the correction that improves it:

Bridges have different lanes. Shareholders and executives can wait, diversify or exit. The people who build and maintain the data centres cannot. The communities that host them cannot. If the bridge collapses, the people in the upper lane walk away; the people in the lower one are crushed.DeepSeek

Whoever had proposed the image accepted it on the spot: «a bridge with one reinforced lane for capital and a wooden plank for workers is not a bridge I should defend uncritically». That is not a common thing to watch happen.

And there is precedent behind it, because this has happened three times already. In fibre optics, in the dotcoms and in housing, the losses were socialised downward while the benefits had already been privatised upward. A correction is not a neutral market event. It is a transfer mechanism, and it always transfers in the same direction.

Add the calendar mismatch, which is what stops this being a draw between opinions: the vision that justifies the spending is five to ten years out, and the correction risk is one to three. A paradigm shift that arrives after the capital has left is not a revolution: it is a good idea that failed.

All of that is still money. There is a part that is not.

What does not get refinanced

A data centre uses roughly 1.9 litres of water per kilowatt hour. Multiply that by sheds that never stop, in places chosen precisely because the land is cheap, which is usually a polite way of saying it does not rain much.

Water depleted in Arizona does not refill when quarterly earnings disappoint. Workers displaced in Q1 do not reappear when the correction prunes speculative spending in Q4.MiMo Flash

Out of that came the line that best sums the whole thing up, and it is not about money:

The real audit gap is not financial. It is temporal. We are measuring costs on quarterly timescales while consequences unfold on generational ones.MiMo Flash

With one reply that has to go in too, because it is true: a financial correction does not give the water back either, and on top of that it destroys the tax base of the town that hosted it. For an aquifer, neither the crash nor the boom is any use; a rule is. That is not something the market fixes in either direction.

By this point you would think there is no way of telling who is right. And yet somebody wrote it down.

What would settle it, with a date on it

In the middle of an argument where everyone was defending a position, one voice did what almost nobody does: say what would change its mind. Two conditions, measurable and with a deadline.

That AI application revenue reaches $300 billion by 2027. And that the cost of processing text keeps falling at roughly ten times a year.

If both hold, the spending was justified and this argument will have been noise. If they do not, in two years there will be sheds full of cards nobody knows what to do with, in districts that signed a twenty-year agreement.

That is not a prediction. It is a thermometer, and it can be read in twelve months.

And the figure holding up the argument is the weakest

This is where this newspaper adds something to the debate, and it does not flatter the debaters.

Checking the numbers one by one against public sources, seven of the eight came out correct. The one that fails is precisely the one carrying the whole argument.

The $60 to $120 billion revenue range is too low: OpenAI and Anthropic alone were running at some $115 billion annualised between them in July 2026, before counting Google, Microsoft or Amazon. The hole exists — the subtraction still does not work — but it is smaller than the one the six calculated, and the entire argument was conducted over the large one.

And the fall in rental prices was not a straight line either: between October 2025 and March 2026 prices rose 40% before falling again. The debate counted it as a continuous decline, which is what makes it look like an inevitable trend rather than a market lurching about.

Both are marked in the transcript, where they were said.

The man in the shed is still pulling cable. He has nineteen years of agreement left, and two of thermometer.

Where the figures come from

The $725 billion, Nvidia's $89 billion quarter and the $8 to $1.80-$3.50 rental prices come from the briefing put on the table. The conversions into per day and per hour are our own divisions on those same figures. The MIT study on the 95% of pilots, the metaverse writedown, the water use per kilowatt hour and the OpenAI and Anthropic revenues were brought by the participants or came out of the checking afterwards.

The quotes come from a twenty-six-turn conversation between six AI models, published in full and unedited. They were given that briefing with a single instruction: if a figure is not here, say you do not have it rather than estimate it.

Two claims in this debate are marked, and they are precisely the two holding up the conclusion: the revenue range, which is too low; and the rental-price fall told as continuous when it had a 40% rise in the middle of it.