"Our children will enjoy in their homes electrical energy too cheap to meter."
— Lewis Strauss, Chairman of the US Atomic Energy Commission, 1954
“At particular times a great deal of stupid people have a great deal of stupid money."
— Walter Bagehot, 1856
"In economics, things take longer to happen than you think they will, and then they happen faster than you thought they could."
— Rudi Dornbusch
Everyone is arguing about whether AI is a bubble. I think that’s the wrong question. A bubble is an argument about valuation, and you can be right about valuation for years and never get paid.
The question that matters is narrower.
Intelligence now has two prices. The price of a unit of thinking is collapsing, and the price of the machine that produces it is climbing, and both are happening for the same reason. Somebody has to sit between those two prices. Who that is, and what happens to them when the gap closes, decides what you’ll pay for intelligence, which companies win, and what your mortgage rate and your retirement account do over the next two years.
Thinking Is Getting Cheap
Somewhere there’s a rack the size of a fridge with 72 of Nvidia’s newest chips inside. You can rent one.
You download Kimi, a model built in Beijing and given away free, and the free software that serves it. You list yourself on OpenRouter, where developers buy thinking by the token. You wait. Dylan Patel at SemiAnalysis, who counts these racks for a living, describes what happens next like a man tired of explaining it. You’ll make more from the tokens than you pay for the compute. Anyone can do it. It’s not that hard.
I keep coming back to that rack because it’s the floor of a market, and the floor is where you find out what something is worth. A megawatt of compute rents for $10-15m a year at the cheap end, and at that rate a stranger with no research team and no brand can sell intelligence to the public at a profit. Every time a Chinese lab uploads new weights, the floor drops. And the floor is where the volume is. Fireworks, one company serving open models, handles 40 trillion tokens a day, twice what OpenAI’s API handled in March.
The price is falling to meet it. Citadel Securities keeps an index of what a token costs, weighted by what people actually use, and between late June and mid-August it fell 40%. Sean Maher at Entext, who has done this arithmetic more carefully than anyone I’ve read, thinks another 60% is the base case, not the bear case.
Thinking is getting cheap, fast, and it isn’t going back up. That’s the first thing to know about where AI is going, and it’s good news.
The Two Prices Of Intelligence
The second thing is that the machines that make it are getting dearer, for the same reason.
A year ago, serving the frontier model lost money on every token. Today Anthropic’s revenue has run as high as $50m a year per megawatt, against a rental cost of ten or fifteen. Token prices fell the whole way through that. The useful work you get out of a watt rose faster than the price fell, and each new model earned more than the last.
A margin like that lets two companies outbid the world for the machine. SpaceX built more compute than it needed, sold the spare to Anthropic and Google for $25-40m a megawatt. Patel thinks the labs will keep paying like that, 25, 30, 50 million a megawatt, to take the share of the world’s compute they’ve already signed for.
So intelligence has two prices, going opposite ways.
The token, what you and I pay for a unit of thinking, is getting cheaper. The megawatt, what a lab pays for the capacity that makes tokens, is getting dearer. Each causes the other. The labs earn so much per watt that they bid up the factory; the free models run on cheap factories and bid down the product. Today the labs are winning that arrangement by a distance. They hold the blades.
Trotsky had a chart for this. In 1923 he stood before the party congress with industrial prices going up and farm prices going down, two lines that had crossed and opened like a pair of blades. He called it the scissors. The peasant sold grain into a falling market and bought nails from a rising one, and the space between the blades was how hard he was being squeezed. Nobody in that hall needed telling that a gap like that closes in the end, one way or the other.
Revenue = tokens per megawatt x price of token
The scissors close on whoever is paying the frontier’s price for compute without earning the frontier’s value from it. And that isn’t the frontier.
I spent 27 years in institutional finance, and the thing I learnt to look for first, before the strategy, before the people, is the mismatch. It’s the one sin finance commits every cycle. You own an asset that behaves one way and owe a liability that behaves another.
The subprime mortgage of 2006 was a two-year teaser rate and then a reset to a payment 30-50% higher, sold to people who couldn’t afford the reset and didn’t plan to be there when it came; the house would be worth more by then and they’d refinance. WeWork took fifteen-year leases on buildings and let the space out by the month. A long fixed liability against a short floating asset. Everything else about that company was a distraction.
The frontier labs have the WeWork structure. Their compute comes on take-or-pay contracts, signed at today’s price for capacity that won’t switch on for two or three years, because a campus has to be sited, powered and built first. Once it’s on, they pay whether or not anyone’s buying. Nothing bills until delivery, and that gap is the teaser. Their revenue, meanwhile, is a spot market in tokens whose floor is set by that rack. One blade is a contract. The other is a price.
There’s one difference from WeWork, though, and it’s a BIG one. WeWork’s asset didn’t compound. A frontier lab’s does.
Revenue per megawatt is tokens per megawatt times price per token.
Maher’s base case has the price down 60%. Patel has tokens per megawatt, the efficiency of models and machines together, rising about 3x a year. Multiply them and revenue per megawatt rises about 20% while the price of the product collapses. That’s how token prices can fall and compute prices rise at the same time.
But it only works at the leading edge. The rack running free weights and the Anthropic megawatt are the same hardware. One earns $13m a year, the other $50m, and nothing about the chips explains that; the model does. The labs aren’t paying an arbitrary price for compute. They’re paying what it’s worth to whoever gets the most out of a watt, and right now that’s them.
So the question that matters is less whether the labs can afford the contracts than whether the frontier stays the frontier. That’s a race between a clock and a gap.
The Clock & The Gap
The clock is diffusion. SemiAnalysis measured how long open models take to catch the first closed model of each era, and it halves every time. Thirteen months in the early scaling era, eight and a half in the reasoning era, five in the agentic one, and they expect under three for whatever’s next.
The gap is compute. The two labs took about a quarter of the world’s new compute this year, will take close to half next year on contracts they’ve already signed, and on Patel’s forecast are running most of the usable compute on the planet by 2028, with models that increasingly help build the next models.
Diffusion gets faster every era. Concentration gets steeper every year. You have to decide which wins. Patel’s bet is the gap. Neil Movva, who runs Sail Research, a company whose entire purpose is making tokens cheaper, has bet the company on the clock, that nothing stays secret, and being three months ahead is worth a little less each time. There’s a ceiling even on winning. Damodaran points out that flat token prices across the two labs’ contracted capacity would mean $1.1trn of revenue a year between them, nine cents of every dollar of American wages. The frontier doesn’t need flat prices, though. It only needs to stay ahead.
Then there are the bills.
In March 2007 Credit Suisse published a chart of a trillion dollars of adjustable-rate mortgages contractually due to reset over the next two years. Every desk in New York and London saw it. Every loan on it was still performing, because every borrower was still inside the teaser, and the market kept buying because it believed the exit would come before the reset.
Compute has the same wall. The signing boom of 2025-26 guarantees a billing boom in 2027-28, the way the 2005-06 mortgages guaranteed the 2007-08 resets, and for the same reason, which is that the delay was written into the instrument. More than $2.3trn of compute contracts sit on the books of the four biggest clouds, signed and not yet billing. Oracle’s grew 360% in one fiscal year. Groundbreaker, a Substack that built the wall contract by contract in a piece called The Teaser Period, has about $700bn switching from signed to billing in 2027 and $800bn in 2028.
Rating agencies have counted take-or-pay obligations as debt for thirty years, in pipelines, shipping and power. Nobody has done it for compute. Do it and the AI complex carries about $2.1trn of obligations against $470bn of reported debt, more than the entire subprime market at its peak.
I wrote in The Last Ship about the gap between when something breaks and when we feel it. This is that gap with the sign flipped. A teaser period doesn’t just hide the reset wall, it manufactures the evidence people use to dismiss it. Record backlog, sold-out capacity, vendors beating estimates, all measured on the way up, all necessarily true while what’s been delivered lags what’s been signed. The backlog can only grow, so it tells you nothing.
Watch conversion instead, the share of contracted capacity actually plugged in and earning. CoreWeave is at 36%. Nebius is under 20.
So whose mailbox do the bills land in?
Who’s Paying The Bills?
Not the owners of compute, or not yet. Hold an energised megawatt today and you hold something scarce and perishable that you sell to the highest bidder, and the highest bidder is the frontier.
A new gigawatt costs about $42bn to build and needs $18.5bn a year to make 15%, and contracts are being signed at 20-25. A lab earning frontier value per megawatt can fund that. A lab one step behind, paying today’s compute price for yesterday’s value per watt, can’t. The credit market has worked this out already. CoreWeave borrowed twice this year against the same chips, 225 basis points over the benchmark in March, when the customer behind the contract was investment grade, and 550 over in August, when it wasn’t. Three hundred basis points, five months apart, and the only thing that changed was who pays.
Groundbrkr ran the wall lab by lab, which puts the argument in two names. On its own plan, OpenAI’s committed compute is more than 200% of revenue at the 2027 peak, and its plan for the reset is the 2/28 borrower’s plan, raise the next round before the bill comes. Anthropic’s commitments peak near 60% of revenue and fall from there.
When a bill like that can’t be paid, I doubt it looks like a default. It’ll be called partnership, with capacity rephasing, efficiency-linked pricing and a deferral that comes with a press release. But the moment one anchor contract is amended, “contracted” stops meaning certain anywhere in the system.
Meanwhile, watch what the labs are doing.
They’re cutting the metered price of a token, raising the limits on flat-rate plans, and selling to enterprises by the seat rather than the unit. A subscription is a fixed price for a floating amount of thinking. A take-or-pay contract is a fixed cost. Sell enough fixed prices against a fixed cost and the mismatch closes from the revenue side. That’s where AI is going for you as a customer. You’ll buy intelligence the way you buy a phone plan, flat rate with a ceiling, and the meter will be something only the free models bother with.
They Have Found Things To Do With It
Maher’s nephew bought a lab-grown diamond and spent the savings on the honeymoon.
He’s not unusual. Lab-grown is now 60% of engagement rings. More carats are sold than ever; the dollar value of the diamond market has fallen; De Beers lost control of the price; and the retailers, who buy cheap stones and set them, make 70-80% gross margins. The stone got cheap. The setting is where the money went.
Cheap thinking works the same way. Cheaper tokens do make people think more; the machines are filling faster than prices are falling, and that elasticity is the bull case.
But look at who’s doing the extra thinking. Citadel has card data from 70,000 firms. In July, with token prices down about 40%, spending on AI per employee still rose 49% in a month at the top 1% of companies, which means they were consuming roughly two and a half times as much. At the median firm, spending rose 9%. Since late 2023 the top 1% has added about $6,500 a month per employee. The median firm has added $9.63. The gap between the 90th percentile and the middle has doubled, from 27 times to 54.
Whether the top firms earn more from all that intelligence, the card data can’t say. What it says is that they’ve found things to do with it, and most firms haven’t.
Cheap thinking goes to whoever already knows what to do with it. The model is the stone. The margin sits with whoever owns the setting, which means the workflow, the customer, the taste. The top 1% of firms have worked that out. The median firm has spent $9.63 finding out.
It Is Your Problem Too
The scissors close on everyone’s cost of capital, which is why this is your problem too.
About half of this decade’s AI buildout, roughly $5trn on Patel’s numbers, gets borrowed. And the borrowers don’t care what it costs. A megawatt that rents for $13m earns the frontier $50m; SpaceX got its outlay back in a year; a trading firm running these models can pull $300-500m a year out of a single megawatt. Meta borrowed at 5 or 6% and would cheerfully pay 8, because on a $50bn gigawatt two extra points is a billion dollars a year, and the compute earns that back in weeks. But if Meta pays 250 basis points more, so does everyone else, and banks reprice what they owe faster than what they own. Johnson & Johnson, which earns steady money and can’t pay 8% for anything, gets repriced without doing a thing. Countries with heavy debt and short maturities get repriced faster; the economist Basil Halperin calls it a second Volcker shock, and forty countries defaulted in the first one.
Your mortgage rate and your pension’s discount rate are counterparties to this trade whether you chose to be or not. So is the pension itself. Companies exposed to AI are about 45% of the S&P 500, and the ten biggest names are about 40% of it, against 27% at the top of the dot-com boom. A passive retirement account is a levered, concentrated bet that booked compute becomes billed compute and that two cash-burning labs pay for it.
Nobody chose that allocation, and almost nobody holding it knows.
And that is Patel’s own bear case on the labs. He doesn’t think they run out of money or ideas. He thinks they run out of permission. The limiter, in his words, is how much the rest of the world lets it happen, and everyone who’s elected is going to hate AI. New York is banning data centres, Texas has paused interconnections, and the best model in the world was trained in February and still hasn’t shipped. The bills heading for your mortgage are how that permission gets withdrawn.
Which leaves the question the next two years will answer. Cheap thinking is coming. Does it reach the median firm and the median household, or does the surplus stay where the July data says it’s going, inside two companies and the top percentile of their customers, reinvested in the next model before it gets near a payslip? Patel expects the second. Movva, building a token factory from scavenged megawatts and chips nobody else wanted, has bet on the first.
The rack is still out there, earning its margin on free weights. Above it, two prices are pulling apart, one written into contracts and one written nowhere. Next time someone quotes you Nvidia’s market cap, ask them two things instead.
What does a token cost this month, and how much of the contracted capacity is switched on. Most people with the market cap to hand can’t answer either, and those two numbers will tell you more about the next two years than the one they’ve got.
If you remember five things:
Thinking is getting cheap, fast, and it isn’t going back up. Token prices fell 40% this summer and the base case is another 60%. That’s good news.
The machines that make it are getting dearer, for the same reason. The frontier earns $50m a year from a megawatt that rents for $13m, so it outbids the world for capacity. The two prices are pulling apart because each causes the other.
The squeeze lands on whoever pays the frontier’s price without earning the frontier’s value. The labs bought fixed, multi-year compute against floating token revenue, the WeWork structure, and the bills start arriving in 2027-28 on a schedule fixed at signing. Watch conversion, not backlog.
Cheap intelligence goes to whoever already knows what to do with it. The top 1% of firms are consuming two and a half times more; the median firm has added $9.63 a month. The stone got cheap. The setting is where the money went.
Everyone pays through the cost of capital. Borrowers who don’t care about the rate set it for everyone who does, and 45% of the S&P 500 is now a bet that booked compute becomes billed compute. The frontier’s real risk is permission, not money.
If this was useful, send it to the one person you argue with about AI. And hit reply and tell me where I’ve got it wrong. The replies are where the next essay starts.
If you allocate capital seriously and want the version of this work with positions in it, that’s the Curious Mind Research Letter.
Sources. Neil Movva, Sail Research, on Invest Like the Best (August 2026). Dylan Patel, SemiAnalysis, on the Dwarkesh Podcast (August 2026). Sean Maher, Entext Market Insight, 24 August 2026. Frank Flight, Citadel Securities, Elastic Expectations, 18 August 2026. SemiAnalysis, Are Open Models Catching Up?, 21 August 2026. Groundbrkr, The Teaser Period, August 2026.
Nothing here is investment advice. I may hold positions in companies mentioned.
My view remains that the cheap statistical correlations enabled by LLMs, though useful, is not thinking and cannot do what thinking does. So these models will actually prove able to do far less than the hype has claimed.