The Two-Story Bubble
Here is a magic trick. Watch closely, because it’s being performed on you every earnings season.
Nvidia takes an equity stake in an AI lab. The lab uses that money — plus a compute contract — to rent data centers. The cloud provider running those data centers uses the proceeds to buy chips. From Nvidia. The same dollar has now been counted as revenue, or investment, or backlog, at three different companies. It looks like three healthy businesses. It’s one dollar, walking in a circle, wearing three different coats.
That’s the whole trick. And the entire trillion-dollar debate — is AI a bubble? — is being fought on the wrong question, because everyone is arguing about whether the revenue is real and nobody is asking the two questions that actually decide whether you lose money: whose dollar is it, and how fast does the thing it bought fall apart.
This letter is an investigation into those two questions. The first half — the part you’re reading free — hands you the tools to run the investigation yourself. Because once you can see the trick, you can’t unsee it, and you’ll never read an AI “run-rate” headline the same way again.
Movement 1 — Follow the dollar
Start with the loop, because the loop is the tell.
The bull case is a statement about a single number: AI revenue is real, and it’s growing faster than the internet did. Both halves of that sentence are true. What the sentence hides is where the revenue comes from.
Roughly $800 billion of the current AI economy sits inside interlocking arrangements — a vendor taking a stake in its own customer, a chipmaker backstopping the cloud that buys its chips, a lab committing to spend at the very company that just funded it. The money moves in a circle among a small group of firms, and each pass around the circle books as growth somewhere. From the outside, manufactured demand and organic demand look identical. That’s the point of the trick — they’re supposed to look identical.
So here is the tool. One question, and it’s free, and it’s the most useful thing in this letter:
Would this revenue exist if the seller had not first funded the buyer?
Run it on any AI revenue figure you’re handed. If a cloud provider’s growth traces back to a lab that a chipmaker capitalized, and that lab’s compute bill is the cloud’s revenue, and that cloud’s chip order is the chipmaker’s revenue — you are not looking at three data points. You are looking at one dollar, three times. Strip out the circular portion and the underlying business is meaningfully smaller than the headline.
The test doesn’t tell you the revenue is fake. Some of it is entirely real — people pay for these products because they’re useful, full stop. The test tells you how much of the number you’re allowed to trust, which is the only thing that matters when you’re deciding what to pay for it.
That’s Movement 1: the numerator — the revenue everyone fights about — is partly a hall of mirrors. But even the real part has a second problem, and it’s the one nobody models.
Movement 2 — Find the hidden clock
Every dollar of AI infrastructure is a countdown. The moment a GPU is installed, it begins losing value — not slowly, like a building, but fast, like a phone, because a better chip is already on the way and the better chip makes the old one worth less.
This is the number nobody watches, and I want to show you exactly how much is hiding in it. Not with a table of names — with one company, dissected, because one honest dissection teaches more than four tidy summaries.
The Oracle problem
Oracle is the cleanest example in the market of what happens when you own the countdown and pretend you don’t.
Oracle levered itself, hard, to build data centers for essentially one tenant. By the end of its FY2026, its free cash flow had gone negative — around minus $24 billion — against roughly $130 billion of outstanding debt, before you even count the hundreds of billions in lease commitments that are signed but haven’t started. Its remaining performance obligation — the backlog everyone points to as proof of demand — swung from about $138 billion to over $500 billion in a single year. That backlog is the bull case. It is enormous, and it is real, and it points, overwhelmingly, at one customer.
Now watch the clock. Here is the accounting sleight-of-hand that the entire industry is running, Oracle included: the chips in those data centers are being depreciated over five to six years. The frontier models that make those chips valuable turn over every twelve to fifteen months. You are amortizing a fifteen-month asset across a six-year schedule, which means the profit you report today is borrowing from a write-down you haven’t taken yet.
How much? If the big AI infrastructure players moved to an honest three-year depreciation schedule, it would erase something on the order of $780 billion of their combined enterprise value. That is not a rounding error. That is three-quarters of a trillion dollars sitting in the single line item that nobody argues about.
So stack Oracle’s two triggers. If its one tenant stumbles — misses a payment, slows its spend, renegotiates — the revenue doesn’t just soften. It softens while the capex is still there, still on the books, still decaying on a clock that was set too slow. Both sides move against you at once, and the drawdown is always worse than your model said, because your model flexed the revenue and quietly held the depreciation constant.
That’s the two-story structure. Story one is the numerator — the revenue, partly circular. Story two is the denominator — the capex, dying faster than the books admit. A disaster is what happens when the market prices the first story and ignores the second, and then the second comes due.
Movement 3 — The bet at the end of the chain
Here’s what makes the whole edifice fragile in a way no balance sheet shows: all of it rests on a bet about technology that isn’t yours to control.
The buildout only pays off if the models keep getting dramatically better — good enough that demand eventually grows into $700 billion of annual spending. Pull that assumption and every number above collapses, because circular revenue can’t loop forever and the depreciation clock never stops.
And the assumption is under strain. Inside the labs, a quiet consensus has formed that the raw scaling that drove the exponential years has hit diminishing returns. The arithmetic is stark: in 2021, doubling the compute you threw at pre-training roughly doubled measurable capability. By 2025, doubling it might buy you ten to twenty percent on the hardest tasks — and less above a threshold. Even the field’s own founders have said the pre-training era, as we knew it, is ending.
Call it the desert. If capability plateaus, revenue can’t grow into the buildout, the loop can’t manufacture demand indefinitely, and the clock does the rest. The infrastructure doesn’t need a crash to fail. It just needs the models to stay roughly as good as they are now while the depreciation keeps running.
The only thing that crosses the desert is a genuine breakthrough — a new axis of improvement beyond brute scale. There are candidates: test-time compute, sparse architectures, synthetic data. That’s the real bet buried inside every AI valuation on the tape. Not “is AI useful” — it plainly is. The bet is narrower and sharper: does a new scaling axis arrive before the current capex finishes depreciating? Almost nobody is pricing it as the binary it is.
And now the part that should worry you — including about this thesis
I’ve just handed you a machine for seeing the bear case as arithmetic. So let me be honest about what I haven’t handed you, because the gap is where people lose money — and this time it cuts against the bears.
The clock is starting to get answered. For every dollar of AI infrastructure that depreciates, roughly $1.19 of hyperscaler and neocloud revenue is now coming in to cover it — up from below 1.0 a year ago, the first time the revenue side has pulled ahead of the capex curve. JPMorgan’s midyear read is that the hyperscalers are profitable, the debt markets are holding, and the cycle has room to run. That is the denominator getting answered, at least at the top of the stack, and it is not nothing.
And here is the distinction that kills lazy versions of both cases: four profitable giants spending $725 billion while running highly cash-generative core businesses is a completely different animal from a neocloud taking on billions in term loans against chips that obsolete before the loan amortizes. Conflating them into one scary number is its own error. The loop is dangerous in some names and survivable in others, and telling them apart is the entire job.
“AI is a bubble” is not a thesis. It’s a mood. The most prominent bears have drifted from economics toward accusation precisely as the economic case got harder to make — the costs-aren’t-falling argument stopped working because costs are falling, and the nobody-uses-it argument stopped working because people do. A screen that shows “huge capex, thin revenue” is worthless on its own, because it can’t tell you which side of that distinction a given name sits on.
That distinction — which names are load-bearing and which are lethal — is the whole game. The free half gave you the two questions and the tools to ask them: follow the dollar, find the clock, name the bet. What it can’t give you is the scorecard.
🔒 The Premium Edition — Macro Notes
The free half gave you the investigation. Here is the case file.
The Circularity Ledger. Every major AI revenue line scored on the one question — real, or funded by the seller. The roughly $800 billion loop, mapped node by node, so you can see which “run-rates” survive the test and which evaporate when you strip the circular dollars out.
The Desert Clock — a downloadable model. Enter a name’s capex, its honest depreciation life, and its genuine third-party revenue, and it returns one number: how many quarters of a capability plateau that name can survive before the arithmetic turns against it. The four traps already handled — circular revenue stripped, depreciation normalized off the stated schedule, one-time equity gains removed, off-balance-sheet leases pulled back on.
The Oracle dissection in full. The balance sheet walked line by line — negative free cash flow, single-tenant concentration, the backlog that’s really one signature, the depreciation schedule that flatters today’s earnings. The exact reasoning trap, and the position-sizing rule that follows when both stories can reverse at once.
Load-bearing vs. lethal. The scorecard. Which names have real revenue and a survivable clock — and the three that look safe and are the fragile node in disguise. The neoclouds are the free preview of that logic. The other names are not free.
The lifetime offer — this week only
Macro Notes is a Substack Bestseller.
The free editions are why people subscribe. The paid editions are why they stay.
The annual subscription is $300/year.
This week, and only this week, $500 once buys lifetime access.
The arithmetic isn’t subtle: it pays for itself in twenty months. But the discount isn’t the real reason. This method doesn’t resolve on a one-year horizon. The whole thesis of the Two-Story Bubble is that the second story arrives on a lag — depreciation is slow until it isn’t, and the desert shows up quarters after the last triumphant press release. A framework built on watching the delayed number can’t honestly be sold as an annual product.
Lifetime access is the version consistent with its own time horizon. You’re not buying a year of calls. You’re joining a portfolio built in public — one dollar traced, one clock read at a time, with every position’s reasoning published before it works, not after.