The Spread

Two Meters, One Engine

Elon got on the SpaceX earnings call last night and said the thing that should have stopped every AI short cold: old H100s are renting for more than they did when they were new. The three-year-old chip now rents above what the current one fetched at launch. When is the last time the used version of any technology cost more than new? Your phone, your car, all of it starts dying the second you touch it. Nvidia's silicon walks the other way.

Here's why, in his words:

The chip is a fixed engine. The models are the fuel, and every new one is higher octane, wringing more intelligence out of the same watt. So the engine that burns it is worth more every quarter. Even the old one.

Now hold that against the number everyone else is quoting. A single correct answer from a frontier model has fallen from 30 cents to four hundredths of a cent in thirty months. Two meters, opposite directions: the thing compute produces is nearly free, and the compute itself keeps appreciating. The crowd reads the first meter and calls the top. Wrong meter.

The real question was never whether there's demand for AI. It's who owns the engine and how they rent it. Elon owns his outright. He had the capital to buy the hard assets and rent them back short, so every renewal reprices up and he keeps the climb. He is the anti-WeWork. Adam Neumann leased buildings for fifteen years and rented out desks by the month, long liability against short revenue, dead the second the short side blinked. Elon did the opposite: own the asset, rent it short, stay long the option on a rising price.

The neoclouds couldn't. CoreWeave, Oracle and the rest lacked that balance sheet, so to finance the hardware they pledged long-term contracts, selling the option on rising prices to their lenders to get the loan. Now they're stuck selling appreciating compute at yesterday's rate, buried under forward commitments they can only meet if their customers keep growing into them. Let that growth flatten and the equity is crushed under the debt, even with revenue still climbing. That's why CoreWeave's five-year credit-default swaps just blew past 1,000 basis points, pricing in roughly 60% odds it defaults within five years, while the cost of insuring the whole AI-tech complex hit record highs, all in the same week compute is the scarcest it's ever been. The swaps aren't doubting AI. They're doubting a stack of cash-burning labs grows fast enough to carry the debt before the debt carries them.

Give the other side its due: Elon is right on this trade today, not forever. Short duration cuts both ways, and if the cycle rolls over he eats depreciation on chips he owns. He just eats less than anyone, because he can feed his own compute to xAI, Tesla and Starlink when the renters go quiet. The levered middleman has no such floor.

And here's the part the compute debate keeps forgetting. A data center is worthless the instant it loses connectivity, and the hardlines carrying the world's bits have never looked more cuttable. SpaceX owns launch and Starlink. Sever the cable and one system keeps the bits moving. The others own compute. Elon owns compute, the power to run it, and the one pipe nobody can cut. That's not a gas station. That's the whole road.

You don't get rich selling the gasoline. You get rich owning the engine, and keeping the option on what it's worth tomorrow.

The full Signal/Noise — the duration trade, the WeWork inversion, and why Starlink is the asset nobody's counting — is live at getcoai.com.

Outsider Labs Advisory →

0.04¢. That's what a correct answer from a frontier model costs today. Thirty months ago it was 30 cents, a 750x collapse. Intelligence is now the cheapest input in the building. So why did a used H100 just rent for more than a new one did at launch? Because the cheaper and higher-octane the fuel gets, the more the engine that burns it is worth. Don't price the answer. Price the machine that prints it.

For the investor — buy the asset, not the lease. The compute trade isn't "long AI." It's long whoever owns the hard assets and rents them short, short whoever financed them long. Before you touch a neocloud, find the forward commitments and ask one thing: do the counterparties grow fast enough to carry them? If the answer leans on cash-burning labs, the CDS is pricing a risk the equity isn't.

For the business owner — check your vendor's duration, not just its uptime. If your AI stack runs on a provider that bought its chips with debt against long contracts, you're exposed to their balance sheet, not only their SLA. Ask who owns the hardware you depend on and what happens to your service if their credit cracks. The cheapest quote today is sometimes the one that vanishes in a flat quarter.

For the parent — teach them to own the engine. The lesson under tonight's tape is the oldest one going: the people who got rich in the gold rush sold the pickaxes, then owned the mine. Intelligence is becoming free; the durable value sits in the hard, physical things it needs. Point your kid at what can't be copied overnight, not at the task everyone can now do for four hundredths of a cent.

Wednesday, October 21

Where the cycle caught up to us this week.

The refinery was always the asset. We said the winner would be whoever owns the assembled, powered stack and competes on cost-to-serve, not whoever ships the smartest model (The Low-Cost Provider, us, Oct 15), then Elon Musk, on the SpaceX earnings call (Oct 20), said compute-per-watt keeps rising, that he rents short to capture it, and that power is the only thing holding him back. Our thesis, read back off his own P&L five days later.

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: