The Boom Bills You Twice

The AI boom already sends one cost down the chain. It just started sending a second, and it lands on the same desk.

A while back I wrote about the AI boom's first bill, in "The Ledger Runs Both Ways." The leading players run extraordinary margins, and those margins are a cost to everyone who buys from them. That cost doesn't vanish. It travels down the chain, marked onto every component, until it reaches a company with no one left to hand it to. E Ink was the clean example. Its shelf-label business, selling to retailers who can absorb a price rise, kept growing. Its consumer e-reader business, selling to shoppers who simply buy less when prices climb, took the hit. Same company, same memory-price shock, two outcomes, split by one thing: who could pass the cost on and who couldn't. That was the ledger. The boom's winnings on one side, and on the other, the firms who quietly pay for them.

Here's what's new. There's a second bill now, and it lands on the same people.

The second bill is the cost of money

The first bill was a component tax, a scarcity markup riding down the supply chain. The second is a capital tax, and it comes from a different place. The thirty-year Treasury just hit its highest level since 2007. Borrowing costs are at multi-decade highs, and not only in the United States. That raises the price of money for everyone who has to borrow.

At first glance this looks like a broad, evenly shared cost, nothing like the targeted markup of the first bill. Rates go up on everybody. But watch where the weight actually settles. It settles on the same axis. The company that couldn't pass a component markup through to its customers also can't pass a rising interest bill through. It is usually the least creditworthy name in the chain, the one with the thinnest margin and the heaviest borrowing, so it feels the rate move hardest of anyone. Two different costs, from two different sources, landing at the same address.

And the second cuts deeper than the first, because the first already thinned the margin the second has to cut into.

The company that is nothing but the two bills

To see both bills on one balance sheet, look at a neocloud, and the clearest one is CoreWeave.

CoreWeave rents out AI computing power. To do that, it buys Nvidia's chips, at Nvidia's prices. That is the first bill in its purest form. Nvidia's roughly 75% gross margin is not an abstraction to CoreWeave. It is the largest line in its cost of goods. CoreWeave pays the leading player's pricing power directly, as the price of being in the business at all.

Then it borrows to pay for those chips. That is the second bill. CoreWeave carried about thirty-five billion dollars of debt at the end of its latest quarter, borrowing that exists almost entirely to cover the cost of Nvidia hardware. Its net interest expense reached roughly six hundred forty million dollars in the quarter, more than double a year earlier, and now eats around a quarter of every revenue dollar. Its gross margin is high, above 70%, and it still posts net losses, because interest and depreciation take what the gross margin earns.

This is the point. CoreWeave has no other business to cushion either bill. A hyperscaler paying more for chips and more for money can lean on a search engine, a cloud franchise, a software cash machine to absorb the hit. CoreWeave has none of that. No cash engine to reallocate, no margin cushion to compress. Its capital structure isn't one input among many, it's the whole company. So both bills land on it undiluted, at the same time, on the same thin margin.

The tell is the engineering

Here's the part that makes it sharper rather than just grim. CoreWeave has actually been driving its cost of capital down, not up. Its landmark financing this year was the first GPU-backed debt to earn an investment-grade rating, priced well below what a company this leveraged would normally pay. It got there by ring-fencing the debt in a bankruptcy-remote subsidiary, backed by its GPU hardware and by Meta's contracted payments, so the rating agencies were really grading Meta's credit and the resale value of the chips, not CoreWeave itself.

Read that as what it is. CoreWeave survives by out-engineering the two bills, quarter after quarter, financing each round of chips on ever more clever terms. And that is exactly where rising rates bite. Every new facility prices off a higher base. The engineering has to work harder each round just to stand still, because the ground under it keeps rising. Meanwhile the collateral beneath the whole structure is fast-depreciating silicon, and analysts already flag that its loan covenants could trip as soon as 2027 if chip values fall faster than the models assume. The escape hatch narrows from two directions at once, the rate underneath and the collateral within.

Same winner, both bills

Now trace both bills back to their source, because they lead to the same door.

The first bill is Nvidia's margin. The second is the cost of capital, and the cost of capital is climbing in part because the AI buildout itself is borrowing on a scale that competes with governments for the same lenders. Nvidia sits at the center of that too. It sells CoreWeave the chips, it is an equity investor in CoreWeave, and its ecosystem's enormous appetite for debt is one of the forces lifting the rates CoreWeave now pays. The same winner is on both invoices. One is the price of its product. The other is a byproduct of how the whole boom it leads is financing itself.

The ledger had one column. This is the second, and it is addressed to the same signer.

The spectrum, and the mispricing

Line the two bills up and you get a spectrum. At one end, a device maker like E Ink feels mostly the first bill, the component tax, because it carries little debt. At the other end, a neocloud like CoreWeave feels both bills at full strength, because its whole model is buying the leading player's product with borrowed money. In between sits everyone else in the chain, paying some mix of the two, depending on how much they buy and how much they owe.

Which brings the mispricing into focus. The market still sorts these companies by the label the boom hands them. Neoclouds get filed as AI winners, riding the demand wave, their backlogs quoted as proof. But the same firm wearing the winner's label is the one signing for both bills, paying the boom's component tax and its capital tax at once, on a margin the first tax already thinned. The backlog is real. So is the double invoice underneath it. The question isn't whether the demand is there. It's whether the company booking the demand keeps any of it after both bills clear.

Who signs for both

The boom's first bill was easy to miss, because it hid inside component prices and landed on companies most people don't watch. The second is easier to see, because it shows up as interest on a balance sheet. What's hard to see is that they are the same story, sent to the same address, and that the firms carrying the AI-winner label are often the ones carrying both.

You don't remove a bottleneck. You relocate it. The cost of this boom relocates the same way, off the leading players and down the chain, and it now relocates twice, to the same desk, on the same margin, from the same source.

So when you read the next neocloud's blockbuster backlog, or the next device maker's guidance cut, ask the question the label is built to stop you from asking. Not who's winning. Who's signing, and for how many of the bills.