By the spring of 2026, insuring a data center had become a strange kind of routine. A broker who does this for a living described the shift plainly: putting ten to twenty billion dollars of value in a single location creates capacity problems, because only so many insurers will stand behind that much on one patch of ground. In 2023, reasonably insuring a twenty-billion-dollar campus was nearly impossible. By 2026 it was a weekly conversation.
What changed was not the risk. The market had learned to paper the risk before it had learned to price it.
Concrete, steel, cooling towers, backup generators: ordinary construction risk has a market, and insurers have priced those exposures for a long time. The harder part is the contents. The most valuable things in the room are the chips, and at that size and scope, pricing each unit one by one would be, in the broker's word, "a nightmare." So the market did the next best thing. Because the assets depreciate fast, it wrote bespoke policies that pre-agree how much a fast-aging asset will be worth, before anyone has to find out what it is actually worth in the open.
Hold that image, because the whole story is compressed in it. Someone has to put a price on a room of chips that may be essential, obsolete, or both. And the price has to hold until the debt that paid for the chips comes due, which in the marquee deal of that spring is a full six years away.
The object
In late March 2026, a cloud company called CoreWeave closed a loan. The credit agreement was dated the thirtieth; the announcement came the next day. The size was 8.5 billion dollars. The structure carried a name only a lender could love, a delayed draw term loan, and one feature nobody had pulled off before: it earned an investment-grade rating while being secured by high-performance computing infrastructure and a single customer contract. Two ratings agencies signed off, A3 from one, A (low) from the other. The lenders were not a fringe outfit reaching for yield. The anchor was one of the largest credit arms in the world, alongside a broad syndicate of global banks, asset managers, and insurance investors.
Strip away the vocabulary and the deal is plain. A company borrowed billions to buy graphics chips and build them into working clusters, and the loan is secured by those assets and the promise of a customer paying to rent them. This is what "GPU-backed financing" points at. You can now pledge a room of processors something like the way you pledge a house.
Read the agreement and you can feel the caution in the fine print. The money does not come out all at once. The borrower draws only as assets are installed, powered, and running under contract, and each draw is capped at ninety cents on the dollar of the chips and gear that draw is buying. The collateral itself is broader: substantially all the assets of a bankruptcy-remote subsidiary set up to hold the deal, plus the customer contract that gives it cash flow. A spreadsheet built into the loan writes down, in advance, how fast the chips lose value: straight line, six years. Within sixty days of each draw, the borrower must hand the lenders serial numbers for the racks, and if it ever receives a list of the serial number of every individual chip, it must pass that along too. The broker's nightmare, valuing each unit, becomes the lender's paperwork.
Everything about it is careful. And the whole careful apparatus leans on one number. The loan matures at the end of March 2032. Six years out. The accounting life of the chips is six years. The depreciation math inside the loan is six years. The debt, the books, and the collateral are all pinned to the same figure.
Now set that six against the number that actually governs the chips. Two years before this loan closed, the dominant chipmaker launched a new architecture and said it succeeded the prior one, which had launched about two years earlier still. It claimed the new generation could do up to thirty times the inference work of the same number of the previous top chip, at up to twenty-five times lower cost and energy. You do not have to swallow the marketing to take the point. The problem is not that the chips break in six years. Something more awkward happens: a newer part can arrive that does the same work so much cheaper that it cuts what the old one commands, even while the old one is still earning. That recent leap is why a six-year useful life is not a neutral accounting line. It is a live assumption.
This is a mortgage on a thing that ages like a phone.
Ruling out the comforting explanations
There are three easy ways to make the anomaly go away. All three fail, and watching them fail is the argument.
First: this is reckless lending, a subprime shop that will finance anything. It is not. The rating is investment grade. The lender group is broad and sophisticated. The covenants are tight enough that a draw can be blocked if projected cash coverage slips below a set ratio. If you want to dismiss the deal as one credulous fool handing out money, the closing binder will not let you. These are careful people, and they signed.
Second: there is no real demand, it is a bubble, the buildings will sit empty. That one is harder to hold than the skeptics admit, and the other side deserves its strongest case. The demand is real and it is large. CoreWeave's backlog of contracted work ran to 66.8 billion dollars at year end, up from fifteen billion a year earlier. One large cloud provider reported 625 billion dollars of contracted future revenue; another reported a backlog of 240 billion. And the fact that cuts hardest against the obsolescence fear comes from the chipmaker's own finance chief, who said in February that not only were the newest parts sold out, but "even Hopper and much of the six-year-old Ampere-based products are sold out in the cloud." A six-year-old chip, still fully rented. If that holds, the collateral does not fall to scrap the day a new generation ships. The bull case is not stupid. It is the best reason to believe the loan is money good.
Third: the chips barely matter, the real collateral is the building and the power and the site, and those last for decades. There is something to this. The same broker will tell you the chips are interchangeable, and operators now build modular halls so boards can be swapped. But the loan does not lean on that comfort. It sizes each draw against the chips, ninety cents on their cost. It schedules depreciation for the chips. It demands serial numbers for the chips. Whatever else secures the deal, the people who wrote it were counting the processors.
So the comforting explanations fall away, and one thing is left standing. The deal works if, and only if, the chips keep earning across the full six years of the debt. Not the concrete. The chips. And whether a graphics processor's earning life stretches to match the life of a loan is not a fact anyone has in 2026. It is a hope that the depreciation curve behaves.
That is the contestable claim, and reasonable people will fight it. They will point to the sold-out old chips, the scarcity of power, the contracts already signed, and call the bet sound. They may be right. The point is narrower and colder: no one has yet shown what a pledged room of these chips could be sold for in 2032, and the structure is built as if that number were known.
Three ways this has gone before
Whether a bet like this is wisdom or self-deception is not a new question. Lending against fast-moving technology has a history, and it turns on four questions. Does the demand last at a price that services the debt? Does the asset's earning life reach the debt's maturity? Is there a resale market deep enough to re-value the collateral and sell it? And does the operator have the revenue and balance sheet to carry the cycle even if one of the others wobbles? Run those four questions across three old cases and the shape of the GPU mortgage comes into focus.
Start with the case nobody wants to name. In the late 1990s, companies laid fiber-optic cable across oceans on the belief that internet traffic was doubling every three months. One built a network reaching more than two hundred cities in twenty-seven countries and never turned a profit. It filed for bankruptcy in January 2002 with 12.5 billion dollars of debt. The network was not fake. Its chief executive testified that it carried more than five trillion dollars of financial transactions every business day and moved television signals under the Atlantic. The traffic was real. The demand simply arrived too late and too cheap to service what had been borrowed. On the four questions it failed three: the demand did not pay, there was no buyer at the borrowed value, and the operator had no profit to fall back on. A telecom consultant gave the warning that belongs taped to every credit desk financing chips today. The failed carriers, he said, "can't sell their assets to anyone else because they overvalued them to borrow more money in the first place." He expected the survivors to pick those assets up for pennies on the investment dollar.
Keep that story at the level of balance sheets and you miss where it landed. In March 2002 a congressional hearing room filled up to sort through the collapse. A congresswoman read into the record an account of former workers from a public forum in Rochester, New York, where more than two hundred and fifty people had come. Some had spent their life savings. Some had cashed in retirement accounts that had lost most of their value, after the severance they were promised stopped. Some had filed for personal bankruptcy or were about to lose their homes. Some figured they would have to leave town to find work. The fiber in the ground still worked. It just was not worth what the paperwork had said, and the gap between those two numbers came out of their lives.
Now the opposite case. There is a mature version of exactly this kind of lending, and it works. Airlines and leasing companies finance jets with asset-backed debt, and the lenders sleep at night, because a used aircraft answers all four questions well. Demand for air travel is durable. A plane's service life runs decades, well past any single lease. And it has a deep, boring resale market: registries, standardized inspections, appraisers, and buyers in every country, with decades of price data behind them. When demand shifts, older planes hold value; in the recent shortages, carriers flew retired jets back into service. A leasing company can raise billions against a fleet whose average aircraft is four years old and whose leases run past eight, and the math is not a leap of faith. It is a market.
Between the two sits the case that matters most, because it looks like the GPU mortgage and did not end in a hearing room. In 2010, one of the first big cloud providers was financing technology infrastructure it wrote off over three years, partly through leases, while building the service that would become its profit engine. Three years is a brutal asset life, shorter than any GPU schedule. By the fiber logic, that should have been reckless. It was not, because the other three questions answered yes: the cloud demand was real and recurring, the resale question mattered less while the boxes stayed busy earning, and the operator had the revenue and balance sheet to roll the cycle. Short useful life, on its own, did not make the financing foolish. What made it sane was everything around the short life.
Where the new deal lands
Put the GPU mortgage against the four questions and it splits.
On demand, it looks strong at first: contracted backlog, sold-out old chips, power scarce enough that last generation's parts still earn. But demand here does not mean a full room. It means a room that rents for enough. A cluster can run flat out every hour and still miss the debt if the price per hour falls faster than the loan pays down, and the rate a given chip commands does slide as cheaper parts arrive. On the operator, it is mixed but not naive: the borrower loses money, but the cash flow is contracted and the lenders are careful. On the asset's earning life, it is the open question, six years of debt against a two-year leap. And on the resale market, it is simply unproven. The deep, boring secondary market that makes a jet financeable, a liquid place to sell a used cluster at a known price, does not yet exist for these chips. The loan does not pretend otherwise. In place of a resale market it does not have, it substitutes the things it can write down: contracted cash flows, serial-number tracking, insurance endorsements, manufacturer warranties, coverage ratios, and a depreciation schedule agreed in advance. That is a serious answer. It is not the same thing as a market.
So the test the reader can now carry is not "will the buildings fill." It is more specific. A GPU mortgage behaves like a jet, and rationally, only if all four questions keep answering yes at once. The contracted customers keep paying, and the biggest of them stay creditworthy. The work keeps flowing out of one-time training runs into the steady, recurring business of inference, search, ads, and everyday software. Power stays scarce enough that older chips still earn. And the operator, or the structure built around it, stays strong enough to carry a single generation of hardware for the full term. Miss one and the deal drifts toward fiber, not because the chips stop working, but because the value on the paper stops matching the value in the world.
This is not one clever loan, which is why the answer reaches past CoreWeave. The four largest American technology companies were expected to spend close to 700 billion dollars on this buildout in 2026, against roughly 200 billion in combined free cash flow the year before. One bank reckoned the year's spending would consume nearly all the operating cash flow those firms generate, against a long-run norm near forty percent. A veteran analyst, asked whether the revenue would show up to justify it, gave the least reassuring answer available: "it's really hard to know the sustainability of top line." And the institution that serves as a bank for the world's central banks had begun calling the off-balance-sheet vehicles behind these projects "shadow borrowing," debt in all but name, sitting outside the balance sheets of the companies whose demand it depends on.
None of that forecasts a collapse. The history supports a risk, not a prophecy. Fiber is the warning, jets are the counterexample, and the 2010 cloud build is the reminder that a short-lived asset can be financed sanely when everything around it holds. What the history does prove is that a signed contract can comfort a lender for twenty years and still not save him, and that the people closest to the numbers tend to see the gap first.
Back at the desk
Which returns to the broker, and the small honest thing at the center of the job.
At that scale the individual chips cannot be priced one by one, so the industry agrees, in advance and in writing, what a fast-aging asset will be worth, and everyone signs, and the loan closes, and the rating holds. This is either the most sophisticated risk management the business has produced, a way to make an unpriceable thing bankable, or it is a very careful way of not looking at the one number you would rather not check.
From the desk, in the spring of 2026, the two are hard to tell apart. The policy is written. The commitments are signed. And as the chips come online in the dark, cluster by cluster, doing work someone is paying real money for, the paperwork answers every question but one: whether they will still be worth the loan in 2032, when the note comes due and someone finally has to find out what the room is worth.