GPU-backed loans hit investment grade in three years. Now insurers and pension funds hold paper tied to chips that may outlast their value. Here’s the risk.
getty
On March 31, 2026, CoreWeave closed a loan worth $8.5 billion that was secured against computer chips, and Moody’s rated it A3 while DBRS rated it A (low). Both grades sit in investment-grade territory, the tier that pension funds and insurance companies are allowed to buy. In August 2023, the same company had borrowed against the same type of hardware at a floating rate of about 15%, the kind of price a lender charges when it is nervous. The hardware is comparable across the two deals but the price of the money is not.
A high-end graphics chip can lose roughly half its resale value within three years. The loan that closed in March runs until 2032. For that math to work, the chips have to hold enough value, or the customer contracts behind them have to keep paying, for the back half of a six-year loan. This piece explains how the price of that debt fell so fast, who is buying it now, and how closely the setup resembles the mortgage securities that blew up in 2008. It resembles them in some ways and differs in others, and the differences matter as much as the similarities.
How the price fell
The numbers come from CoreWeave’s own filings with the Securities and Exchange Commission, so they can be checked rather than taken on faith.
CoreWeave raised $2.3 billion in debt in August 2023, in a deal led by Magnetar Capital and Blackstone. Coverage of the sector put the floating rate at around 15%, a level normally reserved for risky borrowers and unfamiliar collateral.
In July 2025, the company closed a $2.6 billion facility it labeled DDTL 3.0. The 8-K it filed states the rate: 4.00% over SOFR, the benchmark that replaced Libor as the reference rate for most floating-rate loans. That facility was tied to a long-term contract with OpenAI and comes due in 2030.
The March 2026 deal, DDTL 4.0, priced the floating portion at SOFR plus 2.25% and the fixed portion at about 5.9%. CoreWeave’s announcement described it as the first investment-grade-rated financing secured by this class of computing hardware together with a customer contract. The spread over the benchmark dropped by more than 175 basis points between the 2025 facility and the 2026 one. Over the full stretch from 2023, the decline is much larger. Lenders repriced this borrower in under three years.
Two definitions are worth setting down, because the words matter here. A term loan is money that a bank or a group of lenders hands to a borrower directly. A securitization pools many loans, cuts them into bonds, and sells the bonds to investors. CoreWeave’s deals are term loans rather than securitizations, a distinction that a lot of coverage runs together and that comes back later in this piece.
How a chip turns into rated debt
The mechanism behind these deals is borrowed from older corners of finance. A lender that does not want to fight through a bankruptcy if the borrower fails will insist that the assets it cares about be moved into a separate company built for that one purpose. The loan is then secured against that company alone. The DDTL 4.0 filing names the entity holding the collateral: CoreWeave Compute Acquisition Co. VIII, LLC. If the parent company runs into trouble, the lenders are meant to still have the chips and the contracted revenue sitting inside that walled-off subsidiary.
Lawyers call this non-recourse, ring-fenced lending. Airlines borrow against their planes the same way, and shipping companies borrow against their vessels. The thing the lender can seize is the plane, the ship, or in this case a room full of Nvidia chips plus a signed contract with a large cloud buyer.
The rating agencies were slow to get comfortable. Sector coverage notes that S&P kept a ceiling of A+ on data-center asset-backed securities until 2025, on the grounds that the market had too short a track record. After that ceiling came off and the first investment-grade marks arrived, a different set of buyers could participate. Insurers and pension funds are usually prohibited by their own rules from holding much debt rated below investment grade. An A rating opens the door to that money.
The disagreement at the center of the deal
The value of the collateral is disputed, and the dispute runs all the way up to some of the biggest names in finance.
In November 2025, the investor Michael Burry, who bet against the housing market before the 2008 crash, accused the large cloud companies of flattering their profits. His argument was that they spread the cost of their chips across five or six years on their books, while the hardware actually loses most of its economic usefulness in two or three. A longer schedule means smaller yearly charges and larger reported earnings. Burry put the gap at roughly $176 billion of understated depreciation across the industry between 2026 and 2028.
The companies’ own accounting shows the question is live. Meta lengthened the assumed useful life of certain servers and network equipment to 5.5 years starting in fiscal 2025, a change it said would cut its depreciation expense by about $2.9 billion that year. In the same window, Amazon went the opposite way, shortening the life of some of its servers from six years to five and taking about $920 million in accelerated depreciation, and it named the pace of AI development as the reason. Two of the largest buyers of this hardware looked at the same equipment and reached opposite conclusions. A lender holding those chips as security has to pick a side of that argument, whether it wants to or not.
THE STANDOUT NUMBER: Meta stretched the assumed life of certain servers to 5.5 years in 2025. Amazon cut a slice of its fleet from six years to five in the same period. The lifespan of the collateral is a judgment call, not a settled fact.
CoreWeave makes the case for longer lives. The company has pointed out that older chips keep renting and keep reselling at real prices. By one tracker’s count, used H100 systems in their third year have sold for about 45% of the cost of a new unit. That argument carries weight. A chip too slow to train the newest model often gets moved to cheaper work, where it keeps earning. The optimistic reading is that the value steps down rather than falling off a cliff.
There is also a timing problem. If a chip earns most of its keep in its first few years and the loan against it runs close to a decade, the later years of the loan depend on the customer contract holding up and on the chip still being worth seizing. The same dollar of chip value tends to get counted more than once. It shows up as a sign of demand, then again as booked revenue, then again as the collateral behind a loan. Stack enough of those dollars together and the structure grows fast.
Who is holding the paper
This is where the term-loan story meets the larger one.
CoreWeave’s facilities are loans. The same approach is moving into the bond market: bundle the chips and the contracts together, get a rating on the bundle, then sell the risk to big institutions. JPMorgan’s CMBS research team, headed by Chong Sin, told Bloomberg in early February 2026 that data-center securitization could reach $30 billion to $40 billion a year in both 2026 and 2027, up from about $27 billion in 2025. That would be 7% to 10% of the combined commercial-mortgage and asset-backed issuance for those years.
The projection deserves a careful reading, because the real figure is narrower than the scary one. JPMorgan is measuring data-center securitization in general, not GPU-backed paper in particular. The CoreWeave term loan does not flow directly into those bond pools. What ties the two together is direction. Lenders have decided this class of asset earns an investment-grade rating, and once that view holds, it carries over, first into term loans and then into securitized bonds. The repricing in the loan market came first, and the bond pipeline is following behind it.
The buyers at the end of the chain are the reason this matters beyond finance desks. The March deal was anchored by Blackstone’s credit and insurance arm, with asset managers and insurance investors taking part. That is annuity money and retirement money, the cautious end of the market, now a step or two away from a warehouse of chips that lose value every year.
The 2008 comparison
The comparison to subprime mortgages is tempting, and it holds up partway.
The shared part is the machinery. You take an asset whose future value is uncertain and put it inside a legal structure. You get a friendly rating on it. Then you sell it to investors who trust the rating more than they understand the asset underneath. That pattern ran straight through the 2008 crisis, and the resemblance is real.
The differences cut the other way, in the lenders’ favor. The subprime pools were full of borrowers nobody had verified. These chip deals carry named customers with deep pockets, the large cloud companies, on long contracts, along with credit cushions and walled-off collateral. The rating agencies that held the sector at A+ until 2025 were not handing out top marks casually. A stack of Nvidia chips under contract to a trillion-dollar company is sturdier collateral than a no-documentation mortgage from 2006.
One more boundary is worth drawing clearly, because it is easy to blur. CoreWeave is defending a securities-fraud class action filed in February 2026 in federal court in New Jersey, Masaitis v. CoreWeave. The suit exists and sits over the company at the center of this story, so it belongs in the picture. Its claims, though, are specific: that CoreWeave misled investors about its ability to meet customer demand and about its dependence on a single data-center supplier. The case does not concern the chip collateral, the depreciation question, or the structure of the loans. Folding it into the financing argument would claim a link the court papers do not support.
The story has no ending yet, and pretending otherwise would be a stretch. The chips might keep their value long enough to back loans that mature in 2032, or they might lose it faster than the paperwork assumes. The contracts behind them might hold or might weaken. Years will pass before anyone knows. For now, the bet has been placed in public and on the record, using money that belongs to a great many people who have never heard the phrase “delayed-draw term loan.”
The Bottom Line
Lenders now treat chip-backed debt as investment grade, and cautious institutional money is buying it, even though the loans outlast the chips’ likely useful life by years. Whether that proves sound will not be clear for some time. If you hold an annuity, a pension, or a bond fund, it is reasonable to ask your provider how much exposure the portfolio has to data-center or AI-infrastructure credit, and what depreciation assumptions sit behind it, before deciding how much of this risk you are comfortable carrying.