Onpode
Cover art for Why lenders bet on CoreWeave's customer contracts rather than its Nvidia chips

Why lenders bet on CoreWeave's customer contracts rather than its Nvidia chips

November 21, 2026 · 6 min

Cole Brennan & Malcolm Reeves

CoreWeave carried $35.6 billion in debt as of June 30, 2026, backed primarily by take-or-pay customer contracts — not Nvidia chips. Lenders underwrite the customer's creditworthiness, not hardware resale value. The real risk is whether contracts renew at comparable economics before the 7–10 year debt matures.

CoreWeave has pioneered a contract-backed financing model for AI compute infrastructure in which lenders fund Nvidia GPU deployments primarily against contracted future cash flows rather than the market value of the hardware itself. The company reported approximately $35.6 billion in total debt as of June 30, 2026, alongside $5.5 billion in cash and roughly $104 billion in revenue backlog.

0:006:28
Get the next episode on Artificial Intelligence →

Follow it free — new episodes land in your feed.

Or make your own — any topic, in minutes

More Onpode episodes on Artificial Intelligence →

About this episode

CoreWeave went public carrying $35.6 billion in debt, backed in large part by Nvidia chips. The obvious question is whether those chips hold their value. This episode argues that's the wrong question. What lenders actually underwrote was something closer to project finance: a separate legal entity holds a customer's take-or-pay contract alongside the hardware, and the loan amortizes as that contracted cash flows in. The chips are a fallback, not the bet. The episode traces how the structure works — from the August 2023 $2.3 billion delayed-draw term loan led by Magnetar and Blackstone, where each tranche unlocked only when a signed contract hit the table, to the covenants that require contracted cash flow to exceed debt service by a defined multiple. It also sits with the real tension: GPU economic life is estimated at two to three years; the debt runs seven to ten. And roughly $104 billion in backlog represents payments still owed under existing contracts — a number that shrinks every month unless new deals replace it. The episode doesn't resolve whether the structure survives its first big renewal cycle. It identifies exactly what to watch: does a coverage covenant trip before the end of 2027? Does a contract re-sign at a lower rate? The good outcome, as the episode puts it, is nothing happening. That's a useful frame for following this story.

Frequently asked

How does CoreWeave's GPU-backed debt actually work?

CoreWeave's debt is structured as project finance: a special-purpose entity holds a customer's take-or-pay contract, Nvidia chips, and a data center lease. Lenders lend against the contracted cash flows, not GPU resale value. Each loan tranche is released only when a specific signed customer contract is in place.

What is a take-or-pay contract in the context of CoreWeave?

A take-or-pay contract requires CoreWeave's customers to pay for reserved GPU capacity whether or not they use it, transferring demand risk from CoreWeave to the buyer. As of mid-2026, approximately 98% of CoreWeave's revenue came from committed take-or-pay arrangements, underpinning its $104 billion contract backlog.

Why are CoreWeave's Nvidia H100 chips considered a backstop rather than the main collateral?

CoreWeave lenders treat Nvidia H100 GPUs as a secondary backstop because GPU resale values can depreciate faster than loans amortize — Nvidia releases new generations quickly, compressing used-chip prices. The primary collateral is the customer's contractual payment obligation, not what the hardware fetches at auction.

What is CoreWeave's biggest debt risk?

CoreWeave's core debt risk is contract renewal: its 7–10 year loans can outlast the initial customer contracts, and no public evidence yet shows major contracts renewing at comparable economics. If customers re-sign at lower rates, contracted cash flows shrink while the debt schedule stays fixed, potentially tripping debt-service coverage covenants.

How much debt does CoreWeave have and who are the major lenders?

CoreWeave carried $35.6 billion in total debt as of June 30, 2026, against $5.5 billion in cash. Blackstone and Magnetar led the original $2.3 billion delayed-draw term loan in August 2023, secured by H100 GPUs. Reported interest spreads range from SOFR plus 2.25% to SOFR plus 5.50%, reflecting borrower credit quality differences.

Grounded in 12 sources
Markets for Compute and Inference:Structure, Financing, and Risk Transfer ↗ · arxiv.org
Here’s What CoreWeave’s Credit Woes Tell Investors About the AI Trade | Morningstar ↗ · morningstar.com
S-1/A ↗ · sec.gov
Medium ↗ · jrodthoughts.medium.com
CoreWeave: The Correction's End Is Near (Rating Upgrade) (NASDAQ:CRWV) | Seeking Alpha ↗ · seekingalpha.com
How to Underwrite AI Infrastructure Investments and Why GPU Financing Fails | American Compute ↗ · amcompute.com
Private Credit AI Datacenter Boom Hidden Concentration Risk ↗ · angelinvestorsnetwork.com
Nate Gregory - Briefs Finance ↗ · briefs.co
Insuring the AI Boom: Hyperscalers, Neoclouds and the Credit Markets – BSIC | Bocconi Students Investment Club ↗ · bsic.it
AI FINANCING PRIMER - 5 Min Read | Technology and Innovation Community ↗ · connect.cfauk.org
Data Center Primer ↗ · crefc.org
CoreWeave faces $10B debt wall as Nvidia rethinks funding role | Dealroom News ↗ · dealroom.co
Read transcript

Cole Brennan: Hey, you ready for this one? I've got a confession, there's a line from CoreWeave's CEO that's been stuck in my head: 'They pay the box. They don't pay me.' And the company carried $35.6 billion in debt as of June 30, 2026, so I genuinely can't tell who's on the hook.

Malcolm Reeves: Ready. That sentence is a door. Ask it plainly.

Cole Brennan: Plainly: if the money never touches CoreWeave, who is the lender actually lending to?

Malcolm Reeves: The box is a separate company. It holds three things: the customer's contract, the Nvidia chips, and the data center lease. Now, picture a bank lending against a tenant's signed lease. The tenant owes rent even if the rooms sit empty, so the bank cares whether the tenant pays, not what the building is worth.

Cole Brennan: Wait, empty rooms, still pays?

Malcolm Reeves: Still pays. And the promise is exactly that, the customer pays whether or not they use the

Cole Brennan: Hang on. If they pay either way, the chips are almost decoration. So why was the first big loan secured by Nvidia H100s at all?

Malcolm Reeves: Backstop, not foundation. August 2023: a $2.3 billion delayed-draw term loan, Magnetar and Blackstone leading, H100s as security. I'd call it the founding template, with a little caution. And the CEO described the negotiation as matching how fast the chips depreciate to how fast the loan gets repaid.

Cole Brennan: Delayed-draw. Meaning the money doesn't show up all at once?

Malcolm Reeves: No. Picture a closing lawyer in a conference room, a signed customer contract on the table. She checks that the collateral requirements are met, and only then does one tranche get wired. Each draw ties to one specific contract. No contract, no cash.

Cole Brennan: So the loan grows only as paper arrives.

Malcolm Reeves: Draw by draw. And the paper is take-or-pay: the customer reserves capacity and pays whether it uses it or not. Demand risk moves to the buyer, and the lender gets a cash flow it can bond. Now, that's project finance. The loan is sized to specific contracted cash flows, amortizes as customers pay, and releases excess cash only after debt service. Same logic as a toll road with a guaranteed minimum, or a pipeline.

Cole Brennan: I mean, so basically it's a guaranteed NBA contract for a player whose knees are going. The customer pays for idle servers, and the chips are the aging athlete.

Malcolm Reeves: Hold the knees for a moment. Scale first. About 98% of CoreWeave's revenue is committed take-or-pay. As of June 30, 2026: $35.6 billion total debt, $5.5 billion cash, and roughly $104 billion in backlog, the remaining payments owed under signed contracts. You'll see about $25 billion quoted elsewhere. That's a different snapshot, not a contradiction. Quote a debt figure with its date.

Cole Brennan: Wait, backlog is payments still owed? Then it shrinks every month a customer pays, unless new contracts replace it.

Malcolm Reeves: It does shrink. And that's where the clean version grows a clock. Bear cases, FutureX Capital's analysis among them, put a GPU's economic life at two to three years. The debt runs seven to ten. Bull cases say chips last longer, so the gap is overstated. The sources don't settle it.

Cole Brennan: But you said the contract is the point. Why does the chip's age matter?

Malcolm Reeves: Because the chips are the backstop, and Nvidia ships a new generation fast. An H100's resale value can fall quicker than the loan amortizes, and then the backstop is thinner than the paper assumed. And there's a covenant, the debt-service coverage ratio. Contracted cash flow has to exceed debt service by a defined multiple. Easy in year one, every contract fresh. Later years are the worry, if contracts expire and re-sign cheaper.

Cole Brennan: Whoa, wait. Say that again. The contracts expire while the debt's still outstanding? I'd been picturing the chips just getting old, but no, it's the customer who can walk before the lender's paid.

Malcolm Reeves: Yes. And here the history cuts the other way. The 1860s railroads used the Philadelphia plan: a trustee held title to the rolling stock until the debt was paid. But that debt amortized faster than the cars wore out. Here the tenor can outlast the chip. I'd not stretch the analogy past that.

Cole Brennan: Huh. So the railroads had the cushion built in. Now what if it's one customer?

Malcolm Reeves: Then it bites harder. One Meta contract reportedly backs $8.5 billion of debt. Take a hypothetical: Microsoft cuts external compute spend. Contract cash flows weaken, and GPU values drop at the same moment. Forced sales then compress GPU prices market-wide. That risk sits with Blackstone and Magnetar, and the lenders who joined them.

Cole Brennan: That's fragile, man. Like, the infrastructure lead staring at a cluster at 40% still owes the full amount, and the lender's rereading the customer's balance sheet. I mean, some people would say that's a ticking clock.

Malcolm Reeves: Or a known variable. The offtaker's credit is the credit, and Meta is investment grade. Lenders already price that: SOFR plus 2.25% against SOFR plus 5.50%. Stijn Van Nieuwerburgh's NBER paper tracks this capital stack shift, and Oracle and CoreWeave sit inside AI debt aggregations nearing $500 billion. Nobody's hiding it.

Cole Brennan: Priced for today, though. Not for the renewal.

Malcolm Reeves: Yes, and that's the whole test, I think. The first wave of contracts runs out while the debt is still on the books. Then the question is plain: does the customer re-sign at the same rent? Nothing in what I have says that has happened yet, at comparable economics or any other.

Cole Brennan: Huh. So I came in asking the wrong thing. I kept circling what the chips are worth, like, what's an H100 fetch at auction. But that was never really it. It's whether the contracts stay good as long as the debt does. The lender underwrote the contract, so the loan is only as long-lived as the paper.

Malcolm Reeves: That's it. And the signal is dated, which I like. Through the end of 2027, you watch GPU-backed loans pass that two-to-three-year window. Does one get downgraded? Does a coverage covenant trip? If they get through quietly, the structure held.

Cole Brennan: So the good outcome is nothing happening.

Malcolm Reeves: Quite. A quiet year is the evidence. And the bad outcome has a shape too: the box left holding older chips and lower rents. Same debt, same schedule, a smaller check coming in.

Cole Brennan: That's the line from the start, then. They pay the box. I figured it meant the money skipped CoreWeave. Now it sounds more like the box is the one sitting there when the lease runs out.

Malcolm Reeves: Mm. It carries more than it seemed to. I'd say that's a decent sentence to have had stuck in your head.