CoreWeave Issued Yet More Debt
CoreWeave’s latest two DDTLs show that GPU debt isn’t really backed by GPUs. They're backed by customer credit, with the hardware as a recovery floor.
A disclaimer, prompted by some DMs I’ve received: I make extensive use of AI tools when doing research for, and writing, these posts. If that concerns you, this might not be the newsletter for you. This disclaimer has for a long time been displayed on the About page of this newsletter, but apparently many people have not seen it; therefore, this new disclaimer is presented front-and-center.
On March 31st, CoreWeave closed an $8.5 billion delayed draw term loan facility, dubbed DDTL 4.0, which was rated A3 by Moody’s and A(low) by DBRS, and which was secured against a contract with a single investment-grade enterprise customer. On May 18th, CoreWeave closed a $3.1 billion delayed draw term loan facility, dubbed DDTL 5.0, rated Ba2 by Moody’s and BB+ by Fitch, secured against contracts with two large, non-investment-grade customers.
Same issuer. Same lead arrangers, Morgan Stanley and MUFG. Same delayed-draw template, designed to match capital deployment to the GPU installation schedule. Same SPV architecture.
The facilities sit 225 basis points apart in spread and four notches apart in rating. That gap between these two debt facilities is the story.
The two deals
DDTL 4.0 priced at SOFR + 2.25% on the floating tranche, with a fixed tranche at approximately 5.9%. It matures in March 2032, was issued through CoreWeave Compute Acquisition Co. VIII, LLC, and was structured around a single anchor customer whose credit profile assuaged the rating agencies’ concerns. Goldman Sachs and JPMorgan joined MUFG and Morgan Stanley as coordinating lead arrangers on a privately syndicated transaction that was meaningfully oversubscribed.
DDTL 5.0 priced at SOFR + 4.50%. It matures in approximately 5.5 years, was issued through CoreWeave Financing DDTL V, LLC, and is backed by contracts with two large, non-investment-grade customers whose identities CoreWeave has not disclosed. Morgan Stanley and MUFG ran the books alone this time. The deal was also oversubscribed. And CoreWeave is calling it the first publicly syndicated HPC infrastructure-backed financing vehicle. It is designed, in their words, to expand the addressable investor base and enable secondary market trading.
The facilities differ in size, arranger group, covenant package, and critically, in their recourse posture. DDTL 4.0 was described as a first-of-its-kind non-recourse facility; DDTL 5.0’s debt is guaranteed by CoreWeave and key subsidiaries and is secured by substantially all assets of the borrowing group. But the core inputs are unusually close by structured-credit standards: same issuer, same collateral class, same delayed-draw deployment logic, same broad macro window, overlapping lead arrangers. And the recourse asymmetry, if anything, strengthens the signal. DDTL 5.0 has more structural protection, including parent guarantees and subsidiary support, and yet it still prices 225 bps higher than DDTL 4.0. Whatever is driving the gap, it’s powerful enough to overwhelm extra credit support.
The 225 basis point spread
Most spread comparisons in structured finance are noisy. Different issuers introduce management risk. Different arrangers introduce syndication-execution risk. Different vintages introduce market-timing risk. Different collateral introduces asset-quality risk. When you’re comparing two deals and trying to extract a signal about one specific variable, you’re usually doing it through several layers of fog.
DDTL 4.0 and DDTL 5.0 strip most of that fog away. The remaining differences, including size, syndication format, and recourse posture, matter, and I’ll come back to them. But the dominant variable separating these two facilities is the credit quality of the underlying customer contract. The 225 basis points is therefore the cleanest observed signal the GPU-backed lending market has produced on the contract-credit component of this asset class.
The inversion
The conventional narrative about GPU-backed lending is something along the following lines: It is an asset-backed product, in which GPUs collateralize the loan, loan economics track the asset’s useful life, and the main question to be answered by the underwriters is what the residual value curve looks like. The entire analytical apparatus treats the GPU as the credit engine.
DDTL 4.0 and DDTL 5.0 break that framing. If the GPU collateral were doing most of the credit work, you would not expect a four-notch gap between two facilities secured by broadly similar current-generation infrastructure inside the same issuer complex. Yet that’s what you get. What moves the rating is the customer on the other side of the take-or-pay contract.
This is the inversion: GPU-backed debt is not actually an asset-backed product in the way market participants describe it. It’s a contract-backed product with GPU collateral providing a recovery floor. The rating agencies are pricing the off-taker’s ability to pay through the contract term. The asset is the backstop, not the engine.
The distinction changes how you think about the entire stack of GPU-backed lending that’s been issued to date, and everything that’s coming next. If the asset were doing the credit work, the relevant analytical framework would be an obsolescence model—something like the Three-Curve Obsolescence Model I’ve discussed in prior work, tracking compute performance decay, secondary-market price decay, and power-efficiency decay. That framework matters for recovery analysis, but it’s a second-order consideration for initial pricing and rating. The first-order consideration, as the agencies have now demonstrated twice in eight weeks, is: who’s the customer, and can they pay?
The structural consequence
If the contract is the credit, then the cost-of-capital landscape for GPU infrastructure is set primarily by the credit quality of the customers you can win.
This creates a stratification problem. CoreWeave, with its anchor relationships—Microsoft, Meta, OpenAI—can access investment-grade debt pricing when the customer is investment-grade. That gets them SOFR + 2.25%. A second-tier neocloud without those relationships is structurally pinned to the Ba2/BB+ track or worse, regardless of operational quality. Their floor is SOFR + 4.50%, and probably wider, because CoreWeave’s track record, public-company transparency, and arranger relationships are themselves compressing the spread on the non-IG track in ways that a newer issuer wouldn’t benefit from.
The compounding effects are punitive. Higher financing costs mean thinner spread to GPU lease revenue, which means less margin to absorb obsolescence risk, which means less room to underprice incumbents to win the very investment-grade anchor tenants that would let them break onto the cheaper track. I described this dynamic with Luke Mellor in “The Trophy Deal Trap”, with the observation that getting the trophy contract is the only path off the high-spread track, but the trophy contract is itself the instrument that forces issuance under terms the issuer may not fully control.
CoreWeave has not named the two non-investment-grade customers behind DDTL 5.0. The obvious inference, given recent contract announcements, is that these are frontier AI lab credits rather than generic enterprise customers. The ambiguity is itself part of the point: the agencies are being asked to rate cash flows from counterparties whose strategic importance may be enormous but whose standalone credit machinery is still immature. The unrated AI lab is a genuinely new credit category. These are entities with massive contracted revenue commitments, substantial venture backing, and limited operating history as standalone credit risks. The agencies don’t have good machinery for them yet. The Ba2/BB+ rating on DDTL 5.0 is their first attempt at pricing that ambiguity, and 225 basis points above DDTL 4.0 is their answer, at least for now.
The public syndication question
CoreWeave is calling DDTL 5.0 the first publicly syndicated HPC infrastructure-backed financing vehicle. If that claim holds (and the deal’s mechanics actually enable secondary market trading in a way that prior DDTLs didn’t), this is a substrate-level change for the asset class.
Every prior CoreWeave DDTL was a private syndicate. The investors bought in at closing and held to maturity or negotiated bilateral transfers. DDTL 5.0, at least in principle, introduces price discovery between primary issuance and maturity. That’s the precondition for everything downstream: a tradeable benchmark, observable spreads, basis relationships against other instruments, and eventually the kind of liquidity infrastructure that an index or derivatives market would need to reference.
What’s interesting is that the publicly syndicated facility is the lower-rated one, not the higher-rated one. If that pattern holds, the tradeable market for GPU-backed credit develops first at the bottom of the rating stack, which is the opposite of how most corporate credit markets evolved. It’s worth watching whether that’s a one-off structural decision or the beginning of a pattern.
One caveat on the 225 bps in this context: public syndication is a pricing variable. A publicly syndicated high-yield-ish instrument has to clear a different investor base than a privately placed IG facility. Some portion of the spread differential may therefore reflect the costs of distribution technology and investor-base segmentation, rather than pure customer-credit pricing. That doesn’t break the thesis. Customer credit is still likely doing the heavy lifting. But the 225 bps premium above DDTL 4.0 is more precisely read as “contract-credit plus market-format plus rating-bucket effect,” not a clean single-variable isolate.
What would falsify this
Four things would tell us the bifurcation framework is wrong, or at least less durable than we think.
First: a non-IG-customer DDTL that prices materially inside SOFR + 3.50%. If a future facility backed by an unrated AI lab clears closer to DDTL 4.0 levels than to DDTL 5.0 levels, the bifurcation is a transient pricing inefficiency that the market closes as it learns to underwrite AI lab credit. This is plausible. The agencies are learning in real time; their models for unrated frontier AI companies will improve, and spreads could compress as the credit histories lengthen. But it hasn’t happened yet.
Second: an IG-rated facility backed by a diversified pool of mid-sized enterprise contracts rather than a single anchor customer. If diversification can substitute for single-customer IG quality in the eyes of the rating agencies, the bifurcation collapses into a more general credit-quality-of-cash-flows framework, and the specific identity of the customer matters less than the portfolio construction. This would be the more conventional structured finance outcome, and it’s the direction that the market probably migrates toward over time.
Third: a structural innovation, such as credit enhancement, insurance wrap, or SPV reorganization, that pulls non-IG-customer GPU debt into IG ratings without changing the underlying customer. If that happens, the contract isn’t the credit anymore; the structure is. This is the natural next move for arrangers if the bifurcation premium proves persistent, because there’s 200+ basis points of arbitrage sitting on the table for whoever figures out how to bridge the gap synthetically. Monoline-style wraps, overcollateralization, or reserve-account mechanics could all be candidates. The question is whether the rating agencies would look through the enhancement to the underlying customer credit, and based on what we’ve seen so far, the answer is probably yes.
Fourth: a stressed sale, refinancing, or early termination where GPU collateral materially protects lenders despite weak customer credit. If the recovery floor turns out to be high enough, because secondary-market GPU values hold, or because the infrastructure can be redeployed to new customers faster than the depreciation schedule assumes, that would tell us the asset is doing more credit work than the ratings currently imply. This would reintegrate the obsolescence framework into the primary underwriting story rather than leaving it as a second-order recovery consideration. No such event has occurred yet in the GPU-backed lending market, but the first credit stress will be the test.
The forward question
CoreWeave has now closed five DDTLs in roughly eighteen months. The cadence implies a sixth before year-end. Each one has introduced a structural innovation, from the initial GPU-collateralized template through IG ratings on 4.0 and public syndication on 5.0. The question for the next iteration is whether the bifurcation widens, narrows, or gets structured away.
Widening means the market is learning to discriminate between customer credit tiers with increasing precision. The 225 bps we see today is a first print; a second print, from CoreWeave or from another GPU infrastructure issuer, would begin to establish a curve, and with it, a more granular understanding of what specific customer-credit characteristics the market rewards.
Narrowing means AI lab credit is converging on enterprise credit, which would be a much bigger story than the structural finance angle. It would mean the agencies and the buy side are concluding that the frontier AI labs, despite being unrated and venture-backed, carry credit risk comparable to large-cap technology enterprises. That’s a statement about the permanence and bankability of AI demand, not just about one issuer’s financing strategy. And it’s the deeper theme lurking underneath the entire bifurcation: the AI infrastructure buildout is not capital-constrained by GPUs alone. It is constrained by the credit form in which future demand appears. DDTL 4.0 says: when AI demand is converted into an investment grade enterprise payment stream, capital treats it like infrastructure. DDTL 5.0 says: when the same demand arrives as frontier-lab, non-investment grade payment streams, capital still wants a high-yield spread.
I’d watch for narrowing first. The AI labs are building operating histories, generating revenue at scale, and in some cases moving toward public-market structures of their own. If the agencies upgrade their frameworks for this credit category, the 225 bps will look wide in hindsight, and the bifurcation will have been a snapshot of a market in transition, not a permanent structural feature.
Either way, the DDTL series has produced something the GPU-backed lending market badly needed: an empirical anchor. Two facilities, forty-eight days apart, 225 basis points of daylight between them. Whatever happens next starts from that number.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
What's interesting is that Moody's use an Expected Loss approach while Fitch use a default probability method. Setting aside the one notch difference in favour of Fitch, Moody's would normally be expected to produce a higher rating for a non diverse portfolio except where they have applied extreme haircuts to the same underlying GPUs, which we can assume here. All of this is speculation of course.