This post was written in collaboration with Luke Mellor, Partner & Head of Structuring at Optio Investment Partners.
On March 31st, CoreWeave closed an $8.5 billion delayed draw term loan facility–DDTL 4.0–rated A3 by Moody’s and A(low) by DBRS. It’s the first investment-grade rated financing secured by high-performance compute infrastructure and an associated customer contract. The press coverage focused on what this signals about institutional confidence in AI infrastructure.
The press missed the point. The real story isn’t that GPU-backed debt achieved investment-grade status. It’s how it achieved investment-grade status, and what that mechanism tells us about the GPU debt market that everyone assumes is coming.
What the Rating Agencies Are Actually Rating
The rating agencies looked at Meta’s creditworthiness and the robustness of the Master Service Agreement, applied a haircut for the operational layer that sits between Meta’s payments and debt service on the facility, and called it a day.
Meta’s own credit rating is Aa2. This facility is rated A3. The spread between those two ratings is the market’s pricing of CoreWeave’s operator dependency risk, which is the possibility that CoreWeave fails to deliver on service level agreements, trips a force majeure clause, or otherwise disrupts the contracted cash flows that service the debt. That spread is the analytical story, and it’s one the press has entirely ignored in favor of the simpler narrative about GPU-backed debt arriving as an asset class.
An interesting wrinkle: the investors aren’t even exposed to GPU implementation risk. Lenders’ funds can only be drawn after CoreWeave passes Nvidia’s Level 3 testing, which ensures the GPUs perform as designed in a stable operating environment.
The Trophy Deal That Can’t Be Replicated
This is the kind of deal that every bank and institutional lender wants on its book: a high investment-grade obligor, take-or-pay cash flows, no re-leasing risk, no refinancing risk. It’s the deal a credit committee approves without breaking a sweat. But trophy deals are trophy deals precisely because they’re atypical.
DDTL 4.0 earned its rating through a specific and narrow set of structural preconditions. Understanding those preconditions is the key to understanding whether this deal is a template or an artifact.
A single mega-cap counterparty. The facility is backed by contracts with Meta worth at least $19 billion. This isn’t a portfolio of enterprise compute contracts. It’s a single-name credit exposure to one of the most highly rated corporate counterparties on the planet. The take-or-pay structure means payments are fixed and independent of actual usage, and the MSA cannot be terminated for convenience. The credit analysis, in essence, is: under a single-A stress scenario, will Meta honor its contractual obligations? The answer is straightforward: Yes, Meta will honor its obligations.
Step-in rights that make the operator replaceable. If CoreWeave fails operationally or financially, the project either continues to operate or lenders can access their collateral. Both scenarios are factored into the stressed cash flow analyses performed by the ratings agencies.
Full amortization within the MSA term. The debt gets paid off entirely within the life of the Meta contract. No maturity wall. No balloon payment reliant on asset sales. No refinancing risk. No re-contracting risk. The debt service coverage ratio, net of essential OpEx and power costs, averages 1.26x through 2031 under the borrower’s base case, with a covenant floor of 1.15x. Adequate coverage, but not enormous: the leverage is being pushed fairly tight against the net contracted cash flows.
Hedged power costs against fixed revenue. Top-line payments from Meta are fixed. CoreWeave has incorporated power price mitigation strategies that materially reduce cost-side variability, a prerequisite for the deterministic cash flow modeling that rating agencies require for investment-grade. The $4.459 billion in floating rate commitments are 95% hedged over the expected tenor of the financing.
Taken together, these features produce something that looks less like a technology financing and more like a power plant with GPUs instead of turbines. The contractual structure converts a volatile, rapidly evolving technology asset into a stable, predictable infrastructure cash flow stream. That’s the achievement. It’s also the limitation.
Solving the Collateral Bifurcation Problem for One Deal
One of the fundamental challenges of GPU-backed debt is what might be called the collateral bifurcation problem: how do you separate facility-layer asset value (the physical GPUs) from compute-layer cash flow value (the revenue generated by running workloads on that hardware)? In most GPU-backed financing, lenders face an uncomfortable question: If the borrower defaults, what are the GPUs actually worth? The answer depends entirely on whether there’s a contract attached to them, and if so, whose contract, at what rate, and for how long.
DDTL 4.0 solves this problem by refusing to separate the two layers. The collateral package bundles the hardware, the contracted cash flows, the step-in rights, and the power hedging into a single structure. The rating agencies aren’t being asked to value GPUs in isolation, as the data points for GPU residual values are either insufficient or too volatile to sustain an investment-grade rating. They’re being asked to value GPUs secured by investment-grade revenue from Meta. That’s a fundamentally different question. A GPU without a contract attached to it is a depreciating technology asset with uncertain residual value. A GPU with a $19 billion Meta take-or-pay behind it is infrastructure.
This is elegant financial engineering. It is also, by its nature, bespoke.
This Doesn’t Generalize
For anyone who can’t convince Meta, or another investment-grade counterparty, to commit to billions in take-or-pay payments over the life of the financing, this structure is unreachable. And that’s what the “landmark” narrative obscures.
The cost-of-capital trajectory tells part of the story. CoreWeave was paying roughly 9% fixed in July 2024. DDTL 4.0 carries a floating-rate tranche at SOFR + 225 bps and a fixed-rate tranche at approximately 5.9%, or about 300 basis points of compression in under two years. But this escape from the GPU debt treadmill is only available because the Meta contract eliminates refinancing risk entirely. Without a take-or-pay long enough to support full amortization, you’re back on the treadmill.
Most neoclouds operate with shorter-duration contracts spread across a variety of compute buyers. No single enterprise customer provides the kind of credit anchor that Meta represents. Typical off-take providers may not access the capital markets or may have nascent trading histories. Either way, they often lack credit ratings. Operator dependency can’t be easily structured away. Smaller providers lack the standardized, transferable operational infrastructure that makes step-in rights credible. And power costs are often variable against revenue streams that are themselves uncertain.
Here’s the counterintuitive implication: DDTL 4.0 may actually widen the cost-of-capital spread in the neocloud sector, rather than narrow it. CoreWeave can now borrow at investment-grade spreads. Everyone else remains in high-yield territory. The gap between the two just became more visible, more quantifiable, and more structurally entrenched.
This is the opposite of the prevailing narrative, which is that DDTL 4.0 sets a precedent for GPU-backed financing. It opens the door for one very specific type of GPU-backed financing. For the rest of the market, the door remains closed, and now there’s a benchmark on the other side showing exactly how much cheaper capital could be if you could get through it.
What Would It Take to Make This a Market?
If trophy deals like DDTL 4.0 are the ceiling, what does the floor need to look like? What would it take for the rating agencies to underwrite not just a single Meta-backed facility, but a portfolio of GPU obligations? What does it take to make a true GPU lease ABS market?
The answer probably lies in the history of collateralized loan obligations. Three capabilities would need to exist that don’t exist today, but they aren’t equally difficult, and they aren’t equally important.
The first two are extensions of existing methodology. The third is the binding constraint.
Multi-obligor risk modeling. A pooled portfolio of GPU leases backed by corporate obligations, analogous to leveraged loan CLOs or equipment lease securitizations, requires statistical models for default correlation, recovery rates, and prepayment behavior across a diversified set of compute customers. The major rating agencies already have collateral analysis tools for this: S&P’s CDOEvaluator, Fitch’s PCM, and their equivalents. Portfolios of debt obligations are analyzed for intra- and inter-industry correlation and obligor default risk using Monte Carlo analysis to derive scenario default rates across rating stresses from AAA to CCC. It’s conceivable that the agencies adapt these tools for portfolios of GPU leases.
One complication worth flagging, and one we intend to explore more fully in a future piece: data center operating risk may not be independent of off-taker credit risk. The ongoing solvency of both GPU operators and their customers is tied to the same underlying compute economy. These risks may be correlated in ways that existing CLO frameworks don’t capture.
Secondary cloud compute provider credit analysis. The rating agencies need a framework for assessing operator risk across the neocloud tier. These frameworks are in their infancy, but the agencies will eventually need to assign GPU operator ratings across the industry, much like those used in project finance transactions.
Both of these capabilities are tractable. They require effort, data, and methodological development, but they’re evolutionary extensions of things the ratings agencies already know how to do.
A forward curve for compute pricing. This is where the problem gets hard. Rating agencies modeling a five-year GPU lease ABS need to project what on-demand compute rates will look like across different SKUs–H100, H200, B200, and whatever comes next–over the life of the security. Those projections determine residual value assumptions, collateral coverage ratios, and debt service capacity under stress scenarios.
Today, no such forward curve exists. There is no liquid market for forward compute pricing, although a number of entrants are widely anticipated in the next 12 months. There is no standardized benchmark that rating agencies, lenders, or investors can reference to model how GPU lease revenues will evolve over time. Without it, the rating agencies are flying blind on the single most important variable in GPU-backed ABS: what the collateral will be worth in three years.
The Missing Primitive
A functioning compute derivatives market would solve a significant portion of this problem. If you can’t get Meta to sign a $19 billion take-or-pay, the next best path to bankable GPU debt is a liquid, transparent pricing benchmark that lets lenders model residual value risk on the collateral.
That benchmark would provide mark-to-market capability for GPU collateral, replacing book depreciation schedules with live index pricing. This is similar to the way auto lenders reference residual value guides or commodity lenders reference spot prices. It would produce forward curves by SKU as a byproduct of trading activity, giving ABS analysts the inputs they need for residual value modeling. And it would provide hedging instruments that allow neoclouds to transform their risk profiles: hedged cash flows are more predictable, less volatile, and more amenable to leverage.
The Bottom Line
DDTL 4.0 proves that GPU-backed debt can achieve investment-grade status. It also proves that the conditions required to get there–a $19 billion take-or-pay from one of the most creditworthy counterparties on the planet, full amortization within contract terms, step-in rights, hedged power–are extraordinarily narrow.
With over $1 trillion in GPU capex estimated to come online by 2030, the market can’t scale on trophy deals alone. The financing infrastructure that would let it scale–multi-obligor risk models, neocloud credit frameworks, and above all, a liquid forward curve for compute pricing–doesn’t exist yet. Until it does, DDTL 4.0 is a proof of concept for a market that needs to be built.
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:
Nice piece and good observation about the predictability of the cash flow stream!
I've also noticed that the DDTL 3.0 post-draw condition to maintain at least 85% of the depreciated GPUs value is gone. Does it mean that the 4.0 DDTL is more aligned with what you would expect from GPU-backed financing?