What it is
“Neoclouds” are a class of companies — CoreWeave, Nebius, Crusoe and similar firms — that rent out GPU computing capacity to AI labs and enterprises, competing with the cloud divisions of Amazon, Microsoft and Google but specializing narrowly in AI infrastructure. Building that capacity is enormously capital-intensive: a single large GPU cluster can cost billions of dollars in Nvidia hardware before it earns a dollar of revenue. Neoclouds have financed this buildout mostly with debt rather than equity, and the debt is structured in a specific way.
The typical structure works like this. A neocloud sets up a special purpose vehicle (SPV) — a legal entity walled off from the parent company’s other obligations — to hold a batch of GPUs. Lenders extend a loan to that SPV, secured by two things: the physical GPUs as collateral, and the revenue from a customer’s “offtake” contract, a multi-year commitment to rent that capacity. The loan is often non-recourse, meaning if the SPV fails, the lender can seize the GPUs and the contract but cannot pursue the parent company’s other assets. CoreWeave’s $8.5 billion facility closed March 31, 2026, illustrates the model at scale: it was rated A3 by Moody’s and A(low) by DBRS — the first investment-grade rating on this kind of infrastructure debt — and is secured by “substantially all assets” of a dedicated CoreWeave subsidiary, tied to contracts with a leading AI enterprise customer, maturing in 2032.
What makes the debt investment-grade is not the neocloud’s own balance sheet, which typically carries a speculative-grade rating on its own. It is the creditworthiness of the customer that signed the offtake contract. Lenders are effectively betting on whether a company like Meta or Microsoft will keep paying for compute through the life of the loan; the GPUs are the fallback if that bet fails.
Why it matters for AI governance and narratives
This financing structure sits underneath the industry’s boom narrative, and it has become a point of contention in how observers frame the sustainability of the AI buildout. Proponents present GPU-backed debt as a maturing, investment-grade asset class — evidence that AI infrastructure demand is durable enough to attract insurers and pension-fund-adjacent capital (Blackstone Credit & Insurance anchored CoreWeave’s facility). Critics, including short-sellers and some Russian- and Chinese-language technology press, describe the same structure as a “house of cards”: GPU collateral is assumed to depreciate on a six-year schedule, but rental prices for H100-class chips fell from roughly $7–10 per hour in early 2024 to $2–4 by late 2025, and each new Nvidia generation re-prices the fleet behind it. If real-world depreciation runs faster than the loan term assumes, the collateral backing tens of billions of dollars in debt could be worth less than the debt itself well before maturity.
A second framing concerns circularity: Nvidia does not just sell GPUs to neoclouds, it also backstops some of the financing and, in at least one disclosed arrangement with CoreWeave, has agreed to buy unused capacity through April 2032. That makes Nvidia simultaneously the equipment supplier, a credit backstop, and a demand-side buyer of its own customers’ unsold output — a structure that lets one company’s results validate the others’ balance sheets, which is exactly the dynamic the observatory watches for when tracking whose framing of AI’s economics is advancing and whose interests that framing serves. It also means demand-side signals, such as a large customer reconsidering its GPU commitments, cannot be read as market feedback alone; they may propagate directly into the credit quality of debt held by lenders far from the underlying AI business.
Key facts and dates
CoreWeave’s total debt grew from under $8 billion in 2024 to over $21 billion by mid-2026, funded through a sequence of Delayed-Draw Term Loan (DDTL) facilities. The $8.5 billion DDTL 4.0 facility, closed March 31, 2026 (arranged by MUFG, Morgan Stanley, Goldman Sachs and JPMorgan), was the first HPC-infrastructure financing to receive investment-grade ratings. A subsequent DDTL 5.0 facility was structured as the first publicly syndicated vehicle of this type, intended to widen the investor base beyond specialist credit funds. Separately, Nvidia’s agreement to purchase CoreWeave’s unsold capacity through April 13, 2032 was disclosed as an arrangement with an initial value of $6.3 billion. Industry estimates put more than $20 billion in loans across the sector collateralized by Nvidia accelerators as of mid-2026. The credit quality of at least one A3-rated facility became more uncertain after Meta signaled in July 2026 that it was exploring selling its own GPU capacity externally — a move that would reduce its need to lease from neoclouds and, by extension, weaken the offtake contract underpinning that debt.
Where to learn more
- CoreWeave: Closes Landmark $8.5 Billion Financing Facility — the company’s own announcement, with deal structure, ratings and lead arrangers
- CoreWeave: Closes $3.1 Billion Loan Facility, Expanding Access to Public Markets — the follow-on publicly syndicated facility
- SemiAnalysis: Nvidia GPU Debt Backstop Unleashes the AI Project Trinity — technical explainer on the capital/offtake/datacenter financing loop
- Quartz: GPU-collateralized debt explained — accessible overview of collateral and depreciation risk