GPUs as Collateral: How Neoclouds Finance Compute

In the summer of 2023, CoreWeave borrowed $2.3 billion at SOFR + 8.75% from funds managed by Blackstone and Magnetar. With fees, the loan’s effective rate comes to about 15%, the company says. Three years later, in June 2026, IREN closed a $3.6 billion financing at an average cost of 6%, and Fitch gave the deal an A rating. Both deals were secured by the same thing – NVIDIA GPUs and the revenue from renting them out.

Neoclouds build capacity ahead of demand, and much of it is paid for with debt. As we noted in our first article, CoreWeave alone had about $35.1 billion of debt at the end of Q2 2026. The trouble is that GPUs make awkward collateral. Each new generation of chips pushes down the value of the last one, and there was no public price for an hour of GPU time until price indexes started to appear. So lenders mostly look past the hardware to the contract that pays for it.

This is the fourth and final article in our series on neoclouds. The second article looked at how neoclouds sell capacity to customers. Those contracts are already written with future lenders in mind. The provider can pledge the contract, for instance, and the data center has to sign an agreement giving lenders access to the equipment. The third article showed how thousands of private deals feed into a GPU-hour price index, why the market needs exchange-traded futures, and how futures could make neocloud lending work more like lending to oil producers.

Those futures are going to take longer than planned. The CFTC’s review of CME’s contracts was set to wrap up on November 9, 2026, but on November 4 the exchange agreed, at the regulator’s request, to push the deadline to January 8, 2027. Unless the CFTC issues a notice of non-approval by then, the contracts will be deemed approved. The deadline moves up if another exchange self-certifies, wins approval for, or lists comparable contracts before then.

Lenders aren’t waiting, though. GPU-backed loans are already getting done, and that’s what this article is about. First we cover who the parties are, where the money goes, and what lenders take as security. Then we walk through three public deals by CoreWeave, Nscale, and IREN.

The Building Blocks of a Deal

Deal structures vary, but the parties are usually the same.

Neocloud. The group’s parent company. It finds customers, builds or leases data centers, and runs the equipment. A loan for a particular GPU cluster usually goes to a separate project company, but the parent stays involved. It may sign the customer (off-taker) contract itself and then assign it to the project company. It may guarantee the loan. Or it may set things up so that an intermediate holding company borrows the money and lends it down the chain.

Special purpose vehicle (SPV). The company that buys and owns the servers. It either signs the customer (off-taker) contract or takes it over by assignment, and customer payments go into its accounts. Lenders usually limit the SPV to a single project and don’t let it raise other financing without their consent. The SPV can still have more than one lender, as long as they all fund the same project and share the same collateral. IREN’s SPV, for example, has both a bank loan and bonds. The next project, though, will usually get its own SPV. All of these restrictions are there to ring-fence the project from the rest of the group.

Customer (off-taker). The party that has agreed to buy capacity years in advance on a take-or-pay basis. It pays for the agreed volume whether it uses it or not. In the deals below, the customers are Microsoft and a Singapore subsidiary of ByteDance.

Lenders and collateral agent. Private credit funds, banks, or bond investors put up the money. The security usually goes to a collateral agent, which holds it for all the lenders and enforces it if the borrower defaults.

Supplier and integrator. Manufacturers like Dell sell the servers, and integrators like Computacenter assemble them. Loan proceeds are often paid straight to them, bypassing the borrower.

Data center and manager. The servers sit either in the neocloud’s own data center or in someone else’s. The neocloud usually runs them itself under a management agreement.

The security package is similar from deal to deal. The collateral agent takes security over the GPUs, 100% of the borrower’s equity, and the rights under the customer contract. It usually also takes the SPV’s bank accounts, where the revenue comes in. Cash in those accounts is paid out through a waterfall. Operating expenses and interest come first, then principal and reserves, and whatever is left goes to the neocloud.

If the borrower defaults, the lenders have two options. They can (i) take over the whole SPV, contract and revenue included, bring in a new operator, and keep collecting from the customer. Or they can (ii) sell the GPUs and other equipment. The first option usually recovers more, since a running cluster with a signed contract is worth more than used hardware.

The SPV sits in the middle of the diagram. It owns the GPUs and is the party to the customer (off-taker) contract. The lenders lend to the SPV, but the money goes straight to the supplier, which delivers the servers and transfers title to the SPV. The servers are housed in a data center under a colocation agreement, and the customer pays the SPV for capacity. The SPV uses that revenue to pay the lenders interest and principal through the waterfall.

The neocloud owns 100% of the SPV and pledges that equity. It also runs the GPU cluster, either directly or through another group company under a management agreement. The servers, the customer contract, and the SPV’s bank accounts (inside the dotted perimeter) are pledged to the collateral agent, which acts for the lenders.

The dashed arrows show two arrangements that sit alongside the credit agreement. The loan proceeds are paid directly to the supplier, and the agent signs direct agreements with the customer and the data center on behalf of the lenders. The customer agrees to pay into the collateral account. The data center agrees to give the lenders access to the servers and not to block their removal if the lenders enforce.

Two Models: A Loan to the SPV or to a Holding Company

The first decision in structuring a deal is which entity will be the borrower. Publicly disclosed deals use one of two models, both shown below. The dotted line marks the ring-fenced project perimeter.

Model 1. A Loan Directly to the SPV

Here the borrower is the SPV that buys and owns the GPUs. That gives the lenders a direct claim against the owner of the equipment, and they take security over both the SPV’s assets and its equity.

Rating agencies find this structure easier to analyze and investors find it easier to buy, so it’s used for both bonds and syndicated loans. The drawback is that the SPV itself has to be a party to every key contract, and the lien on the GPUs usually has to be taken under the law of the country where the servers are located.

Model 2. A Loan to a Project Holding Company, Which On-Lends It

Here the lenders fund a holding company set up for the project, which on-lends the money to the company that buys the GPUs. The borrower is an intermediate holding company created for this one project, not the group’s top company, so the project is still ring-fenced.

This puts the lenders one step further from the assets, and they close the gap in three ways. The GPU-owning company guarantees the holding company’s loan, the rights under the intercompany loan are assigned to the lenders, and the borrower’s shares are pledged to them. In exchange, the lenders get a borrower in a jurisdiction they know, and if things go wrong, they can sell the holding company and everything under it in one go. Lenders tend to prefer this structure when the servers are in a jurisdiction with unfamiliar secured lending rules, currency controls, or a burdensome process for perfecting liens on equipment.

What Drives Loan Terms

In practice, the price and size of a loan turn mostly on two other factors.

One is recourse to the parent, meaning whether lenders can go after the group if project revenue isn’t enough to repay them. This varies widely. The parent may guarantee the whole debt, be liable only for its own failures (as cluster operator, say), or give nothing more than a bad-boy guarantee that covers misconduct. The less recourse lenders have, the more they care about the customer’s credit.

The other is how the borrowing limit is set. The loan may be capped at a percentage of equipment cost or tied to contracted revenue, and many credit agreements use both. The stronger the customer (off-taker) and the longer its contract, the more of the cost lenders will fund and the cheaper the loan.

Three Deals

We’ve put the deals in chronological order. Taken together, they show GPU-backed lending moving in three years from expensive private credit to bank loans and investment-grade bonds.

1. CoreWeave, Blackstone, and Magnetar: The First Multibillion-Dollar GPU-Backed Loan

In July 2023, CoreWeave was still private. It didn’t go public until March 2025. Microsoft was already its biggest customer, accounting for 35% of revenue in 2023 and 62% in 2024, according to the S-1.

Lending against computing equipment wasn’t new. Crypto miners, for one, had borrowed against their ASICs. What set the CoreWeave deal apart was its scale and structure. Nobody had borrowed billions against GPUs by putting the servers and the assigned customer contracts together in one ring-fenced company. The diagram below shows how it was set up.

CoreWeave, Inc. formed a subsidiary for the loan, CoreWeave Compute Acquisition Co. II, LLC, and became its sole member. CoreWeave signed the customer contracts itself and assigned its rights under them to the SPV before the first draw. From that point on, customers paid the SPV instead of CoreWeave.

The SPV signed the $2.3 billion credit agreement on July 30, 2023. It’s a delayed draw term loan (DDTL), meaning the money is drawn in tranches as equipment is bought rather than all at once. CoreWeave calls it DDTL 1.0, and it kicked off a whole series of loans. Funds managed by Blackstone and Magnetar led the deal, with Coatue, Carlyle, BlackRock, and PIMCO also participating. The SPV used the money to buy GPU servers from its suppliers.

Lenders got a lien on all of the SPV’s assets (servers, customer contracts, and bank accounts), and CoreWeave pledged 100% of the SPV’s equity. U.S. Bank Trust Company holds the collateral as administrative agent and collateral agent. CoreWeave also gave the agent a full payment guarantee of the loan, so the ring-fencing was mostly on paper.

The weak spot in the structure was the data centers. CoreWeave, Inc. had signed the leases for space in third-party facilities, but the servers inside belonged to the SPV, which had no agreement of its own with those facilities (the dashed line in the diagram). In a default, the lenders would have had to retrieve servers from sites where the SPV had no rights, and the operator could have refused them access or held on to the equipment over CoreWeave’s unpaid rent. So CoreWeave, as manager, agreed to try to secure access for the lenders and, where possible, to transfer the relevant portion of its leases to the SPV.

The most interesting part of the agreement is how the lenders handled the two big risks in a deal like this. Servers lose value fast, and customers can walk away and stop paying. A single mechanism covers both: the debt is always tied to what the servers are worth and how much customers still owe under their contracts.

When the SPV bought servers, it couldn’t borrow their full price, only a set percentage of their depreciated cost. If a batch cost $100 million and the advance rate was 80%, the SPV could draw no more than $80 million against it. Tranches like these were available until June 30, 2024, up to the $2.3 billion total.

After that, the SPV pays on a quarterly schedule, but before each payment it tests the debt against two limits. One is based on the revenue customers still owe under existing contracts, the other on the current value of the servers. If the debt is within both, the SPV simply makes the scheduled payment. Now suppose a customer terminates and projected revenue drops from $130 million to $85 million while the debt stays at $100 million. The SPV then has to prepay $15 million that same quarter. That way, at any given moment, the lenders can count on being repaid from contract revenue or from a sale of the servers.

Pricing was steep, as you’d expect for a new product. The loan bore interest at SOFR + 8.75%, which rose to SOFR + 9.62% in May 2024. The SPV also paid a 1.75% fee on amounts drawn and 1% a year on the undrawn commitment. The loan matures in March 2028, and $1.3 billion was outstanding as of June 30, 2026, according to CoreWeave’s filings.

CoreWeave’s later loans show how quickly the market moved. DDTL 3.0 in 2025 priced at SOFR + 4%. DDTL 4.0, an $8.5 billion loan in March 2026, was rated A3, priced at SOFR + 2.25%, and came with only a bad-boy guarantee from CoreWeave.

How cheap the money gets also depends on who the customer is. DDTL 5.0, a $3.1 billion loan that CoreWeave syndicated in the public market in May 2026, funded capacity for two customers without investment-grade ratings. It was rated Ba2 and BB+ and priced at SOFR + 4.5%. The $2.6 billion DDTL 5.5 in August priced at SOFR + 5.5%, well wide of its initial price talk of SOFR + 4.25–4.5%. The likely reason is a tenor mismatch. The loan runs about five years, but the customer contracts average three. Lenders bear the risk that customers won’t renew, and the final pricing reflected that.

2. Nscale and Macquarie: A Holding Company Loan Backed by a Contract with a ByteDance Subsidiary

Nscale, a British neocloud, was spun out of Arkon Energy’s crypto mining business in May 2024. Its first site was a former crypto mining data center in Glomfjord, Norway. The details of the Macquarie deal are public because Nscale filed for a U.S. IPO in September 2026 and included its loan documents in the prospectus. The diagram below shows the structure.

Nscale used the second model. The borrower is a holding company that sits above the GPU owner, and the chain has five companies in all. At the top is the English parent, Nscale Global Holdings. Under it is a Jersey holding company, Nscale Intermediate Holdings Limited, and then the borrower, an English company called Nscale Drift II Holdings Ltd. Below the borrower are a Norwegian intermediate company, Nscale Drift II Holdings AS, and finally Nscale Drift II AS, the Norwegian company that owns the equipment.

The customer contract came first. On May 15, 2025, the Norwegian project company signed an English-law cloud services agreement with Spring (SG) Pte. Ltd. covering the capacity of 2,304 NVIDIA B200 GPUs. The prospectus doesn’t mention any link between Spring and ByteDance. It says only that Nscale’s largest customer accounted for 73% of its 2025 revenue. According to press reports, that customer is Spring, ByteDance’s Singapore subsidiary.

A month later, Macquarie Bank’s London branch lent against that contract. Under an agreement dated June 12, 2025, it provided a $105 million facility, $28 million of which could be used for letters of credit. The borrower, Nscale Drift II Holdings Ltd, sits above the project company and on-lent the money to it through an intercompany loan. None of the cash actually passed through the group. Macquarie paid it directly to the systems integrator Computacenter, which built the servers from NVIDIA, Aivres, and Kaytus hardware and delivered them to the project company. The facility was fully drawn by January 30, 2026.

The equipment cost about $140 million all told. Nscale could borrow up to 75% of that and had to put in the other $35 million as equity before the first drawdown.

Both Norwegian companies guaranteed the loan, and the Jersey holding company pledged the borrower’s shares. Macquarie holds the collateral itself and plays every role on the lending side: lender, agent, collateral agent, and letter of credit issuer. There was also a second-lien lender, Sandton Capital, but its loan was repaid in October 2025.

Unlike CoreWeave, Nscale kept every other role inside the group, so server access was never a concern. The servers are in the Glomfjord data center, which is owned by another Nscale company and leases space to the project company under a colocation agreement. Nscale Operations provides management services.

The loan bears interest at SOFR + 7.5%, and the bank also received warrants for $6.3 million of Nscale stock, according to the prospectus. The more unusual feature is the cash sweep. On top of scheduled payments, Macquarie takes almost all of the project’s free cash flow, leaving Nscale no more than 10%. The sweep runs until the bank has collected the loan amount plus a preset premium. If the loan is repaid in the first year, that comes to 1.15 times the loan, or about $121 million on $105 million. The multiple then steps up and reaches 1.40, or about $147 million, after three years.

So the bank shares in the project’s income almost like an equity investor while keeping a lender’s priority. Macquarie appears to have been satisfied, and in July 2026 it repeated the structure in Iceland, lending an Icelandic Nscale company $331.9 million.

3. IREN and Microsoft: A Limited-Recourse SPV and A-Rated Notes

IREN is an Australian, Nasdaq-listed company moving from bitcoin mining into AI cloud, though mining still made up about 82% of its revenue in fiscal 2026. We covered its Microsoft contract in our second article. Here we focus on how it was financed. The diagram below shows the structure.

IREN used the first model, so the loan went straight to the SPV that owns the GPUs. Through its subsidiaries, IREN formed a project holding company, IE US Hardware 3 Holdings LLC, and under it a Delaware SPV, IE US Hardware 3 LLC. Another group company, IE US Development Holdings 3 Inc., runs the cluster and provides the data center for it under a management agreement.

The SPV signed the Microsoft contract, a Partner Statement of Work dated November 2, 2025, on its own, without the parent. Microsoft gets capacity on NVIDIA GB300 GPUs, which the SPV delivers in four tranches. Each tranche runs five years from go-live, and Microsoft prepays 20% of the value of each one. In total, Microsoft will pay up to $9.67 billion across all four tranches. If the SPV goes bankrupt, Microsoft can, with the lenders’ consent, step in and make the SPV’s loan payments itself. The step-in right protects both sides. Microsoft doesn’t lose its capacity, and the lenders keep getting paid.

To deliver on the contract, the SPV is buying servers from Dell for about $5.8 billion. To pay for them, it signed a two-part $3.6 billion financing on May 29, 2026. Banks led by Goldman Sachs and JPMorgan provided a $1.5 billion delayed draw term loan at SOFR + 2.25%. Investors bought $2.1 billion of 5.96% notes due December 31, 2031, sold in a U.S. private placement. IREN says it’s the first GPU-backed financing in the private placement market. Fitch and DBRS gave the financing ratings of A and A (low), respectively.

All the money went to the SPV, as borrower under the loan and issuer of the notes. The SPV pledged all of its assets and its cash flow under the Microsoft contract, and the holding company above it pledged 100% of the SPV’s equity. CSC Delaware Trust holds the collateral as administrative agent and collateral agent for both the banks and the noteholders.

IREN doesn’t guarantee the SPV’s debt. It’s on the hook to lenders in only two cases. One is if the group’s management company fails to perform under the management agreement. The other is if Microsoft refuses to accept or terminates a tranche and selling the GPUs for that tranche doesn’t cover the share of the debt allocated to it. It’s a performance guarantee, not a payment guarantee. The lenders take the risk that Microsoft won’t pay, which, given Microsoft’s credit rating, is a sensible bet.

The loan can be drawn in tranches until May 29, 2027, and until then the SPV has to keep its debt service coverage ratio (DSCR) at 1.05x or higher. A mandatory prepayment kicks in if the DSCR stays below 1.10x for six straight months, if the projected DSCR for the third and fourth tranches drops below 1.20x, or if debt goes above 65% of costs. As of June 30, 2026, the SPV had drawn $413 million of the loan and $525 million of the notes.

Microsoft’s prepayment drives the deal’s economics. Together with $1.94 billion in prepayments from Microsoft, the financing covers about 96% of the GPU costs under this contract, so IREN is putting up almost none of its own money. The debt costs about 6% on average. Counting the prepayment, which carries no interest, IREN puts the all-in cost of the financing at 3.31%.

What the Parties Negotiate

These deals were done in different countries with different lenders and customers, yet the agreements have a lot of terms in common.

The customer contract runs at least as long as the loan. Lenders want to know that revenue will pay off the entire debt. If the contract ends first, the rest of the loan term carries renewal risk, and lenders price that in, as they did with CoreWeave’s DDTL 5.5. At IREN, each Microsoft tranche runs five years, and the notes mature at the end of 2031.

Equity first, then the loan. In the deals we looked at, lenders fund no more than 65–75% of costs, and the borrower’s equity or the customer’s prepayment goes in first. Loan proceeds are paid straight to the supplier or integrator against transfer of title to the servers, in tranches as the equipment is delivered and accepted.

Customer prepayment. A prepayment from the customer (off-taker) cuts both the amount of debt and its cost. At IREN, Microsoft’s 20% prepayment plus the financing covered almost the entire cost. As we noted in our first article, most of Nebius’s new deals include prepayments too.

Direct agreements with the customer and the data center. Lenders need access to the servers and the right to remove them if they enforce. They also need the data center to waive any lien on the equipment for the neocloud’s debts, which is why the data center signs a direct agreement with them. The second article showed the same approach in Nscale’s contract with WhiteFiber, which requires the data center provider to sign agreements with Nscale’s lenders. The customer (off-taker) agrees to pay into the collateral account and not to terminate without notice to the lenders. At IREN, Microsoft also has a step-in right.

The right to replace the operator. In every deal we reviewed, the neocloud or one of its affiliates runs the cluster. If the operator fails to perform or goes bankrupt, lenders need to be able to bring in a new one without losing the customer contract. At IREN, the parent also guarantees its management company’s performance.

Coverage tests and cash sweeps. Lenders monitor whether revenue is enough to service the debt. The minimum DSCR is 1.05x at IREN and 1.35x on CoreWeave’s DDTL 5.5, and if the ratio falls too far, the loan has to be prepaid. Macquarie went a step further with Nscale, sweeping nearly all of the project’s free cash flow until it recovers the loan plus a preset premium.

Export controls. When the GPUs are outside the United States, lenders require representations that the borrower complies with U.S. export controls, and a breach usually lets them call the loan.

What Futures Will Change

Today, lenders mainly fund capacity that’s already sold under long-term contracts, and the loan terms show it. CoreWeave’s first loan is paid down whenever the debt exceeds projected contracted revenue. IREN’s has to be prepaid if projected coverage on future Microsoft tranches falls short. Uncontracted capacity and the years after a contract runs out either don’t get financed at all or cost more, as DDTL 5.5 showed.

As we explained in the third article, oil producers get around the same problem by hedging. The bank values the collateral against a conservative price deck, the producer has to hedge part of its output two to three years out, and the bank credits the hedged volumes at the hedge price. Once CME launches GPU rental futures, neoclouds will be able to do the same with uncontracted capacity and with the tail after a contract ends. Lenders could then finance hedged hours, not just signed contracts.

The way lenders value the hardware will probably change too. Credit agreements now value GPUs on a depreciation schedule, as CoreWeave’s first loan does. With a public GPU-hour price index, that value could be marked to market, the way banks periodically revalue oil producers’ reserves.

What Comes Next

In three years, GPU-backed lending has gone from an experiment for private credit funds to a mainstream tool for banks and bond investors, and the best deals now get investment-grade ratings. But the best terms go only to borrowers with a strong, highly rated customer (off-taker) behind the contract. Loans backed by unrated customers still cost noticeably more. That is because what lenders actually finance is the contract under which customers pay for the chips, so the price of the loan depends on who is on the other side of it.

This wraps up our series on neoclouds. We’ve looked at how the market emerged, how neoclouds sell capacity, how compute is turning into an exchange-traded commodity, and, finally, how neoclouds pay for it all.

The market’s next stage will probably start when neoclouds can hedge uncontracted capacity with futures, and when loan agreements and customer contracts begin to reference a compute price index.