GPU-backed debt in the capital markets: what gets rated

A single GPU cluster is financed the way any equipment is: a lender advances against it, takes security, and looks to a contract for repayment. That works until the amount needed is larger than any one lender will hold, and the buildout the market is now financing is far past that point.

When the ticket outgrows a balance sheet, the debt climbs a ladder that is well worn in other asset classes and new only in its collateral:

• A syndicated or club facility spreads one loan across several lenders.
• A delayed-draw term loan lets the money be committed once and drawn as the cluster is actually deployed, rather than funded in a lump before it can earn.
• A private placement sells the debt to institutional buyers — insurers, credit funds — outside the public markets.
• An asset-backed security pools the cash flows, divides them into tranches of different seniority, and sells those tranches to buyers who each want a different point on the risk curve.

Each rung reaches a larger pool of capital than the one below it, and each demands more standardisation, more disclosure and — at the top — a rating. The move up the ladder is not sophistication for its own sake. It is the only way to fund a buildout of this size, and it is why compute financing became a capital-markets subject rather than a bank-lending one.

The delayed-draw term loan is worth isolating, because it addresses the problem most specific to compute: the gap between committing capital and being able to earn on it.

A cluster is not built in a day and not deployed all at once. Accelerators arrive in tranches, get installed as power and space come available, and only begin generating contracted revenue once they are running. Funding the whole amount up front means paying to carry capital that is sitting in a warehouse or a half-built hall — on an asset that is losing value on the product cycle the entire time it is idle.

A delayed-draw term loan commits the full facility at signing but releases it in stages, against deployment milestones. The borrower draws as the hardware is installed and contracted, so the debt scales with the earning asset rather than running ahead of it. The lender, in turn, advances against equipment it can see deployed rather than against a purchase order.

The structure matters beyond its own efficiency. It is what makes the debt legible to the market above it: a facility that funds in step with deployment produces a cleaner, more predictable cash-flow profile, and a cleaner profile is what can be placed privately or securitized later. The draw mechanics are plumbing; the reason they recur is that they make the layer above possible.

The instinct on hearing "GPU-backed" is that the collateral is the hardware. It is not, or not primarily, and the distinction governs the whole structure.

What is pooled and tranched is the contracted cash flow — the payments a creditworthy counterparty is obliged to make for the compute, whether or not it consumes it. The accelerators sit underneath as recovery value, and recovery value on this asset is exactly the weak second line of defence the rest of the pillar describes: movable, hard to secure cleanly, and declining on someone else's roadmap. A structure that relied on the chips as its credit would price like a warehouse of depreciating inventory. A structure rated off the contract prices like the counterparty behind it.

The securitization mechanics are conventional once the collateral is understood correctly:

Two features do most of the work of turning compute cash flow into a rated instrument, and both are answers to the residual problem rather than to the credit one.

Over-collateralisation. The pool backing the notes is larger than the notes themselves, so the cash flow can absorb some contracts underperforming before the senior notes are touched. On an asset whose recovery value is uncertain, the cushion is sized against the cash flow, not the hardware — the equipment is not trusted to fill the gap.

Tranching. The cash flow is divided into layers of seniority. The senior tranche is paid first and is the one that can reach a high rating; the junior tranches absorb the first losses and are priced for it. This lets a single pool serve an insurer that can only hold investment-grade paper and a credit fund that is paid to hold the risk beneath it, out of the same deal.

The amortisation is where the tenor mismatch reappears in capital-markets form. Notes secured on accelerators cannot amortise on the thirty-year schedule of the infrastructure around them; they have to be repaid over the economic life of the hardware, which means the structure pays down fast and the schedule is set by the depreciation curve rather than by the borrower's preference. A structure that stretched the tenor to lower the coupon would be unsecured for its final years whatever the documents said.

The rating on a GPU-backed structure is, to a first approximation, the rating of the party paying for the compute — adjusted down for everything that can go wrong between the cash flow and the note.

That framing explains what moves the outcome:

• Counterparty credit. A contract with an investment-grade buyer that pays whether or not it uses the capacity is the foundation. The strongest structures are built on exactly that, which is why compute offtake is treated as credit rather than as demand.
• Contract enforceability and tenor. A rating cannot run longer than the contracted cash flow behind it. Termination rights, assignment, and what happens on the counterparty's own distress are read closely, because they are the paths by which the cash flow disappears before the notes are repaid.
• The residual, as a haircut. Because recovery value is uncertain, it enters the rating as a conservative floor, not as a source of repayment. The depreciation curve is why the senior tranche amortises quickly and why the whole structure is capped below the counterparty's own rating.
• Structural protection. Over-collateralisation, reserve accounts and the tranching above decide how much of the counterparty's credit actually reaches the senior note.

The reason this became a capital-markets story rather than a bank one is that debt structured this way has reached an investment-grade rating — rated off the take-or-pay obligation of an investment-grade buyer, and delinked from the borrower through a bankruptcy-remote issuer, rather than resting on the accelerators. (CoreWeave / Moody's, as of August 11, 2026) Where that contract is present, the debt behind it can be rated and sold to institutions that may hold only rated paper; where it is absent, the structure falls back to the bilateral, recovery-led terms the rest of this pillar describes.

The first rated structures proved that contracted compute receipts could carry debt when the customer obligation extended through maturity. The next step is more revealing: lenders have begun financing beyond the average term of the underlying customer contracts, which means the debt case contains an explicit renewal or reletting assumption.

CoreWeave's August 2026 DDTL 5.5 facility is the public example. The $2.6 billion facility carries an approximately five-year maturity against customer contracts averaging approximately three years, permits qualifying renewal or reletting after those contracts expire, and received Ba2 and BB+ ratings rather than the investment-grade ratings achieved by the earlier contract-matched transaction. (CoreWeave Form 8-K exhibit, as of September 20, 2026)

That comparison should not be reduced to "lenders now accept short contracts." They accept a different risk package at a different rating and price. During the initial contracts the structure is repaid from identified counterparties. After expiry, repayment depends on the operator renewing those customers or selling the released capacity to others under criteria written into the credit agreement. The uncovered period is financed merchant exposure, even if demand is expected to remain strong.

The distinction matters for market analysis because both facilities can be described as GPU-backed delayed-draw term loans. Their credit propositions are not the same. One isolates a firm payment stream through debt maturity. The other combines contracted cash flow with a forward view on GPU demand, platform competitiveness, customer replacement and remaining equipment utility. The lower rating is not a verdict on the asset class; it is the expression of additional assumptions between the last firm receipt and the last debt payment.

For sponsors, the lesson is constructive. Shorter enterprise contracts do not automatically make capital unavailable. They require faster amortisation, more coverage, a stronger platform, enforceable reletting criteria and pricing that pays lenders for the tail. For investors, the lesson is equally direct: the contract schedule, not the phrase GPU-backed, tells them whether they own customer credit or future compute-market risk.

Once debt extends beyond contracted revenue, the credit agreement needs rules for what may replace it. Otherwise a borrower can repair an expiring investment-grade contract with a longer but weaker one and claim the same collateral coverage.

A disciplined eligibility test addresses at least six points. Customer quality sets a rating, financial or concentration threshold for the replacement payer. Contract firmness requires a non-cancellable minimum rather than usage-only revenue. Term establishes how much of the remaining debt period the new contract must cover. Pricing prevents a nominally eligible contract from producing too little cash after power and operating cost. Technical fit confirms that the existing equipment and site can perform without a refresh that consumes the proceeds. Commencement timing limits the vacancy period between contracts.

Concentration tests matter differently in this phase. Several short enterprise contracts may diversify payer default while increasing rollover workload and correlated expiry risk. A pool in which half the capacity renews in one quarter is not diversified by customer count. The borrowing base should show capacity and revenue by payer, expiry month, equipment generation and site, then cap how much exposure can sit in any one cohort.

Cash control also changes as expiry approaches. Excess cash that could be distributed safely while long contracts remain may need to sweep debt, fund a renewal reserve or pay for reconfiguration once the tail begins. A forward-looking test can trap cash before coverage fails rather than after. The trigger should use contracted revenue remaining after each future date, not trailing revenue earned under a contract about to end.

Finally, reletting must be operationally possible. The lender needs continuing rights under the hosting agreement, a usable power allocation, access to equipment and enough time to qualify a new operator or customer. A credit agreement can define an eligible contract perfectly and still fail if the site agreement ends first. This is why the capital-markets structure inherits every seam below the compute layer.

A rating is a structured conclusion, not a substitute for understanding what was structured. Three reconciliations expose most of the difference between two GPU-backed instruments.

The contract-to-debt reconciliation compares minimum customer receipts, net of service credits and operating costs, with interest, scheduled amortisation and required reserves by month. It identifies how much debt is repaid before each contract expires and the balance left for renewal, reletting or residual recovery.

The asset-to-contract reconciliation maps equipment, capacity and location to the customer obligation that uses them. It should be possible to identify which serialised systems support each contract, whether capacity is dedicated or shared, and which assets remain available when one customer leaves. Double-allocation of the same capacity is the physical version of counting revenue twice.

The entity-to-cash reconciliation follows receipts from the customer through operating accounts, tax and site costs, reserves and the issuer waterfall. It identifies commingling, set-off, affiliate claims and any point at which cash can sit outside the bankruptcy-remote vehicle. A perfect take-or-pay contract owned by the wrong entity is not the same collateral.

Investors should then read the downside assumptions that sit outside the rating headline: permitted equipment substitution, refresh funding, customer-concentration limits, expiry cohorts, power pass-throughs, insurance, hosting cure rights and the valuation basis for residual recovery. None necessarily weakens the structure. They specify the risks for which the spread is paid.

The result is a more useful classification than bilateral versus securitised or investment grade versus high yield. A GPU-backed instrument can be primarily contract credit, platform and renewal credit, or asset recovery credit. Most are mixtures. Knowing the proportions is the prerequisite to comparing them.

One question sits underneath all of this and is genuinely unsettled: whether a given data-center or compute securitization is an asset-backed security as the securities laws define it, or something outside that definition.

The distinction is not academic. Whether a structure falls inside the asset-backed regime changes the disclosure it must carry, the ongoing reporting it owes, and the liability its sponsors take on. A structure treated as outside that regime can be brought to market faster and with lighter continuing obligations — which is why the treatment is fought over, and why regulatory guidance on it moves the market. Staff have taken the position that certain data-center securitizations — those where the issuer owns the facility itself and pays noteholders out of its operating income — fall outside the asset-backed definition, because a building whose repayment depends on being operated is not the pool of self-liquidating financial assets the regime was written for. (SEC Division of Corporation Finance, as of August 11, 2026)

This is securities-law territory, and it belongs to counsel on any specific structure — the analysis turns on the exact assets, the exact cash flows and the exact rights being sold, and it is precisely the kind of confident-sounding general statement that is wrong in the particular case. What is durable, and what this page will assert, is only the shape of the question: the same pool of compute cash flow can be structured on either side of that line, the two sides carry materially different obligations, and which side a structure sits on is decided before it is sold, not after.