Compute offtake as credit: contracted capacity vs pipeline

On assets with dependable resale, recovery does real work — a lender can accept moderate cash-flow uncertainty because the collateral is a genuine second line of defence.

Compute does not offer that. The equipment is movable, the value moves on someone else's product cycle, and distress is correlated across holders. Recovery is therefore a weak backstop, and the consequence is direct: the contract has to carry what the collateral cannot.

This is why a cluster with strong hardware and indicative demand supports so much less than sponsors expect, and why a cluster with unremarkable hardware and a firm contract from a solid counterparty supports so much more. The asymmetry is not conservatism. It reflects where recovery actually comes from.

The first test is what the counterparty is actually obliged to do.

Committed capacity, paid regardless of use. The customer pays for reserved capacity whether or not it consumes it. This is the structure that functions as credit, because the payment obligation does not depend on the customer's own demand materialising.

Minimum commitment with consumption above it. A firm floor with variable upside. The floor is creditworthy; the upside is not, and should not be financed as though it were.

Consumption-based with no minimum. The customer pays for what it uses and is obliged to use nothing. This is a commercial relationship, not a contracted cash flow, however large the expected volumes.

Framework agreement. Terms under which orders may be placed, with no obligation to place any. This is a sales channel.

All four are legitimate commercial arrangements and all four are routinely described as offtake. Only the first two support capital in a meaningful way, and the distinction is the first thing tested in any serious review.

The second test is duration, measured against the capital rather than in the abstract.

A contract shorter than the period over which capital is repaid leaves an uncovered tail. Someone carries the risk that the capacity is recontracted at all, and at what price. That is a real position and it should be identified rather than assumed away with a renewal probability.

Three things make the tail more tolerable:

• Renewal economics that favour the customer. A counterparty that has integrated the capacity into its own operations, and would face switching cost and disruption to move, is more likely to renew than one that treats capacity as fungible.
• Redeployability. Capacity that can be recontracted to other customers without relocating equipment has a genuine alternative. Capacity purpose-built for one counterparty does not.
• Amortisation inside the contracted term. The cleanest answer to a tail is not to have one. Structures sized to repay within the firm period avoid the question entirely, at the cost of a heavier profile.

Where the contract is materially shorter than the equipment's economic life, the mismatch is the structuring problem rather than a detail of it — see tenor mismatch.

The third test is whether the obligation is worth having.

A contract is only as good as the party behind it in the conditions where it matters, which is not the conditions in which it was signed. The questions are ordinary credit questions, and they are frequently skipped because the counterparty is well known:

• Can it pay from operations, or does payment depend on continued fundraising? A well-capitalised company that is not yet self-funding is a different credit from a profitable one, whatever its profile.
• Is the contracting entity the entity with the resources? A thinly capitalised subsidiary signing on behalf of a substantial group is a common structure and a common surprise. Where the parent is the real credit, that should be documented rather than assumed.
• What are the termination and suspension rights? A firm-looking commitment with broad termination for convenience is not firm. Change-of-control, force-majeure and service-level provisions all deserve reading for the same reason.
• How correlated is the counterparty with the sector? A customer whose own demand collapses in the same conditions that depress compute prices provides less protection than its standalone credit suggests.

Counterparty prestige is not a substitute for any of this. Some of the least enforceable arrangements in the market carry the best-known names.

Yes, but the word commitment has to describe an enforceable payment floor rather than a forecast of usage. A multi-year agreement can support debt when its contracted receipts remain available through the lender's repayment period and when the lender can rely on the payer and the contract if the borrower encounters trouble. The exercise is not to assign value to the headline contract amount. It is to identify the portion of that amount that survives credit, legal and operating stress.

A lender typically works through six connected questions:

• What is the firm minimum? The underwritten base is the amount due without regard to actual consumption. Burstable usage, non-binding reservations, forecasted expansion and renewal options can support a commercial case, but not scheduled debt service. Credits, performance rebates and service-level penalties should be deducted from the floor where they can reduce cash receipts.
• Who owes it? The payer's financial capacity matters more than the customer group's brand. If a special-purpose subsidiary signs the agreement, the analysis turns to its assets, funding obligations and any parent guarantee, letter of credit, deposit or other credit support. Support should cover the same obligations and survive for the period the lender is relying on it.
• Does the term fit the amortisation? Contracted receipts should be compared month by month with operating costs, taxes, reserves and debt service. Debt that repays inside the non-cancellable term is structurally different from debt that assumes renewal or recontracting for its final instalments. The latter may still be financeable, but the uncovered balance is exposure to the future market rather than contracted credit.
• Can the customer stop paying? Conditions precedent, acceptance tests, delivery milestones, service levels, outage credits, termination rights and force-majeure provisions can turn a stated multi-year term into a conditional one. A lender will test the earliest date and the most plausible event at which the payment obligation can fall away, not only the date printed on the cover page.
• Can the contract travel with the financing? Assignment restrictions, confidentiality provisions and anti-transfer clauses determine whether the lender can take security over receivables, receive notices and preserve the contract after enforcement. Where the contract is central to repayment, a consent or direct agreement may address notice of default, cure periods, payment directions and the customer's treatment of a replacement operator. The exact form depends on the transaction and the governing documents; it cannot be inferred from a general permission to assign.
• What remains if this customer leaves? A single large customer can make an asset financeable and simultaneously create concentration risk. Underwriting should show the exposure by payer, contract expiry and capacity block, then test how quickly the released capacity could be recontracted, at what price, and with what reconfiguration cost. A pipeline is useful evidence for that downside case, but it is not a substitute for the existing customer's firm minimum.

These tests also explain why two agreements with the same stated value can support different amounts of debt. One may produce a durable, assignable payment stream from a creditworthy payer and amortise the financing before expiry. The other may depend on usage, carry broad termination rights and leave most principal outstanding when the term ends. The commercial totals match; the financeable cash flows do not.

The clearest public evidence that compute offtake can function as credit is a financing in which the debt result was driven by the customer contract and the structure around it, rather than by a generic claim on GPUs. In March 2026 CoreWeave closed an $8.5 billion delayed-draw term loan that received investment-grade ratings, with the disclosed credit case resting on a take-or-pay obligation from an investment-grade offtaker and a bankruptcy-remote issuer. (CoreWeave / Moody's, as of August 11, 2026)

That example should not be read as a market-wide advance rate or as proof that any named-customer agreement can produce the same result. It establishes a narrower and more useful point: when the payer, payment floor, term and legal structure are strong enough, compute cash flow can be separated from the operator's general corporate risk and financed on its own merits. The hardware remains part of the collateral, but the payment contract supplies the predictable cash flow and the bankruptcy-remote structure protects the path from that cash flow to the debt.

It also shows why the contract review cannot stop at commercial value. A financing party needs to know which capacity is dedicated to the contract, when the payment obligation starts, what happens if delivery is late, how service credits affect receipts, where cash is paid, and whether the contract remains in place if the operator defaults under its financing. Those are not secondary legal points. Together they determine whether the lender is underwriting a receivable, a performance-dependent operating forecast, or merely the borrower's expectation that a valuable relationship will continue.

The case is a precedent for the proposition, not a template. Different contracts, customers, equipment generations and structural protections can produce different ratings and debt capacity. The capital-markets structure built around GPU-backed debt is addressed separately; here the lesson is simply that the contract earned its place as the primary credit because its terms survived the tests on this page.

A large signed contract can be commercially transformative and still leave a project without the money required to perform it. The distinction is easiest to see by scheduling three quantities separately.

Contracted value is the sum the customer may pay if every tranche is delivered, accepted, available and used under the agreement's payment mechanics. It is a legal and commercial figure, usually stated over several years.

Accepted capacity is the portion that has passed the conditions required for payments to begin. Before acceptance, the provider may carry equipment deposits, construction cost and operating preparation without a current receivable from the customer.

Funded capacity is the portion for which committed capital is available through equity, equipment finance, a construction facility or another binding source. A contract may help obtain that capital, but it is not itself cash for the equipment unless the customer prepays.

Nscale's preliminary September 2026 prospectus provides a clear public example of the gap. It described Anthropic service agreements with aggregate potential payments of up to approximately $44.6 billion, subject to delivery and service-availability requirements, while stating that binding commitments had not yet been obtained for the financing required to perform them. (Nscale preliminary Form S-1, as of September 20, 2026) The disclosure does not make the agreements weak. It makes the funding condition visible.

A lender or arranger should therefore build a tranche schedule that places, on the same row, equipment order, deposit, delivery, site readiness, installation, acceptance, customer payment commencement and committed financing. The schedule exposes a period that headline TCV hides: capital must leave before contracted revenue can begin. It also shows whether one failed tranche delays only itself or allows the customer to terminate later tranches.

This is the point at which offtake becomes useful but not sufficient. A strong contract can support the financing; an executed financing allows the provider to perform the contract; accepted delivery starts the cash flow. Describing any one of the three as though it proves the other two is how a pipeline becomes mistaken for a funded asset.

The cleanest compute financings repay inside the firm customer term. That remains the benchmark because it eliminates a future sales assumption. It is not the only structure the market will fund.

CoreWeave's DDTL 5.5 facility illustrates the alternative: an approximately five-year debt maturity against underlying customer contracts averaging approximately three years, with qualifying renewal or reletting available after the initial contracts expire. (CoreWeave Form 8-K exhibit, as of September 20, 2026) The facility therefore does not pretend the tail is contracted. It identifies the tail and finances it under defined conditions, at a rating and price that reflect the additional risk.

The underwriting question changes after the last firm payment. Before expiry, the lender asks whether the named customer will pay and whether the operator will perform. After expiry, it asks whether the platform can retain the customer or replace it, how long that takes, what price the capacity earns, what refresh or configuration cost is required, and whether the equipment remains competitive. The credit shifts from a payer to a market and operator.

That shift should be visible in the repayment profile. A structure can amortise aggressively during the contract, leave a smaller tail sized to conservative reletting value, sweep excess cash as expiry approaches and require replacement contracts to meet eligibility tests. It can reserve for downtime and refresh rather than assume an immediate rollover. What it cannot do honestly is capitalise renewal probability into the original contract value and still call the whole balance contracted debt.

For a sponsor, this creates a useful choice. A longer customer term may support cheaper or larger debt but can require lower pricing and less commercial flexibility. Shorter contracts may command better margins and reach a wider customer base, while leaving more renewal risk with the operator and its lenders. The correct structure prices that trade explicitly. Offtake remains credit during the period in which it is owed; beyond that period, the platform has to earn the rest.

The most reliable way to prevent a headline contract value from doing too much work is to translate the agreement into a monthly schedule. Start with the firm payment due in each period, then subtract every contractual path by which cash can be delayed or reduced: implementation periods, ramp allowances, service credits, performance abatements, disputed amounts, termination windows and taxes or pass-through costs the operator must bear. The remainder is not automatically debt service; operating costs, reserves and required reinvestment still come ahead of it.

A useful schedule runs at least three cases. The contract case includes only non-cancellable minimum receipts and the costs required to deliver them. The performance case applies plausible service-level credits, outages and a slower ramp. The loss-of-customer case assumes the earliest credible termination, then adds downtime, reconfiguration expense and a defensible recontracting price. Upside usage and renewals can be shown separately, but they should not quietly repair a shortfall in the contracted case.

The exercise often reveals that the apparent tenor is longer than the effective tenor. A five-year agreement with a delivery ramp, a customer termination right after year three and a six-month notice period may provide materially less than five years of dependable receipts. Conversely, a shorter contract with a hard minimum, strong credit support and fast amortisation may carry more debt because fewer assumptions sit between the payment obligation and final repayment.

The output should reconcile to the physical deployment. Contracted units, reserved capacity, power allocation and equipment serials or clusters should describe the same asset in compatible units. If the commercial schedule says one thing and the deployment schedule says another, the lender cannot tell which equipment earns which receipt or what remains available to another customer. That ambiguity matters most during enforcement, which is exactly when there is no time to solve it.

A compute contract is reviewed as a document, not as a summary. What shortens the process:

• The contract itself, including schedules and any side letters. Payment mechanics, termination rights and service levels are where the substance sits, and they are exactly what summaries omit.
• The contracting entity's position — who is actually obliged, and what stands behind it.
• The capacity being sold, stated so it can be reconciled with the equipment schedule. Contracts written in units that do not map cleanly onto the deployed configuration create ambiguity nobody benefits from.
• Pricing mechanics, including escalation, indexation and any pass-through of power cost. Where power is passed through, the site position becomes part of the contract's credit.
• An honest statement of what is not contracted. Pipeline described as pipeline costs nothing. Pipeline described as offtake and discovered in diligence costs the transaction.

Continuum structures and arranges around these contracts; it does not lend, take capacity positions, or hold client funds.