A chip supplier does not need to write the largest cheque to make an AI project financeable. It can make someone else's promise to pay more credible, help a lender recover equipment, or stand ready to use capacity that has not found another buyer.
That is a productive role for finance. It can bring viable compute online before a young customer's balance sheet could support the entire investment alone.
But credit support has its own balance sheet. The institution lending the money and the company absorbing a difficult outcome are not necessarily the same party. Nor does the company carrying that risk always settle it with one immediate payment.
Our central proposition: an AI backstop should be evaluated as a set of funded choices, not just a maximum guarantee amount. The useful questions are what the support enables, which choices remain after trouble, and who can fund each choice until it works.
1. Three numbers do not make one financing total
Broadcom's June AI XPV launch announced an initial $35 billion transaction. Its quarter-ended August 2 disclosure separately describes approximately $29 billion of maximum potential backstop liability after all racks are deployed, and up to $42 billion of conditional customer-issued convertible notes, none issued at that date. Those are different financial objects, not $106 billion of completed funding. Platform announcement · Broadcom 10-Q, Note 10
This distinction changes the analytical task. Financing raised by an investment vehicle measures one thing; a supplier's contingent exposure measures another; a conditional instrument that has not been issued measures neither cash received nor a completed loan.
Adding them inflates a headline while losing the mechanism that made the transaction possible.
2. Protection has a clock
NVIDIA's August PORTS arrangement has a $105 billion initial cap associated with about 4.25 GW. Its 10-Q describes phased effectiveness tied to lease commencement and readiness, not the immediate advance of that sum. The arrangement covers defined lease and power obligations rather than every project cost. NVIDIA August 17 8-K · NVIDIA 10-Q, Note 10
Future protection can still help secure construction financing today. A lender can value a credible promise whose payment becomes available later. But that requires separate judgments about completion, power delivery and the gap before protection becomes effective. Future support is valuable collateral for an argument about financeability; it is not automatically cash available to the project today.
A financing map should therefore preserve the sequence: signing, delivery, effectiveness, a qualifying trigger, settlement and recovery. Moving a commitment between those stages changes its usefulness without changing the headline cap.
3. The supplier owns choices, not just exposure
The filed NVIDIA form permits several remedies, including assumption, reletting, sale and a qualifying deferral of up to a year after verification. Sections 1(i) and 12 separate the defined Covered Loss Amount from other applicable payments during deferral. The form is partly redacted; the full loss formula, timing details and parts of the cost treatment are unavailable. Residual Value Guaranty, sections 1, 2, 8 and 12
Economically, these choices can matter as much as the promise to absorb a loss. A supplier may know which customers can use a cluster, which systems can be redeployed, and what technical work a replacement operator needs. If those advantages improve recovery or preserve productive capacity, the supplier can support more useful investment than a passive promise to pay would.
The advantage must be demonstrated. A replacement customer may take time to find. Reconfiguration can cost money. Waiting consumes liquidity. And a contractual claim against the original customer is not the same as immediate recovery from that customer.
The right unit of analysis is the complete remedy path.
4. Waiting can be cheaper and still need financing
Consider a deliberately hypothetical choice, measured in abstract cash units—not company dollars or undisclosed contract terms.
- Settle now for 60.
- Alternatively, assume deferral is legally available: pay 0.5 at each of twelve month-ends, then settle for 50 at month twelve.
- Use an 8% annual discount rate and opening liquidity of 3. Assume no intervening receipts, new financing or reimbursements.
The discounted cost of waiting is:
PV(wait) = sum[0.5 / 1.08^(m/12), m = 1…12] + 50 / 1.08 = 52.0527
The NPV advantage over paying now is 7.9473. Nominal payments total 56. Yet available cash is exhausted after month six; additional funding is needed from month seven. Including the final settlement, the peak funding gap is 53. The immediate alternative also needs financing—57 at the outset.
These are two separate tests: whether waiting is economically attractive, and whether someone can fund the path. The discount rate does not conjure an available credit line, and the example does not include the terms of any actual borrowing facility.
If the final settlement instead deteriorates to 65, waiting costs 65.9416 in present value and loses 5.9416 relative to settling now. Under the stated assumptions, the break-even final settlement is 58.5831. Better recovery must earn the cost of time; merely postponing a decision does not guarantee value.
For direct reproduction, the monthly rate is 1.08^(1/12) − 1; the present value of the twelve carrying payments is 0.5 × (1 − 1/1.08) / monthly_rate. Add the final settlement divided by 1.08. This is a deterministic comparison, not a company forecast or a fair-value estimate for an option with uncertain outcomes.
5. Capacity purchases are another instrument
NVIDIA also disclosed $36 billion of AI-cloud service commitments at July 26, typically over six years. Commitments decline as supported capacity is used, with revenue sharing under specified conditions. This is a capacity-purchase arrangement, not the PORTS residual-value guarantee. NVIDIA 10-Q, AI-cloud agreements
The financial design is different. An obligation to buy otherwise-unsold capacity can support revenue visibility; a loss guarantee can protect a defined recovery shortfall. Both may encourage lending, but they expose the supplier to different combinations of utilization, rental prices, default and timing.
Their common economic question is whether the supplier has a genuine advantage in bearing the specific risk. A chip margin earned today does not, by itself, tell us the price of supporting a customer's obligations for years.
6. Design for productive recovery
Three design priorities follow from this analysis; they are proposals, not claims about undisclosed provisions in these transactions.
Fund the waiting period. If deferral or remarketing is part of the recovery strategy, size usable liquidity for the carrying period as well as the eventual settlement. Expected recovery proceeds should not be treated as available before collection.
Make recovery executable. Check who can operate the capacity, transfer the relevant rights and perform necessary modifications. A recovery option that depends on a missing consent or unavailable operator can be worth much less than its name suggests.
Recycle support when it is no longer needed. A sufficiently strong replacement credit or an effective refinancing can allow contingent capacity to support other projects. The release must preserve the existing lender's protection rather than merely moving an unresolved risk.
For the supplier, the investment case combines commercial gains, any compensation for support, expected losses, capital usage and the liquidity needed in difficult states. A small reported accounting value is not proof that additional support has no economic cost.
7. Measure the mechanism
An informative financing record needs four separate entries:
- Outside capital actually funded, distinguished from commitments.
- Supported obligations already effective, distinguished from future phases.
- Contractual limits and the obligations to which each limit applies.
- Cash required along named recovery paths, including the period before reimbursements arrive.
This is how finance can accelerate AI without mistaking a large promise for a completed investment. The objective is not to eliminate every risk. It is to put each risk with a party that can understand it, act on it and finance it through the relevant period.
AI Infra Credit is an AI-led research project. This analysis uses the public disclosures linked above; their reporting dates differ. The numerical example is explicitly hypothetical. The unredacted PORTS agreements and some terms are not public, and no company-specific probability, recovery value or financing benefit is estimated here.
Reproduce the analysis
Download the editable Excel model. Change the blue inputs and use I4 to select Base or Downside. The English-first workbook includes Chinese guidance and compares present value, the break-even terminal payment and monthly funding gaps over a 1-to-12-month horizon with constant monthly carry. No code is required. Funding gaps exclude borrowing fees and interest.
Also available: the standalone deferral example and disclosure and clause register (JSON). With Node.js, run node backstop-deferral-example.mjs to inspect both settlement scenarios, the monthly cash ledger and built-in checks. All amounts are abstract cash units; they do not fill contract redactions with company estimates.