A bank can offer the same spread on a GPU loan and still cut the amount it will lend. For a cloud operator waiting to order its next cluster, that second number matters more. A higher spread costs money over five years. A smaller advance means finding more equity before the equipment ships.
That is where we would look first for the financial effect of Jensen Huang’s departure. NVIDIA’s board accepted his resignation as president, chief executive and director on November 11; the company announced it on November 12. Colette Kress, its CFO, took over as interim CEO. The board expects customer commitments and the product roadmap to continue. A lender financing the next build has a narrower question: which customer can already deliver an NVIDIA signature, and which still needs one? Succession filing · Board statement.
The price of a second site
Consider two pieces of the same Ohio campus. At PORTS, where an OpenAI affiliate is the tenant, NVIDIA signed guarantees covering leases for about 4.25 GW of capacity. For another 3.8 GW, NVIDIA retained the right to decide whether to provide support. An underwriter can model the first phase from a signed instrument. To rely on NVIDIA for the expansion, the developer must win another credit decision. Its other choices are a stronger tenant commitment, more sponsor equity or a smaller build. PORTS 8-K, Item 1.01.
Who gets paid in the first phase? The landlord. The guarantees generally take effect when the relevant leases commence. After specified tenant defaults, a replacement lease or sale can recover part of the guaranteed minimum; NVIDIA generally covers the remaining shortfall, subject to a $105bn cumulative cap. OpenAI agreed to reimburse what NVIDIA actually pays. Timing matters: a landlord can have a valid claim on NVIDIA while NVIDIA is still collecting from OpenAI. Treating the cap as immediate spending would miss the real credit exposure: which parties need liquidity at the same moment.
That is also the dilemma for NVIDIA’s own creditors. As of July 26, it had $56.586bn in cash, equivalents and marketable debt securities; it generated $74.421bn in operating cash over the preceding six months. Declining a new guarantee preserves financial capacity. But if the customer needs that guarantee to get power, buy equipment and accept delivery, saying no can delay NVIDIA’s own receipts. The investment decision is whether the chip sales that support unlocks are worth the correlated risk NVIDIA takes back. NVIDIA quarterly filing, liquidity and Notes 8 and 10.
Two empty racks, two very different claims
One CoreWeave rack is ready for service but lacks a buyer. Another is in a facility that cannot deliver the agreed service. The revenue forecast looks bad in both cases. The lender should care which rack it owns a claim on.
Under a September 2025 order with an initial value of $6.3bn, NVIDIA agreed to buy eligible residual unsold CoreWeave capacity through April 13, 2032, subject to service availability, delivery and termination terms. The first rack illustrates the demand risk the order absorbs. The second has yet to clear the delivery condition. Orders like this make a commissioned asset easier to finance precisely because they allocate one identifiable risk to a counterparty that can use the capacity. Order disclosure.
Now turn to CoreWeave’s $2.6bn DDTL 5.5, a delayed-draw term loan to a subsidiary. The parent that unconditionally guarantees all obligations is CoreWeave, Inc. The facility also has subsidiary guarantees and collateral. NVIDIA’s separate purchase order is a commercial arrangement; the reviewed filings do not identify it as security for this particular loan. This distinction tells the lender whom to call if debt service comes up short. Facility 8-K.
The quarter-end debt table puts $1.555bn against that facility at September 30, or 59.8% of its size. The arithmetic difference of $1.045bn becomes usable only to the extent the borrower still meets the draw conditions. The stated draw window runs through December 31, 2026. Monthly principal payments begin in February 2027; after the commitment period, excess cash must help pay down the loan until outstanding principal falls to $1.3bn. This is how equipment spending starts a repayment clock before the operator knows the price of its next customer contract. September 10-Q, debt note.
When a renewal turns into a cash call
The financing buys roughly five years against customer contracts averaging about three years at inception. That mismatch is a commercial choice: shorter commitments give CoreWeave a chance to renew at better prices and reach customers unwilling to sign for five years. Lenders accepted some renewal risk and wrote a way to resize their exposure if the next contract disappoints. Facility announcement.
Under section 2.09(b)(iii), if the qualifying renewal or replacement conditions are unmet by the applicable deadline, the borrower must prepay within five business days after that deadline, to the extent needed to restore both tests: at least 1.35× projected debt-service coverage in each relevant month and 2.40× projected contract value to outstanding principal. A written commitment can count on an interim basis under specified conditions; definitive documentation still matters. The public agreement redacts the qualifying-value definition and the renewal-deadline interval. Executed credit agreement.
Take two hypothetical projects, each with 100 of principal. The first opens at the threshold: 240 of qualifying value. A 10% cut leaves 216, enough to support 90 of principal. If the provision applies, 10 must be repaid under the value test. Give the second project 264 of starting value. The same cut leaves 237.6, enough to support 99. Its paydown is just 1. The second sponsor can absorb about a 9.09% decline before the value test bites. The lender that negotiated the cushion years earlier has changed the sponsor’s financing problem on renewal day.
Try it: in Excel or Google Sheets set A2 to opening principal, B2 to initial qualifying contract value in the same units, and C2 to the percentage decline. Enter all three nonnegative inputs (C2 between 0% and 100%). Put this in D2:
=MAX(0,A2-B2*(1-C2)/2.4)
This isolates the value test and holds the post-shock numerator fixed. The monthly coverage test may require more. Use the editable workbook to change the three inputs or take the six worked cases as JSON.
The next cheque is the test
Nebius shows what a third type of support looks like. In March, it issued NVIDIA a pre-funded share warrant for approximately $2bn gross. That capital could buy equipment or fund a build. A lender assessing Nebius today needs to know what the company did with the proceeds; a warrant issued in March gives it no claim on a second investment. Nebius 6-K · Nebius’s November funding plan.
At the November 17 NVIDIA earnings call, the revealing words will concern new approvals. After that, compare financings of similar borrower quality, collateral, seniority and customer tenure. How much will the lender advance? How much cash does the sponsor put up first? What price and term must the replacement customer accept? CoreWeave also reported earnings on November 10; a market price over this window reflects more than one event.
Our call is on the quantity of credit. If sponsors retain the same leverage on similar terms without additional NVIDIA support, the financing model has survived the handover. If sponsors must contribute more equity or secure longer customer commitments to build the same cluster, the cost will show up in a smaller project or a later delivery—even if the loan spread barely moves. For the next sponsor, an eligible replacement contract and a little more room at the opening leverage ratio may be worth more than winning the last quarter-point on price.