After GPUs are sold, who still carries capital for AI compute?

Compute expansion is often drawn as a straight product chain: supplier sells GPUs, cloud sells compute, users buy services. It does not identify who pays first during construction, when collection occurs or who carries unsold-capacity obligations. Revenue ownership and contractual capital responsibility are different questions.

The NVIDIA case supplies two verified mechanisms: 90-day to one-year terms for certain large investment-grade equipment buyers, and separate agreements to purchase committed capacity not sold externally. One extends collection waiting; the other can make the supplier a buyer. The shovel-seller label does not resolve either.

This is a case-based explanation, not an industry financing estimate or a claim about every supplier. One disclosure establishes a possible structure, not prevalence. Wider conclusions require multiple parties, matched projects and actual execution.

Extended terms leave the supplier carrying the interval after delivery. Buyers may build or deploy before paying; actual deployment and repayment from operations still require buyer records. Contractual permission to wait is not realized end demand.

Receivables rose 54.9% sequentially against 17.9% revenue growth and absorbed $22.346 billion of cash. These company values establish capital absorption, not the extended-term subset. Management’s attribution covers first-half year-over-year growth, not every quarterly receivable.

Trade credit differs from credit loss. Even payment at maturity leaves a period the supplier financed; overdue invoices and unrecoverability require further evidence. Who waits can be established before who loses.

A separate agreement operates at compute-service sales: NVIDIA purchases residual committed capacity if the cloud cannot sell it externally. This concerns downstream demand, not collection of the same equipment invoice. Some contracts can retain the supplier’s exposure after hardware sales.

Signing is established; quarterly triggering, payment and outcome are not. Purchased capacity may be used or resold, or fail to recover costs if demand, pricing and operational capability are insufficient. Loss depends on those outcomes, not purchase alone.

Equipment customers are not identified as the same AI clouds. A loop from supplier-financed GPUs to purchases of the same customer’s unsold compute is therefore unsupported. Track the responsibilities separately before combining them within verified transactions.

Advances reverse the funding direction: the customer pays first and NVIDIA fulfills later. Advance balances rose from $0.160 billion to about $2.8 billion across the half-year boundaries. This blocks the generalization that suppliers finance every customer, without offsetting quarterly receivable flows or identifying identical counterparties.

Company payment timing can also be wrongly absorbed into a financing story. Other working capital worsened $12.489 billion alongside receivables, and federal-payment timing differed. Tax and liability movements constrain a customer-only narrative; they are not additive supplier-financing volume.

These counterexamples do not erase responsibility. Advances, normal collections and economically valuable capacity purchases can weaken an excessively negative account; longer waiting and conditional obligations weaken the claim that supplier risk always ends at sale.

The next evidence concerns counterparties, matched projects, execution windows, actual payments, who uses or buys capacity, guarantees and cancellation rights. Those links are needed before discussing chain-wide capital recovery and allocation of risk.

Project announcements and news may identify a transaction, but contracts, revenue and funding still require execution records. A construction plan is not deployment; investment is not operating revenue; two company totals are not one transaction. Who funds compute expansion ultimately depends on actual payments, not project lists.

The case yields a clear mechanism: supplier roles depend on contracts as well as products. Waiting for customers, potentially buying capacity elsewhere and receiving customer advances can coexist. They are not a verified loop. New coverage and execution evidence is needed to judge expansion of the mechanism; applying it elsewhere requires that company’s transactions.