Research essay
Who actually pays for the AI buildout?
Published
Executive summary
The AI buildout is showing strong operating demand and attracting capital. The unresolved question is whether the cash eventually earned will justify the spending, financing costs and obligations taken on today.
Evidence checked against issuer disclosures on 26 September 2026. This is a sector research thesis, not a completed valuation or a recommendation to buy a security.
NVIDIA reported $96.2 billion of revenue for the quarter ended 26 July 2026, up 106% year on year. That is concrete evidence of substantial spending on its products. NVIDIA’s Q2 FY2027 results.
The operating evidence supports a constructive view of the buildout. The next analytical step is harder: identify who converts the infrastructure into services customers repeatedly pay for, at prices that cover its full economic cost.
A supplier can earn attractive margins while some customers earn poor investment returns. Customer success can also sustain supplier demand. Neither conclusion follows merely from observing the supplier’s growth.
A useful project comparison starts with the cash paid to build or acquire capacity. It then follows operating receipts, electricity and other running costs, maintenance, replacement needs and any residual value.
Longer assumed equipment lives can improve reported depreciation and near-term accounting profits without changing the original cash outlay. Shorter economic lives can increase the replacement spending required to sustain a service. Distinguish the accounting schedule from the asset’s actual productive life.
Backlog helps describe contracted activity, but timing, delivery, cancellation terms and collection still matter. It is not cash already available to repay capital.
Alphabet’s Q2 2026 operating cash flow was $39.069 billion against capital expenditure of $44.924 billion, producing negative quarterly free cash flow of $5.855 billion. Its trailing twelve-month free cash flow remained positive at $53.273 billion. A weak quarter is not the same as an annual funding deficit.
In the quarter it also raised $49.6 billion net through common and mandatory-convertible preferred stock, and $20.3 billion net through notes. The equity proceeds included funding for AI infrastructure among their stated uses. Alphabet’s Q2 2026 release, financing disclosures and cash-flow reconciliation.
These disclosures show multiple funding channels being used. They do not trace every borrowed or raised dollar to a particular asset, or establish that future capital will remain available on the same terms.
NVIDIA’s filing describes guarantees signed in August 2026, capped at $105 billion, supporting specified lease and power obligations at an SB Energy campus used by OpenAI. The commitments become effective in stages as applicable leases commence, subject to conditions; the buildout was expected to begin entering service in fiscal 2029. NVIDIA’s Q2 FY2027 filing.
A guarantee can help another party obtain financing while moving contingent risk to its guarantor. Its maximum exposure is not an immediate cash payment, booked borrowing or expected loss. Adding all those categories into one headline “debt” number would obscure the economics.
For each obligation, record its payer, beneficiary, timing, conditions and overlap with other commitments. That is more informative than an impressive aggregate that counts the same project several times.
Our working view is constructive on operating momentum and unresolved on the eventual adequacy of returns. The financing question has moved beyond whether these companies can raise money today to what that money costs and whether the resulting capacity earns enough.
That is not proof that funding access can never constrain the cycle. A repriced financing arrangement, slower collection or weaker utilisation could make it binding again. Equally, falling compute costs and expanding paid usage could improve the economics even while aggregate spending rises.
The constructive case strengthens if new capacity supports durable paid usage, healthy margins after running costs, and cash generation that keeps pace with the capital needed to maintain it. Lower unit costs can widen adoption; the key question is how much of that benefit survives competition and reaches owners.
The case weakens if commitments and replacement requirements grow while customer collections, utilisation or pricing disappoint—or if financing becomes materially more expensive before projects earn cash.
Useful monitoring separates three clocks: construction payments, operating ramp-up and capital repayment. A mismatch can create a funding problem even for a project with a plausible long-run return. The cash-timing example shows the distinction.
A stock valuation would require another step: an explicit path for cash flows, a discount rate and a dated market-price comparison. This sector argument does not provide that completed valuation.