The Vendor Is Now the Underwriter
Nvidia's quarter was reported as a supply story. The more consequential work was on the other side of the transaction — making someone else willing to lend against the chips.
Who pays for the build-out, how the bill reaches earnings, and what the disclosures can actually tell us.
The infrastructure behind AI has three different clocks. Cash is committed when equipment is ordered or financed. Accounting expense arrives as assets enter service and depreciate. Customer revenue arrives only if someone buys the resulting capacity. A convincing account of the build-out has to connect all three.
Our starting point is the financing chain: a supplier can participate in a customer’s funding as well as sell to it. That does not by itself establish circular revenue or bad demand. It does make the identity of the ultimate payer, the terms of the commitment, and the allocation of risk worth investigating.
The second question is measurement. Company-wide property and equipment purchases include activities beyond AI. Finance lease additions are different from this quarter’s lease payments. Guidance is an intention rather than spending already incurred. Our disclosure comparison keeps those boundaries visible instead of collapsing them into a league table.
The reading order below moves from financing to accounting, physical constraints, the allocation of grid costs, and competitive advantage. Our electricity analysis distinguishes payment for a connection from the cost of dependable generation and energy: assigning one bill does not settle the others. Use the depreciation scenario to isolate one accounting assumption, then return to the filings: a lower annual expense is not a lower purchase price, and neither tells us whether a new facility will earn its cost of capital.
Nvidia's quarter was reported as a supply story. The more consequential work was on the other side of the transaction — making someone else willing to lend against the chips.
Useful-life assumptions change reported infrastructure costs. They need to be tested alongside utilization, cash spending and the financing behind demand.
The hyperscalers compete fiercely on everything except the thing that now constrains them all: where to find the next gigawatt.
A data center can pay for its connection and still affect everyone else's power bill. Ohio's tariff and PJM's proposals address different parts of that problem.
Falling inference prices and narrower benchmark gaps put model-level pricing power under pressure. Distribution, data and workflow integration deserve a separate test.
The chip supply chain is not decoupling so much as duplicating — at enormous cost, and with no guarantee the copies will ever run at scale.
Microsoft’s cash-flow statement reports $35.802 billion of additions to property and equipment for the quarter ended June 30, 2026. This is a company-wide cash-flow item. Its fiscal-year total is $115.948 billion.
The cash measure provides a traceable starting point for assessing investment. It should be matched to the exact period and compared with earlier observations using the same definition. Financing arrangements and the timing of payment can separate cash expenditure from equipment becoming available.
The line does not identify spending solely on AI, the amount of capacity put into service, or its eventual return. It also cannot be substituted for a management capex measure that includes finance lease additions.
Start with the cash line, then reconcile the additional definitions rather than select whichever number produces the strongest headline. The disclosure comparison keeps these measurement boundaries beside the figures; the depreciation scenario isolates a different question: how capitalized cost reaches earnings.
Read four companies’ infrastructure disclosures without confusing cash spending, leases, and forecasts.
Compare the accounting cost of the same investment over two useful lives.
Who Pays for the Next Gigawatt? extends the dossier’s financing analysis to grid obligations. AEP Ohio’s minimum-demand framework addresses underused infrastructure, while PJM’s July and August filings propose ways to secure supply and manage large loads. We distinguish those proposals from effective tariffs and explain why a contribution to connection costs alone does not establish protection from every electricity-cost increase. The update adds this distinction to our analysis; it does not announce a new regulatory decision today.
The dossier now separates company-wide cash property-and-equipment spending from lease payments and useful-life assumptions. This establishes the measurement baseline for subsequent reviews; it is not a new claim that all of those expenditures fund AI.