AI financing has two buyers left: the customer and the state

PWD's coverage ran to 35 items this week, and the largest of them shared a shape: a named customer on one side, financing on the other. SpaceX's compute leasing book runs at $41.1 billion a year, carried by four counterparties, and the company deepened it with a $13.3 billion contract that added revenue without adding names—close to a third of the annual total. When a buyer of that size has signed, the data hall beneath it, the substation, and the transformer order all become bankable against the lease; when nobody has signed, the financing has nothing to attach to. The constraint on the AI buildout has moved from capex to counterparty.

Concentration of that order is a financing property as much as a revenue one. A book resting on four names is a wager a public market can price and a lender can model, right up until one of the four changes its mind, at which point the revenue line and the collateral underneath it move together.

Oracle showed the structure from the seller's side, tripling its quarterly delivery rate to 850 megawatts while planning less capital expenditure than its cloud peers because the customers are buying the hardware. The megawatts are real; the capex sits on somebody else's balance sheet. Growth underwritten by your own clients works for as long as the clients keep signing.

Anthropic's data center arm pushed the same logic further, hiring the architect of Equinix's hyperscale joint ventures and putting the anchor tenant on Theseus's cap table. Once the tenant sits inside the ownership, the lease risk and the equity risk collapse into a single exposure—the shell gets built, and the credit is double-stacked on one demand forecast—which makes the structure easier to finance and harder to diversify.

Where the customer does not exist, a different balance sheet arrives: the reported $5 billion loan from the Pentagon to Fluidstack is aimed at component manufacturing, the one layer of the AI capital stack with no hyperscaler standing behind it. A data hall can be pre-let and a lease can be priced off a named tenant's demand forecast, but a component line sells to an industry rather than a company, and no single buyer will underwrite capacity that its competitors will also use. If the report holds, the Pentagon is not subsidizing AI compute so much as buying the industrial base the compute depends on, and absorbing the merchant risk that has no offtaker to lean on at that layer.

Microsoft's 38-gigawatt target is where the arithmetic gets tested, because the implied quarterly build rate sits barely above the company's current pace, which makes the headline a lease-and-grid forecast. What actually gets built turns on how much capacity is leased from third parties and on what state power rules allow, and neither of those numbers was settled this week.

NextDC's AU$1.1 billion convertible raise is the same wager from the other direction, with notes priced at a 32.5 percent conversion premium and the company's third raise in four months. A conversion premium is a promise that a buyer will want the stock later, so the demand forecast ends up at the bottom of the capital stack, carried by the noteholder.

The week carried the counterexample in another corner of digital infrastructure, where the owners of Virgin Media O2 are cutting £600 million after a 17-point slide in a $925 million note that trades at 61 cents, a market verdict on the capital structure. When the equity has to defend a bond price, the cut lands on the one line a fiber builder can least afford to trim.