AI's bond bill: $220B issued in 2026, and the spread is now part of the compute price

Seventeen times more paper in twelve months ^[5] is not a funding problem by itself. The problem is arithmetic further down the chain. A rented GPU hour has to cover the depreciation of the machine and the cost of the money that bought it, and most of that money now arrives as term debt against contracted capacity. The coupon is fixed at issue, so a wider spread does not change this quarter's invoice. It changes the next tranche of capacity and the renewal price on the reserved contracts written against it.

The concession detail matters more than the headline total. Amazon is about as strong a credit as this sector produces, and Reuters, as relayed in theneuron.ai's Friday digest, says it still had to give ground on a $25B raise while some large buyers warned they were nearing exposure limits ^[2]. That deal alone is roughly 11 percent of all US AI-related issuance this year ^[6]. When the best name in the queue sets the clearing level, everyone behind it prices off that, and the queue is long: Broadcom is reportedly in talks for a $70B to $80B package tied to AI chip deals ^[4], which is between about 32 and 36 percent of everything issued so far in 2026 ^[7].

Exposure limits are a quantity constraint, not a price one. A buyer at its sector cap stops buying at any spread, which means the marginal dollar for the next mega-raise has to come from somewhere with a less visible price: private credit, structured vehicles, or supplier paper. None of those publish a spread anybody can check.

There is a real disagreement in the reporting worth holding onto. The New York Times account has the borrowing binge helping keep Treasury yields elevated because markets are pricing stronger AI-driven growth and higher rates for longer ^[3], which is a story about confidence. The Reuters account has investors demanding compensation and running out of room ^[2], which is a story about crowding. For anyone buying compute, both readings push the same direction: the risk-free base and the credit spread stacked on top of it are the two halves of the all-in coupon, and they are moving together.

Meanwhile the supply side is committing on a longer clock. Micron has put a planned $10B over the next decade behind a Boise research hub, inside a wider $50B buildout, with its chief executive saying AI has "totally changed" memory's boom-and-bust equation ^[8]. That capital is being pledged in decade units against demand whose financing is repricing deal by deal. If you are signing multi-year reserved capacity, the clause to read is not the discount. It is who absorbs the refinancing.