America's largest data center bonds are sold before power is secured
Pension funds have financed AI campuses in places like Louisiana without any guarantee that the huge promised loads will be supplied by local utilities.
America's technology giants are raising record sums to build AI campuses, as Blue Owl Capital did in Louisiana, where debt was sold to pension funds through custom structures. The lenders signed before there was proof the required megawatts would be delivered.
Morgan Stanley identified a gap between the electricity US data centers need through 2028 and what’s actually contracted, as Fortune reported on August 20. Billions have been spent in anticipation of uncommitted gigawatts, making that mismatch central to the AI trade.
Meta plans to keep its models competitive with a 2026 capital budget between 115 and 135 billion dollars, according to a LinkedIn summary of Meta guidance from January 2026. Amazon pushed its spending plan to roughly 220 billion dollars for 2026, citing uncontrollable memory-chip prices, TradingNews reported in August 2026.
Blue Owl Capital secured fees and a long-term bond product by raising around 27 billion dollars in bonds to fund Meta's Hyperion campus in Richland Parish, Louisiana, in one of the largest private bond sales ever. PIMCO purchased those bonds for teachers’ pension funds, the New York Times reported on July 27. This puts pensioners at the mercy of a construction schedule in rural Louisiana.
Motley Fool relayed Morgan Stanley’s August 20 figures: US data centers require about 68 gigawatts of new power between 2026 and 2028, but only 30 gigawatts are covered by projects or contracted grid capacity. That leaves 38 gigawatts exposed, with interconnection queues taking five to seven years in some areas.
The bottleneck predates the current buildout. Money moves with term sheets’ speed, but electricity follows substations’ pace, and the gap widens every quarter that capital spending guidance outpaces utility planning cycles.
US data center construction hit an annualized rate of 59.3 billion dollars in May, a 23 percent increase year over year, with steel and concrete arriving on schedule for now, according to Census Bureau data via Associated General Contractors in June 2026.
Data Center Watch counted at least 75 projects worth about 130 billion dollars blocked or delayed in the first quarter of 2026, matching all of 2025’s delays in just three months, according to 24/7 Wall St. on August 20. The spending and delay curves are steepening; only one trend can continue.
Bonds pay interest from the moment of issuance, but revenue only comes when racks are powered. Bondholders are paid before transformers are even ordered, and each quarter of grid delay converts equity returns into extra carrying costs. If delays affect core projects, borrowers must refinance under worse terms or require sponsors like Meta to top up collateral, which strains their free cash flow as these companies tell investors the buildation is self-funding.
The most apt history is the telecom fiber boom of the late 1990s, in which Global Crossing and competitors borrowed tens of billions to lay cable based on demand projections that materialized years late, causing debt to sour before assets did.
Key differences mark this cycle: borrowers like Meta, Microsoft, Alphabet, and Amazon have immense cash flows and can absorb some overbuild, unlike Global Crossing, which lacked other businesses. But the contrast is clear—a buried fiber optic cable can wait for traffic almost costlessly, whereas a data center without grid connection accrues property taxes, security costs, and depreciation without income.
Utilities and gas-turbine manufacturers benefit first, as Morgan Stanley’s solution to the electricity deficit is onsite generation and converted bitcoin-mine hookups, which can shave years off queue times—Morgan Stanley via Blockspace, July 14. Counties striking tax deals, like Richland Parish, discover a delayed campus delivers costs but no revenue.
Much of the expansion takes place in joint ventures and GPU purchase agreements, which remain off official debt lines until facilities are operational, so scheduling slips transform accounting entries into visible losses. Power equipment suppliers and independent power producers profit if this analysis holds; each missed quarter raises workaround prices. The bond buyers, including PIMCO, pay out to pension members from a revenue stream that does not exist until the grid connects. Hyperscalers can absorb misses; creditors cannot repossess a substation queue.
The May construction number demonstrates this: spending is converting to buildings at a record-breaking pace, with delays concentrated at the speculative fringe among developers without tenants. If hyperscaler-owned sites keep pouring concrete, the 38-gigawatt gap becomes a future concern, and bonds are repaid by America’s strongest companies. This scenario is plausible and trackable, making physical site activity—not earnings calls—the key indicator.
Confirmation comes from an upcoming Census construction release showing continued spending even as announced capex rises, signaling funds piling up faster than work progresses. Disproving evidence would be Meta reporting Hyperion campus energization on schedule and Morgan Stanley’s shortfall shrinking with more generation contracts.
Dirt movement, not executive commentary, is the test. The AI trade has become a utility-lag trade, a lesson lenders learned only after committing.