Data center site activity offers a check on whether borrowed billions become steel on schedule
The money has already been borrowed against buildings that need electricity no one has agreed to deliver, so the ground itself is where the promise gets audited.
Two things are true right now and they cannot both survive contact with the power grid. America's biggest technology companies are borrowing record sums to build AI campuses, selling the debt on to pension funds through structures like the one Blue Owl Capital assembled in rural Louisiana, while Morgan Stanley counts a gap between the electricity US data centers will need through 2028 and the supply anyone has actually contracted (Fortune, Aug 20). The lenders signed before anyone could prove the megawatts would show up. That sequence is the story.
The actors line up cleanly. Meta wants training campuses fast enough that its models stay competitive, and its 2026 capital budget runs between 115 and 135 billion dollars (LinkedIn summary of Meta guidance, January 2026). Amazon raised its own spending plan to roughly 220 billion dollars for 2026, blaming memory-chip prices it does not control (TradingNews, August 2026). Blue Owl Capital wanted fees and a long-duration bond product, and got both by selling about 27 billion dollars of bonds to fund its stake in Meta's Hyperion campus in Richland Parish, Louisiana, one of the largest private bond sales ever (New York Times, Jul 27). PIMCO bought those bonds on behalf of teachers' pension funds (New York Times, Jul 27). The pensioner is now downstream of a construction schedule in rural Louisiana.
The trigger this month is Morgan Stanley's arithmetic: data centers need roughly 68 gigawatts of new power between 2026 and 2028, only about 30 gigawatts of it covered by projects under construction or contracted grid capacity, leaving 38 gigawatts exposed, with interconnection queues running five to seven years in some regions (Motley Fool citing Morgan Stanley, Aug 20). The slow pressure underneath is older and simpler: money moves at the speed of a term sheet and electricity moves at the speed of a substation, and the gap between those two speeds widens every quarter that capex guidance rises faster than utility planning cycles.
The ground-level check the headline names comes from two places. Actual US data center construction hit a seasonally adjusted annual rate of 59.3 billion dollars in May, up 23 percent year over year (Census Bureau data via Associated General Contractors, June 2026), which says steel and concrete are arriving on time for now. Against that, Data Center Watch counted at least 75 projects worth roughly 130 billion dollars blocked or delayed in the first quarter alone, matching all of 2025 in three months (Data Center Watch via 24/7 Wall St., Aug 20). So the spend curve and the delay curve are both steepening. One of them has to give.
The bondholders were paid before the transformers were ordered, and now the ground itself audits the promise.
The mechanism runs like this. The bonds pay interest from the moment they clear, but revenue starts only when racks hum, and racks hum only when megawatts arrive. The bondholders were paid before the transformers were ordered, and every quarter of grid delay converts equity returns into carrying costs. If delays spread from the fringe to the core pipeline, the borrowers either refinance into a worse market or ask the sponsors, Meta and its peers, to top up collateral, which shows up as weaker free cash flow exactly when those companies are telling shareholders the buildout pays for itself.
The history that fits is the telecom fiber boom of the late 1990s. Global Crossing and its rivals borrowed tens of billions to lay transoceanic cable against demand projections that arrived years late, and the debt went bad before the asset did. What is different this time: the borrowers, Meta, Microsoft, Alphabet and Amazon, sit on enormous cash flows and can absorb their own overbuild, unlike Global Crossing, which had no other business. And what argues the other way: the fiber was buried and waited, nearly free, for traffic to find it. A data center without a grid connection is not waiting patiently underground. It is property taxes, security fencing and depreciation accruing on an empty field.
Follow the consequences down the chain. First, the utilities and gas-turbine makers get pricing power, because Morgan Stanley's answer to the shortfall is onsite generation, fuel cells and converted bitcoin-mine hookups, anything that buys back years of queue time (Morgan Stanley via Blockspace, Jul 14). Second, the counties that cut tax deals, like Richland Parish, learn that a delayed campus still costs them services but pays nothing. Third, much of the buildout sits in joint ventures, leases and GPU purchase commitments that stay off the companies' reported debt lines until facilities go live, so a schedule slip converts accounting footnotes into visible losses rather than staying buried.
Who profits if the read is right: the power equipment suppliers and the independent power producers who own scarce megawatts, because every slipped quarter raises the price of the workaround. Who pays: the bond buyers, PIMCO among them, whose paper pays out of a revenue stream that does not exist until the grid connects, and ultimately the pension members behind them. The hyperscalers themselves can eat a miss. Their creditors cannot repossess a substation queue.
The honest counter-read is that the May construction number proves the machine works: spending converts to buildings at record pace and the delays concentrate at the speculative fringe, the merchant developers who never had a tenant. If the hyperscaler-owned sites keep pouring concrete, the 38-gigawatt figure becomes a 2029 problem priced in 2028, and the bonds get repaid by the strongest balance sheets in corporate America. That possibility is real. It is also testable, which is why site activity, not earnings calls, is the instrument.
What confirms the read: the next Census construction release showing data center spending flattening even as announced capex keeps rising, meaning money is piling up faster than shovels. What breaks it: Meta reporting Hyperion energization on schedule and Morgan Stanley's shortfall shrinking as new generation contracts get signed. Watch the dirt, not the deck. The judgment this run earns: the AI trade has quietly become a utility-lag trade, and lenders discovered it after signing.