Satellites catch data centers falling behind the promises made about them

Two things are true at once and they cannot both survive the year. America's biggest technology companies have told investors their artificial intelligence data centers will arrive on schedule, in time to earn returns on hundreds of billions of dollars of committed capital. And satellites photographing those same sites from orbit show dirt where buildings were supposed to stand. An analysis by the Financial Times using imagery from the geospatial firm SynMax found that nearly 40 percent of American data center projects planned for completion in 2026 are likely to miss their dates by more than three months (Financial Times via Ars Technica, Apr 17).

The contradiction is not accounting fraud. It is a scheduling fiction that everyone in the chain had an incentive to keep alive. Microsoft, Oracle and OpenAI need completion dates to justify capital spending to shareholders and to lock in cloud contracts priced against future capacity. Utilities need announced load growth to defend rate cases and grid investments before regulators. Contractors need signed schedules to win bids. The satellite does not care what any of them filed. SynMax's Vulcan platform pairs orbital and drone imagery, including thermal shots of sites like a Stargate-linked facility in Garland, Texas, with permit records compiled by the research group IIR Energy, so progress can be measured against promises rather than taken from press releases (Ars Technica, Apr 17).

The trigger this spring was the imagery itself, cross-checked against public statements and permits. What it exposed was the slow pressure underneath: the people who pour foundations were never the constraint. Interviews with more than a dozen industry executives pointed to chronic shortages of labor, power and equipment, plus permitting delays (Financial Times, Apr 17). Large power transformers now carry average lead times around 128 weeks, and generator step-up units around 144 weeks, against roughly two years or less before 2020 (Wood Mackenzie transformer survey, 2026). Tariffs on imported Chinese equipment have stretched those waits further (Ars Technica, Apr 17). A building can rise on schedule and still sit dark, waiting for a steel box to connect it to the grid.

Follow who pays when the dates slip. The hyperscalers pay first, in delayed revenue recognition and in carrying costs on land, chips and crews that sit idle. But the bill does not stop there. Data centers draw electricity on the scale of hundreds of thousands of homes each, and utilities recover grid upgrades from everyone connected to the wires (Ars Technica, Apr 17). In Virginia, the self-described data center capital of the world, a recent poll found a majority of residents worried about land use, environmental impact and their own power bills (Ars Technica, Apr 17). The Trump administration responded in March with a Ratepayer Protection Pledge signed by major tech companies, which carries no meaningful enforcement, while Microsoft separately pledged to cover the full electricity costs of its own data centers (Ars Technica, Apr 17). When a voluntary pledge needs a name like that, the political pressure is already real.

The Globe and Mail

The companies themselves are quietly admitting the calendar problem by reshuffling it. In March, Oracle and OpenAI scrapped plans to expand their flagship Stargate site in Abilene, Texas, from 1.2 gigawatts toward roughly 2 gigawatts after financing dragged (Bloomberg, Mar 6). Microsoft has since moved to claim the adjoining campus that OpenAI walked away from, a deal reported across the industry but not yet confirmed in filings, so treat it as the shape of the market rather than a settled fact: capacity does not vanish when a deal dies, it changes owners, and the buyer is the company with the balance sheet to wait out a transformer queue.

History offers one close model: the fiber-optic mania of the late 1990s. Carriers like Global Crossing and WorldCom announced intercontinental routes on investor timetables, and by 2002 only a small fraction of the buried glass was actually carrying traffic. The schedules were marketing documents. Yet the counter-example argues the other way too: the dark fiber laid then became the cheap backbone of the broadband decade afterward, so late delivery did not falsify the demand thesis. It only bankrupted the people who had promised the timing. This time the physical bottleneck is heavier than glass in a trench; you cannot overbuild a transformer by accident, because each one is ordered against a specific substation years ahead.

Capital moves at the speed of a press release; transformers move at the speed of factories, and the gap between them is where the money changes hands.

Now walk the chain forward without the analyst labels. First, 2026 completions slide into 2027, and cloud capacity that customers contracted for arrives late, pushing buyers toward whatever compute exists, which is priced higher. Then the equipment makers and gas turbine suppliers holding real backlog gain pricing power, while developers whose value depends on dated delivery promises lose it. Eventually the politics turn: ratepayers seeing bills rise to subsidize buildings that do not open become voters, and the moratoria already spreading through Virginia spread to Ohio and Texas. The bottleneck migrates from construction speed to permission.

What confirms the read: watch the next round of SynMax-style imagery and utility commission filings. If foundation work at the big Texas and Midwest campuses still shows no vertical construction into the autumn, and interconnection agreements keep slipping past their contractual energize dates, the miss compounds. What breaks it: a genuine acceleration in grid equipment supply, or hyperscalers standing behind their own generation fast enough to skip the queue, visible in on-site turbines actually spinning at announced sites within the year.

The judgment the numbers support is uncomfortable for both bulls and bears. The satellites do not show that the AI buildout is a fraud, any more than the fiber glut showed the internet was fake. They show that capital moves at the speed of a press release while steel, copper, transformers and linemen move at the speed of factories and apprenticeship programs, and the gap between those speeds is where fortunes get transferred. The winners will be whoever owns the scarce physical link, not whoever owns the announcement.