For two years the story of AI and energy has been one of terrifying arithmetic: data centers demanding more power than nations can build, grids buckling, nuclear plants worth of shortfall stretching to the horizon. This week brought a twist. A large share of that demand, analysts now say, is not real.
New projections from Wood Mackenzie, reported via ZeroHedge, estimate that grid operators will ultimately commit to only about 28 percent of the 1,066 gigawatts of electricity currently requested for US data center projects. The rest is what the industry has started calling “bragawatts”: phantom demand. Developers, racing to secure capacity, pitch the same project to several utilities at once, planning to proceed only with whichever offers the best deal. A separate estimate from Rapidan Energy put the share of genuine, buildable projects at just 20 to 30 percent.
The scale of the inflation is almost comic. In Texas alone, interconnection requests to the ERCOT grid reached 474 gigawatts, of which about 90 percent were data centers, against an all-time peak demand of 91 gigawatts. That is more than five times the entire state’s record load. Governor Greg Abbott responded by freezing new data center approvals, unable to tell the real proposals from the speculative ones clogging the queue.
This should be reassuring, and in one narrow sense it is. If most of the demand evaporates, the apocalyptic forecasts of grid collapse soften. But the phantom projects carry a real cost of their own. As Bloomberg noted, inflated projections complicate genuine planning, and the capital utilities spend bracing for demand that never arrives is typically recovered through rate hikes. The everyday customer helps pay for infrastructure built to serve a data center that was never going to exist.
That cost-shifting is the throughline. One technology explainer this week described how the aging American grid, much of it designed in the mid-20th century, was never built for the concentrated, always-on load of hyperscale AI campuses. As the wealthiest firms strike private deals for dedicated power, including a growing push toward small modular nuclear reactors, the public grid is left to absorb the volatility and the maintenance bills.
And then there is the emissions question, where the framing flips entirely. Greenpeace USA seized on new research, published in a Nature journal by the Enabled Emissions Campaign, arguing that the climate threat from data centers is not even AI’s worst. The bigger danger, the group says, is the use of AI to make the fossil fuel industry more productive: helping companies find and extract more oil and gas. The report estimates those enabled emissions run 3.3 to 13.3 times larger than today’s data center emissions.
Put the two findings side by side and a strange picture emerges. The visible, much-feared number, AI’s electricity appetite, may be substantially a mirage inflated by speculative paperwork. The invisible one, the carbon AI unlocks by making drilling cheaper and faster, may be the figure that actually matters. The grid was always the easier thing to count.