Data Centers Are Getting Power-Hungry Faster Than the Grid Can Be Rebuilt

A single ChatGPT query uses 10 to 100 times the energy of a standard Google search, depending on prompt complexity. That is the smallest meaningful unit of the AI power story, the one readers can hold in their heads. Multiply it by every user, every day, and the question is no longer whether AI uses more electricity than the last generation of cloud software, but whether the grid can carry what AI is about to ask of it.

The next rung up the stack is the rack. AI server racks now draw 100 to 120 kilowatts apiece, according to Vittorio Pierangeli, SVP of PowerGen at Rolls-Royce Power Systems. Five years ago the same rack drew roughly 10 kilowatts. The jump is not a tuning problem; it is the physical signature of moving from CPUs and small models to densely packed GPU clusters running much larger ones.

Stack the racks into data centers and the data centers into the national grid, and the cumulative U.S. power supply to data centers is forecast to fall more than 50 gigawatts short of demand by 2030, per the same Power Magazine interview with Pierangeli. Fifty gigawatts is roughly the output of 50 large nuclear reactors, or the continuous load of tens of millions of U.S. households. It is also a single named expert's projection rather than a settled grid-planning consensus, and it reads most fairly as a vendor-adjacent forecast that nonetheless lines up with a build-time problem the trade press has been documenting for years.

The structural reason is the gap between how fast the demand side ships and how fast the supply side can be built. A new data center can be stood up in 18 to 24 months. The transmission lines, substations, and generation that have to feed it run on multi-year permitting and construction cycles. When the demand curve outruns the supply curve by years, the gap is the new baseline, not a temporary lag.

That gap is what turns an electricity question into an energy-mix and energy-security question. Pierangeli points out that conventional coal plants are being decommissioned while renewable generation remains intermittent, and that geopolitical instability is adding pressure on energy security at the same moment AI data center load profiles are growing dramatically more volatile. Each of those constraints tightens the same gap from a different direction, and none of them is a scenery problem with a software fix.

The buildout curve is real, though. The backup power segment is expanding roughly 22% a year and continuous power roughly 24% a year, according to Pierangeli. That is what a market looks like when the demand curve has already broken away from the existing supply curve and operators are buying their way around the gap with on-site generation. The question for the next few years is whether that on-site buildout stays a bridge to a rebuilt grid, or whether it becomes the default answer to a problem the grid was not designed to solve.

The next place to watch is not a chip launch. It is the queue: the line of large loads waiting to interconnect, and the regulator or utility commission deciding which of them gets a hookup date and which gets a price signal instead.