Nvidia chief executive Jensen Huang told a Stanford class in May that AI computing will eventually need "probably 1,000 times more" energy than we have today, before cheerfully admitting he might be off by a couple of orders of magnitude. The number is a shrug dressed as a forecast, but it captured a mood. The same month Goldman Sachs projected that US data-center demand would more than double between 2025 and 2027, from 31 to 66 gigawatts, lifting AI's share of national electricity from roughly 4 percent to well over 8 percent. Bloom Energy thinks the figure could reach 12 percent by 2030. On these numbers, AI is a freight train bearing down on the grid.
Meta is behaving as if it believes them. This week word spread that the company has quietly exited RE100, the corporate pledge it joined in 2016 to run entirely on renewable electricity. The reason is blunt arithmetic: wind and solar cannot deliver firm, around-the-clock power to a data center on the timeline Meta's AI buildout demands. So it has turned to natural gas, commissioning a 200-megawatt plant in Ohio and striking a deal with Entergy for a fleet of gas plants in Louisiana that will feed more than seven gigawatts to a single site called Hyperion. Climate Group, which runs RE100, said Meta could no longer meet its technical criteria and let it go. The gesture is telling. A company that spent a decade branding itself green decided the AI race was worth breaking that promise for.
Meta is not alone in the squeeze. Google, Microsoft and Apple all set headline 2030 climate targets around 2019 and 2020, before anyone imagined generative AI. Those deadlines now look less like finish lines than tripwires, as the same firms pour concrete for data centers whose electricity and water appetites are drawing local opposition. The awkward truth is that the cleaner a hyperscaler wants to look, the harder AI makes it to stay that way.
And yet a study out of the University of Waterloo and Georgia Tech, published in Environmental Research Letters under the title "Watts and Bots," argues the panic is misplaced. Modeling AI adoption across occupations and industries, the researchers estimate it will add roughly 28 petajoules of annual US energy use and about 896 kilotonnes of carbon dioxide. Those sound enormous until you see the denominator: about 0.03 percent of national energy consumption and 0.02 percent of emissions. On that accounting, AI is a rounding error, and one that might be repaid many times over if the technology helps optimize grids, materials and logistics.
So which is it, freight train or rounding error? The honest answer is that both can be true. The Waterloo figures measure AI's share of a vast national total, where almost anything looks small. The gigawatt forecasts measure something narrower and more urgent: the concentrated, local strain of dozens of new plants wired to a handful of campuses, built fast, and increasingly powered by gas because that is what can be switched on in time. AI may barely move the national emissions needle while still reshaping which power plants get built, where, and how clean they are. The planet may not notice much. The county next to the data center will.