AI generation is fast. The energy bill behind it is not something most people think about while they sip coffee and ask a chatbot a question. But the math is catching up, and the United States has a quantified headline figure to prove it. The International Energy Agency’s most recent estimates put US data centers at 224 terawatt-hours of electricity in 2025, which is more than 5% of the country's electricity use. That is a major jump from an estimated 1.9% in 2018, years before generative AI became mainstream.

Part of why the numbers are rising is that AI’s day-to-day “inference” work scales with every user query, not just the one-time training run. Google’s own estimate says processing a median-length text prompt with its AI assistant Gemini consumes around 0.24 watt-hours. On its own, 0.24 watt-hours is easy to shrug off. It’s roughly equivalent to watching TV for about nine seconds. But multiply “nine seconds” by millions of prompts, then by billions of queries across the internet, and the power demand becomes a real constraint.

And the constraint is getting sharper because the AI market is in a build frenzy. To compete for generative AI leadership, companies including Google, Meta, Amazon, OpenAI, Anthropic, Microsoft, and Oracle are investing tens to hundreds of billions of dollars to build AI-focused data centers. In pre-AI days, data centers often consumed around 100 megawatts of electricity, enough to power about 83,000 homes with average demand. The newer “hyperscale” wave can use a gigawatt or more, roughly a tenth of the electrical capacity of Los Angeles. Bigger facilities mean bigger electrical commitments, and that raises the stakes for every sourcing and efficiency decision.

One reason executives are getting nervous is that clean power is not guaranteed. Experts cited in the piece say much of the demand is being met by plants powered by fossil fuels, such as gas. The big practical problem is siting: data centers are often built in locations without abundant renewable energy sources like hydropower, geothermal, solar, or wind. Tech companies sometimes offset emissions by investing in renewable energy elsewhere. But if the renewables installed do not produce more clean electricity than the data centers consume, the offset strategy, at best, freezes CO2 emissions rather than reducing them to near zero. Eric Masanet, a researcher on data center sustainability at the University of California, Santa Barbara, frames it bluntly: for every megawatt of fossil fuel power installed, it sets progress back.

Energy is only one line item, too. The article highlights second-order impacts that boards rarely quantify in spreadsheets until a public backlash or operating constraint forces the issue. Communities near gas plants can face air and noise pollution. Cooling also matters: data centers use water resources, and strain on local water resources is a concern. Even beyond operations, there is the energy and materials embedded in manufacturing the hardware that fills new data centers. So the “treat energy as a technical issue” mindset is getting harder to defend.

Underneath all the headlines is a technical reason energy use keeps ballooning: the way large language models work. Transformers, a design described in 2017 by Google Brain, process text quickly by weighing relationships between every word and every other word it sees. That requires lots of computations, represented by additions and multiplications once words or fragments are converted into numbers. GPUs, mostly manufactured by NVIDIA, enable the parallel processing that makes this fast enough to feel magical. But it comes with two energy phases: the initial training, which can consume enormous electricity, and inference, which is the part that repeats every time someone asks a question.

Researchers are now focusing heavily on inference because it is the compounding cost driver. “You train once, then you inference for a billion people in the world,” Mosharaf Chowdhury of the University of Michigan says, in the piece. Transformer inference is described as surprisingly inefficient because for each generated word, the model reruns a large set of computations. It selects the next word by using the probability given the context, but doing that requires applying parameters from training that number in the hundreds of billions or even trillions. Günter Klambauer of Johannes Kepler University in Austria points to the core issue: lots of calculations for a single word to be added.

This is why the response now looks like a multi-front scramble, not a single fix. The article notes a shift toward smaller language models specialized for specific tasks, which reduces parameter counts and cuts computation compared with larger models. It also points to efforts in energy-saving algorithms, processor designs, and smarter planning of where and how data centers are constructed. Fengqi You, an energy systems expert at Cornell University, argues that AI’s energy cost is not an accident, but a product of how systems are built. With the right mix of solutions, he says, the trajectory can be reshaped.

For decision-makers, this is the strategic pivot: AI expansion is no longer just a product roadmap issue. It is an energy procurement, infrastructure permitting, and carbon risk issue that intersects with real-world geography and the grid. If efficiency gains do not keep up, projections in the piece suggest US data centers could soon release the equivalent of 24 to 44 megatons of CO2 annually, with the upper end compared to Norway’s emissions. That is the kind of range that makes CFOs and boards stop treating “power” as a background utility and start treating it like a market force.