AI video clips look cheap to make — but how much more power do they actually burn than a quick chatbot answer?

This explores how the energy cost of generating a video compares with answering a text prompt. The short answer is that the corpus doesn't give a multiplier. No note here measures video generation against text queries side by side, so any specific ratio ("100x", "1,000x") would have to come from outside this collection. What the corpus does offer is something more useful than one number: a way to think about why video matters for AI's energy footprint even as each individual task gets cheaper.

The key framing comes from the IEA's analysis How do efficiency gains and new AI capabilities balance each other?. AI electricity demand is the net result of three forces that move independently. The energy needed per task is falling by orders of magnitude each year. Usage is surging. And the mix of tasks is shifting toward energy-hungry ones, with video and extended reasoning named explicitly. So the real question isn't only "how much more does one video cost?" but "what happens when a growing share of AI use moves from text to video?" A text query that keeps getting cheaper and a video clip that is inherently far heavier can coexist, and total demand depends on which trend wins.

Why is video heavier? One clue sits in how modern video models work. Lumiere's approach Can generating entire videos at once beat keyframe interpolation? gets smooth, coherent motion by processing the entire clip, every frame across space and time, in a single pass instead of generating a few keyframes and filling in the gaps. That is a quality win, but it also shows the scale of the work. A text model produces a short sequence of tokens. A video model has to produce and coordinate a whole moving image, frame by frame. The trend toward better video points toward more computation per output, not less.

The demand side matters just as much. On a Chinese short-video platform, AI-assisted creators uploaded noticeably more videos than human creators and matched their total engagement through volume, even though viewers liked each AI video less Can AI creators match human creators through posting volume alone?. When producing a video takes little effort, people make a lot of them. That is exactly the "surging uptake" force in the IEA framing, aimed at the most energy-intensive content type.

A surprising counterweight: energy may not be the main cost even when it is the main headline. Epoch AI's model of a one-gigawatt data center finds servers make up about 60% of annualized cost, while electricity is only about $0.6B of an $8.5B yearly total What actually drives AI data center costs?. So the economic pressure from heavy workloads like video shows up mostly as demand for more hardware, and the grid impact follows from that. To get an actual per-video vs. per-query energy figure, look for measurement studies outside this collection. Inside it, the lesson is that the ratio matters less than the shift in what people ask AI to make.