Generating AI videos may be creating bigger problems than deepfakes, a new study warns of a looming global crisis

As AI-generated videos flood social feeds and creative platforms, a sobering reality is emerging: the biggest threat may not be misinformation or deepfakes, but energy itself. A new study by Hugging Face, titled Video Killed the Energy Budget, reveals that text-to-video systems consume staggering amounts of electricity, with costs skyrocketing as clips get longer or sharper.

Energy Demands That Spiral Out of Control

Unlike text or image generation, video generation grows exponentially more expensive for the planet. The researchers found that energy and latency scale quadratically with video resolution and length, and linearly with denoising steps. In practice, a six-second clip may demand nearly four times more power than a three-second one. For high-resolution videos, the consumption shoots up to hundreds of watt-hours per clip—thousands of times greater than generating text or images.

AI’s Power Hunger

This revelation echoes warnings from industry veterans. Former Google CEO Eric Schmidt recently argued that AI’s natural limit is not silicon chips but electricity, predicting the U.S. may need the equivalent of 92 nuclear power plants to fuel its AI ambitions. Already, Microsoft has signed nuclear energy deals and OpenAI has invested in fusion startups to secure power for future growth.

Why It Matters More Than Deepfakes

While society remains preoccupied with the dangers of deepfakes, experts caution that the environmental costs of generative video could spark a crisis of their own. The Hugging Face team observed that GPU usage accounts for more than 80 percent of energy in every model tested, with larger systems consuming 3,000 times more power than lighter ones. At scale, this demand threatens to derail global climate commitments, as seen in Google’s 2024 report showing a 13 percent rise in carbon emissions largely driven by AI.

The Hugging Face report suggests strategies such as diffusion caching, pruning inefficient training data, and quantization to reduce the carbon footprint. Yet, these fixes may only slow the tide. As Sasha Luccioni of Hugging Face notes, “Video diffusion is far more costly than text or image generation… highlighting the need for hardware-aware optimizations and sustainable model design”.

AI videos promise creativity without limits, but they carry an unseen price tag. Unless efficiency innovations catch up, the very technology that powers digital imagination may leave us facing an environmental bill too high to pay. The crisis ahead may not just be what AI creates, but how much power it takes to create it.