There’s a high energy cost to generative AI. But even the massive amount of energy needed to train and operate the large language models pales in comparison to what’s required to run the video models behind tools like OpenAI’s viral Sora app, which are flooding our social media feeds with goofy fake clips.
Generative AI models, on the whole, require a lot of energy to power. The servers that are running your ChatGPT query use a compute-intensive process that requires a lot of electricity to maintain. AI is the «biggest driver» of electricity use in North America, one report found. And that might be showing up in your power bill, with AI datacenters cropping up all over the US, raising the electric bills of households nearby. Some estimates say one AI query uses 10 times more energy than a simple Google search.
While the big AI firms are still hesitant to detail exactly how much it takes to train and run AI models, there’s a growing field of research searching for answers. Sasha Luccioni, the AI and climate lead at Hugging Face — one of the most popular AI platforms and research hubs — is a leading researcher studying the energy demands of artificial intelligence. In a new study, Luccioni and her team examined several open-source AI video models. (Popular video tools such as Sora and Google’s Veo 3 were not included in the study because they aren’t open source.)
The team used the open-source Hugging Face codebase and created AI videos with a variety of models. They measured the amount of electricity required to create those clips as they changed different factors, including making the videos longer, at a higher resolution and higher quality (something achieved through a process called denoising). They ran the test using an Nvidia H100 SXM GPU, a high-powered computer chip that can be used in AI datacenters.
«Video generation is definitely a more computationally-intensive task — instead of words, you’re generating pixels, and there are multiple frames per second to make the videos flow well,» said Luccioni in an email. «It’s complex.»
Take an AI video that’s 10 seconds long and 240 frames per second. That’s 240 images that the AI needs to generate, Luccioni explains. Especially for high-dimensional content, «That really adds up in terms of compute power and energy,» she said.
AI video energy usage
The study found that video diffusion is 30 times more costly in terms of energy spent than image generation and 2,000 times more costly than text generation. Creating a single AI video uses approximately 90 Watt-hours, compared to the 2.9Wh needed for image generation and 0.047Wh for text generation.
To put those numbers into context, an average energy-efficient LED lightbulb uses between 8-10 watts. LCD televisions can use between 50-200 watts, with newer technology like OLEDs helping run them more efficiently. For example, the 65-inch Samsung S95F, CNET’s pick for the best picture quality of 2025, typically draws 146W, according to Samsung. So creating one AI video would be equivalent to running this TV for 37 minutes.
The energy demands of generative AI, particularly for video, are significant. It sets the stage for a huge problem as AI becomes more widely used.
Increasing AI energy demands
Generative video is having something of a breakthrough moment. It’s mostly thanks to Google and ChatGPT maker OpenAI. Veo 3 and Sora, the companies’ AI video models, respectively, each launched to much fanfare and have since gone viral. The Sora app had over one million downloads five days after it launched, and Google said Gemini users made over 40 million videos in the first few months after its debut.
As AI usage grows, the US’s electrical grid might not be prepared to handle future demand. That’s why AI companies and the US government are championing a billion-dollar push for AI infrastructure. Nvidia recently announced it is investing $100 billion in OpenAI to build AI datacenters that aim to produce 10 gigawatts on Nvidia systems over the next few years. Microsoft and Constellation Energy are considering reopening Three Mile Island — the site of the worst US nuclear power plant disaster — to power its AI ambitions. But there are other ways AI energy demands can be mitigated, including using more efficient AI infrastructure.
Individually, we can think critically about whether or not we need to use an AI tool. You don’t always need — or possibly even want — an AI summary every time you look something up, Luccioni said, and using alternative browsers can help with that. But part of the problem is that AI companies aren’t forthcoming about the specifics of the energy demands of their products.
«AI companies should be transparent about their environmental impacts… It’s unacceptable that for tools that we use each day, we don’t have the precise numbers,» said Luccioni. «As users, we should have the information we need to make sustainably-minded decisions, and companies have the responsibility to provide us that information.»
