Asset 20 8 2
Does AI recommend your business? Run the free check →

Join 15,000 business owners, marketers and entrepreneurs. The Sunday newsletter you'll be annoyed only arrives once a week.

Article

Does Making AI Videos Use A Lot Of Energy And Water? Here's What The Numbers Show

If you are skim reading
Yes. Generating an AI video uses far more electricity than a text answer or an image, and that electricity is what drives the cooling water.

Yes. Generating an AI video uses far more electricity than a text answer or an image, and that electricity is what drives the cooling water. The best public measurement, from MIT Technology Review's 2025 tests, put one five-second clip from an open-source video model at about 3.4 million joules, roughly 0.94 kWh, which they compared to running a microwave for over an hour. No company publishes a water figure per video, and the closed tools most businesses use, such as Sora and Veo, publish no energy figure either. Your drafts, not your finished video, are where most of that cost sits.

Why video costs more than a chat message

I've spent a fair chunk of the last year making short AI videos for client campaigns, so this isn't theoretical for me. Text generation is quick because the model is predicting the next word, one token at a time, using a relatively small amount of compute per response. Video generation is a different beast entirely. The model has to predict thousands of frames, keep objects consistent across those frames, render motion, lighting and sound, and do it all through multiple passes of a much larger model. That takes far more GPU time, and GPU time is what turns into electricity, and electricity is what turns into heat, and heat is what the data centre then has to cool down, often using water.

The gap shows up clearly in MIT Technology Review's own measurements. A text answer from a small open model took about 114 joules once cooling and other overheads were counted. A high-quality image took about 2,282 joules. The five-second video took about 3.4 million joules, more than 700 times the image. Text is measured in hundreds or thousands of joules; video is measured in millions.

The actual numbers people quote

Nobody outside the big AI labs has the exact per-video figures, because OpenAI, Google and Runway don't publish them. What we have are independent estimates and studies that give a rough shape of the problem.

  • Video: MIT Technology Review's 2025 investigation measured CogVideoX, an open-source model, at about 3.4 million joules for a five-second clip at 16 frames per second. Divide by 3.6 million joules per kilowatt-hour and that's about 0.94 kWh per clip. An earlier, grainier version of the same model needed about 109,000 joules per clip, so the better model costs more than 30 times as much.
  • Image: about 2,282 joules for one image in the same tests, and about 4,402 joules when the quality was raised to 50 diffusion steps.
  • Text: about 114 joules per answer from Meta's small Llama 3.1 8B model and about 6,706 joules from the much larger Llama 3.1 405B, both including cooling and other overheads. That's roughly 0.00003 to 0.0019 kWh, which is why I've written before about whether ChatGPT uses water in everyday business use as a separate, much smaller question.
  • Water: Google and Microsoft both publish yearly environmental reports with their data centre water figures. The year-on-year percentages that circulate online are often quoted second-hand, so read the latest report itself before repeating one.

Sora, Veo and the other closed video models publish nothing at all. They make longer, sharper clips than the open model MIT tested, so their cost per clip is likely higher, but nobody outside those companies can say by how much.

None of these numbers are exact, and I want to be honest about that rather than pretend I have a perfect figure for you. Energy use varies by which model you use, which data centre it runs in, what the local electricity grid looks like, and what cooling system that particular facility uses. A data centre in Iowa running on wind power is a very different story to one in a hot, dry region cooling itself with fresh water. The honest position is: it's meaningfully more than text, it's real, and it's currently under-reported. If you measure environmental impact for a living and want to share what you found, you can write for us about the environment.

What about the water specifically

Water comes into this because most large data centres use evaporative cooling. Hot server racks need to be cooled, and one of the cheapest ways to do that at scale is to run water through cooling towers where some of it evaporates, the same basic principle as sweating. That water is gone from the local supply, not returned, which is why water-stressed regions like parts of Arizona and the American Southwest have started pushing back on new data centre builds.

The most widely cited figure comes from a 2023 pre-print by researchers at UC Riverside and UT Arlington: training GPT-3 in Microsoft's US data centres can directly evaporate 700,000 litres of clean freshwater. That's a training cost, a one-off, not a per-video cost, but it tells you the infrastructure behind these tools is thirsty before a single customer has typed a prompt. Every video you generate afterwards adds a smaller, ongoing draw on top of that.

What you won't find anywhere is a water figure per video. No AI video company publishes one. It also helps to know there are two kinds of water in these estimates: direct water, evaporated by the cooling towers at the data centre itself, and indirect water, used by the power stations that make the electricity. A per-video number would depend on both, plus the site and the season, so be wary of any "litres per clip" figure you see shared online, and never divide the GPT-3 training figure by the number of videos made. It doesn't work that way.

The client video that made me stop assuming

Here's my actual experience with this, because I think it matters more than the abstract numbers. I was putting together a set of short product intro videos for a client's launch using an AI video tool, testing different openings to see which one hit hardest. I generated something like thirty separate clips before we landed on the final three. Thirty renders, most of them ten seconds long, most of them thrown away.

Nobody on that call thought about energy use once. Why would we, we were focused on the brand, the pacing, whether the product looked right in the light. But sitting down to write this post made me go back and count, and thirty renders is not nothing. On the only published figure, about 0.94 kWh for a five-second clip from an open model, thirty clips would come to about 28 kWh. Most of ours were ten seconds long and made on a commercial tool that publishes nothing, so the real number is unknown and probably higher. That's one afternoon of iterating on a single client project, before we'd even picked the winner. Multiply that across every agency, every solo creator, every brand doing the same thing this week, and you start to see where the real cost sits. It isn't the finished video. It's the pile of drafts nobody ever watches.

Where the footprint piles up

This is the part that gets skipped in most articles on this topic, and I think it's worth saying plainly. The environmental conversation around AI video tends to focus on the individual user feeling guilty about making one video, when the actual scale problem is corporate and structural. It's the labs racing to release bigger, more capable video models every few months without slowing down to make them efficient. It's the marketing departments generating fifty variations of an ad because storage and compute feel invisible and free. It's the "sustainability" pages from these same companies that talk about carbon offsets and renewable energy targets while quietly not publishing the actual per-generation numbers a journalist or researcher could check.

Worrying intensely about your own three-video social post while a company runs data centres on diesel backup generators in a drought region is a bit like feeling guilty about a plastic straw while a container ship idles offshore. Your usage matters, but it is tiny next to the industrial scale of the problem, and the companies selling you the tool would rather you focus on your own guilt than their infrastructure choices.

How to make AI videos with a lighter footprint

You don't need to give up the tool. You need to be less wasteful with it, the same way you'd be less wasteful with printer paper or petrol. Here's what cuts the number of renders, based on what changed my own workflow after that client job.

  1. Write a fuller prompt the first time. Vague prompts get vague results, which means more regenerations. Specify camera angle, lighting, pacing, and mood in one go rather than nudging it across five attempts.
  2. Use image-to-video instead of text-to-video where you can. Starting from a still image you already like gives the model less to guess at, which typically means fewer wasted takes to get something usable.
  3. Batch your prompts and review before generating the next round. Generate three variations, watch all three, decide, rather than generating one, watching it, tweaking, generating another, one at a time.
  4. Repurpose existing footage instead of creating new clips from scratch. If you've already filmed something, cutting it down is far cheaper, computationally and financially, than generating new video. This is exactly the gap that AI video clipping tools built to turn long content into short clips are designed to fill, and they use a fraction of the compute a full generation model does.
  5. Keep clips short. Duration scales the cost roughly in line with frame count, so a three-second clip is meaningfully cheaper to generate than a fifteen-second one.
  6. Download and archive the final version once. Don't keep regenerating a clip because you lost the file. Save it the first time, the same instinct that applies when you're figuring out how to grab a video off social media rather than trying to recreate it, similar to the steps people search for when they want to know how to download an Instagram Story video to their phone instead of screen-recording it badly five times.
  7. Preview small before you render big. Use the lowest resolution or shortest length your tool offers for test runs, and save the full-quality render for the version you've already approved.

Here's what a draft budget saves, using the only published figure of about 0.94 kWh per five-second clip. Ten test clips come to about 9.4 kWh. Three planned clips come to about 2.8 kWh. Same finished video, roughly a third of the electricity. Treat those as estimates for an open model, because the closed tools don't publish theirs.

Work with me

Want AI doing the heavy lifting in your marketing?

I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.

The one habit worth adopting is to storyboard in stills before you render any motion. In MIT Technology Review's tests an image cost about 2,282 to 4,402 joules against 3.4 million for the video, so you can try hundreds of looks for the energy of a single clip. Get the frame right as an image, then animate it once.

Should this stop you using AI video for your business

No, and I'll say that plainly. Used with a bit of discipline, AI video is still far cheaper in time, money and often carbon than flying a crew somewhere, hiring equipment, and reshooting until you get it right. The comparison people rarely make is against the alternative, not against doing nothing at all. A studio shoot has its own footprint too, travel, lighting rigs, air conditioning, catering, all of it.

What I'd push back on is the guilt spiral some business owners fall into, where they read one scary headline about AI and water and quietly stop experimenting altogether. For scale, MIT Technology Review worked through a busy day of AI use: some chatbot questions, ten attempts at an image and three attempts at a five-second video. The total came to about 2.9 kWh, and the three videos made up nearly all of it. That's worth knowing and worth managing, not a reason to stop. If this is the kind of thing that sends you spiralling into self-doubt about every tool decision you make, it might be worth reading through the common ways entrepreneurs beat themselves up over choices like this, because most of the guilt isn't proportionate to the actual harm. Budget your drafts the way you'd budget ad spend: three planned renders, not thirty unplanned ones. If you need a nudge to stay consistent with new tools rather than abandoning them after one uncomfortable headline, my guide to staying motivated every day has more on pushing through that exact kind of hesitation.

When the regenerate button eats your afternoon

The commonest waste is regenerating a whole clip to fix one small thing: a hand with too many fingers, a mangled logo, a misspelt word on a sign. Every retry is a full render, and the model can break something else while it fixes that one detail. Do it this way instead. Approve a still first and use image-to-video, so the composition is locked before any motion is made. Add text, logos and captions in your editor afterwards rather than asking the model to draw them. And once a render is close, freeze the prompt and change one thing at a time, so each new render is a test rather than a lottery ticket.

If your clips look slightly off, my piece on what makes an AI video believable explains why, and my guide to fixing the odd sound of AI video voiceovers covers the audio side.

I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.

If you write about clean power and data centre energy, the renewable energy page is where to pitch it.

Frequently asked questions

How many kWh does one AI video use?

The best public measurement is about 3.4 million joules, roughly 0.94 kWh, for a five-second clip at 16 frames per second from the open-source CogVideoX model, measured by MIT Technology Review. Closed tools such as Sora and Veo don't publish a figure, so for them the honest answer is unknown.

Which uses more energy, an AI video or an AI image?

Video uses considerably more, because it requires generating and keeping consistent thousands of individual frames rather than one static image. In MIT Technology Review's tests a high-quality image took about 2,282 joules, or about 4,402 at 50 diffusion steps, while the newer five-second video model took about 3.4 million joules, more than 700 times the image.

Is an AI video worse than a ChatGPT answer?

Yes, by orders of magnitude. MIT Technology Review's tests on open models put a text answer at roughly 114 to 6,706 joules including overheads, and the newer video model at about 3.4 million joules for five seconds. OpenAI doesn't publish per-answer or per-video figures for its own tools.

Can I find out exactly how much water or energy my AI video tool used?

Not currently, no. Almost no consumer-facing AI video company publishes per-generation energy or water figures, so the numbers available are independent estimates from researchers and journalists rather than official data from the companies themselves.

Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works

Published and maintained by the Lilach Bullock team, covering marketing, AI and business growth.
Your buyers are asking AI who to use. Does it say you?

See for free whether ChatGPT, Claude, Perplexity, Gemini and Google name you, and get the plan to become the answer.

Check my AI visibility →
Sundays only

Get the Sunday newsletter.

One email a week. AI experiments, marketing tactics, and the workflows Lilach is building right now in her own business.

Subscribe free

Let’s get your marketing running on AI.

Book a free 30-minute call

We figure out what you need, where AI fits in, and what working together would look like.

Book the call →

Or take the 30-second calculator

You’ll see the hours and the money quietly leaking out of your week, and the three workflows worth building first.

Take the calculator →

Or grab the free AI resource library

Prompt packs, templates, checklists, and swipe files. The exact tools I build for paying clients. Yours, free.

Get the library →
Keep reading

More from the blog.