Carbon, water, and cost

Prompt vs. camera

This tool lets you compare the environmental and financial costs of using top generative AI models to do the same work as a traditional film production. Comparisons are made using real global production data and the median estimated costs of AI-generated video.

One second of AI-generated video (Seedance 2.5 at 720p)
33Wh
22 gCO2e
70 mLwater
$0.23list price

Each second of video generated is the equivalent of about two full phone charges.

Calculator

Make it both ways

Pick what you're making and we'll compare shooting it for real against generating the same finished video with AI. Every number shows its likely range.

What are you making?

Adjust the details
In seconds

30 s

80:1

Seconds of AI video generated for every second that makes the final cut. Every discarded take still uses energy, so this setting moves the AI footprint more than any other.

What each side counts

Shot for real: the fuel, power, travel, hotels, food, sets, and waste of a typical shoot like this one, taken from industry carbon reports and scaled to your length.

Made with AI: every clip generated along the way (not just the ones that make the cut), including data-center power, cooling, and hardware. The cost also includes the people who prompt and edit.

Units: CO2e, or carbon dioxide equivalent, counts every greenhouse gas as the amount of CO2 that would warm the planet as much. Energy is in watt-hours (Wh), where a full phone charge is about 15 Wh and a kilowatt-hour (kWh) is 1,000 Wh.

Scenes

Scene by scene

Some scenes are expensive to shoot, and some cost almost nothing extra. Each comparison here only counts what the scene adds to a shoot that's already happening, using the AI model you picked above. The badge shows the lowest-carbon way to get the shot, or says it's too close to call when the likely ranges overlap too much to tell.

Access

Anyone can generate. Few people can shoot.

1,000,000
60 s

CO2e per year. Each gridline is 10 times the one before it. Teal is AI video, amber is filmmaking, violet is one person's whole year, and gray is other parts of the economy.

The bigger fix is the power itself

Almost all of AI video's footprint comes from electricity, and it's the same for VFX rendering, so a clean grid shrinks both directly. Wētā's Avatar rendering would've emitted about a fifth as much on New Zealand's grid as it did on Australia's. Clean power won't get rid of the water used for cooling or the carbon from manufacturing the hardware, but it would remove most of the rest.

Filmmaking can't get there on clean electricity alone. Crews and gear have to travel by plane and truck, and there's no electric replacement yet for most of the planes in service. Sets, food, and waste also carry their own emissions. Those can be cut, or properly offset with real reductions elsewhere instead of tree-planting pledges, but the industry is a long way from actually doing it.

So how the world generates its energy is what matters most. Cutting filmmaking's footprint, from discarded AI takes to diesel generators, limits the damage while that work gets done.

In practice

How AI filmmaking actually works

Nobody generates a shot just once. How many takes a production throws away changes the answer more than which model it uses.

How people workTakes per kept shotAI shooting ratioExample

Professionals mostly start with prepared images and animate them (a workflow called image-to-video). In long-form studio work, AI shows up mainly in previz, crowds, environments, and individual VFX shots. House of David generated about 20 takes for every AI shot it actually used.

ByteDance's own quality tests rate clips animated from images lower, clip for clip, than clips made from text alone. However, starting from prepared images helps keep shots consistent with each other, which those tests don't measure, so the calculator leaves the shooting ratio up to you.

"Unlimited" plans deserve suspicion. Artlist, Runway, Higgsfield, and Krea have all run unlimited-Seedance promotions, which make generating another take free to you while the energy it uses stays exactly the same. This site counts the seconds you generate, not the money you spend.

Money is a separate question. Generation is cheap, and people are the expensive part. Svedka's AI Super Bowl ad is the cautionary tale: the brand's CMO said it "didn't save much money and certainly not a ton of time."

Biggest lever

Draft first

65–80%

less energy per kept shot when you explore at 480p and render just the keepers at full resolution. A Seedance 2.5 draft costs 18% as much as a 1080p render. Cheap drafts can invite more drafts, though.

Use carefully

Generate small, upscale at home

0.7 Wh

for a Mac mini's processor to double the resolution of a 5-second clip with a simple upscaler (Real-ESRGAN), about a twentieth of a phone charge. AI upscalers that repaint detail, like SeedVR2, can use about 125 Wh for the same clip, which can cancel out the energy you saved by generating small.

Counterintuitive

Fast beats frugal

21 Wh

for a gaming PC with an RTX 4090 graphics card to make a 5-second clip with the open model Wan 2.1, compared to 50.5 Wh for a Mac mini's processor alone. The PC draws 20 to 30 times the power, but it finishes in under 3 minutes instead of 3 hours 25, so it uses less than half the energy.

Local or cloud?

Running a model on your own computer skips the data center, but that doesn't save much on its own. What really matters is how big the model is and how fast the machine finishes the job. The same small Wan model uses about 3.5 Wh per generated second on a data-center graphics chip, 4 Wh on a gaming PC with an RTX 4090, and 10 Wh on a Mac mini's processor alone. The big savings come from small, stripped-down models, and those have visibly lower quality.

Energy for each second of generated video. Each gridline is 10 times the one before. Teal runs in a data center. Violet runs on your own computer.

Local generation is worth it when all of these are true:

  • The model is small. LTX-Video (a small open model) made 5 seconds of video on the Mac mini for 0.57 Wh, a few hundred times less than Seedance 2.5 uses for the same length. A mid-size Wan model at 720p on a gaming PC uses about 14 Wh a second, as much as MiniMax H3 in the cloud.
  • The hardware is fast. A powerful graphics card that finishes in minutes beats a frugal chip that runs for hours, since the whole computer draws power for as long as the job runs.
  • The quality is enough for the job. Open models (which anyone can download and run) fall behind the cloud leaders once they're small enough to fit on a home computer. In blind side-by-side votes run by Artificial Analysis, viewers pick the leading cloud models (Seedance 2.0, MiniMax H3, and Wan 3.0) over the best of these home models (LTX-2.5 and LTX-2.3) about 70–80% of the time. That makes local models useful for previz, animatics, and look tests, and only marginally useful for finished shots. If a weaker model needs more takes, or you end up remaking the shot in the cloud anyway, the savings shrink or disappear.
  • Your power is cleaner. Seedance's international service runs on Malaysia's power grid, which emits about 600 grams of CO2e per kWh, against about 380 grams for the average US grid and 230 grams for Los Angeles's utility.

The cloud has efficient options too. MiniMax H3 Max (a stripped-down fast model) uses roughly 3 Wh a second and ranks with the best. The one open model in that top group, MiniMax H3, needs four data-center graphics chips, so it's not a home option.

Guided platforms like Krea

Krea offers every Seedance and MiniMax tier alongside cheap tools for exploring ideas. A real-time draft image costs about 1/2,000 as much as a Seedance clip (in Krea's own credits), and its visual workflow editor shows the cost of each step and skips steps that haven't changed. Settling on a shot with those cheap tools before paying for premium video probably saves energy. Nobody has measured it, though, and live video previews and automated assistants keep the servers busy for the whole session.

Voices and music

Voice and music models running on your own computer use under 2 Wh per minute of audio. Turning off a video model's built-in sound only saves the energy it spends on audio, about 21–34% of the work for LTX-2. Seedance 2.5 and MiniMax H3 generate sound and picture together and don't charge any less for silent clips.

Method

How the numbers work

No AI video company publishes energy usage per clip, so the AI side is always an estimate. Here's what those estimates are based on.

AI energy, carbon, and water

Our estimate starts with how long a clip keeps a data center's graphics chips busy. For Seedance 2.5 at 720p, we used three separate methods that all landed at 17–24 Wh per generated second for the chips alone. One measures how long clips take to run through OpenRouter (a service that resells access to AI models). One compares against Jegham, Gamazaychikov, and Luccioni's 2026 academic estimate for Seedance 1.5. And one works backward from ByteDance's reported 70% gross margin, which implies how much they spend running the service. We use 18 Wh. The MiniMax H3 figure comes from MiniMax's own speed test of the version anyone can download.

From there, we add the rest of each server, machines kept idle in reserve, and cooling (a 2025 Google study found a whole data center uses about 1.7 times the energy of its AI chips alone). Then we apply how much carbon the local power grid emits for each kWh, and add a share for manufacturing the hardware and 5% for training the model. Water counts cooling at the data center plus the water used to generate its electricity.

The ranges combine the uncertainty in each of these steps. The AI shooting ratio isn't part of the range because you set that yourself.

We timed MiniMax H3 through MiniMax's own service, which reports when each job actually starts. A 2K clip took about 1.7 times as long as a 768p clip, which is pretty close to the 1.5 times we assume, and 15-second clips took about 15% less time per second than 5-second clips. Timing Seedance 2.5 through resellers mostly just measured time spent waiting in line, so that can't pin down how much computing Seedance actually uses.

Measured on a Mac mini

The most useful result here is the counterintuitive one. For the one heavy model, Wan 2.1, the Mac mini's processor used 50.5 Wh for a 5-second clip. A gaming PC with an RTX 4090 running the same settings uses about 21 Wh for the whole computer, less than half, because it finishes more than 70 times faster. Small, stripped-down models were cheap on the Mac, though. LTX-Video made 5 seconds of video for 0.57 Wh.

To check the figures for running AI at home, we ran five open models on a base M4 Mac mini with 16 GB of memory using the free ComfyUI app, and we recorded the processor's power once a second with macOS's built-in power monitor. That leaves out memory, storage, the fan, and the power supply, so the whole computer uses more. Each figure is the average of up to five runs after a warm-up.

ModelOutputTimeProcessor energy
SDXL base 1.0, 30 stepsOne 1024×1024 image2.6 min0.69 Wh
Z-Image Turbo, 4-bit, 8 stepsOne 1024×1024 image3.1 min0.89 Wh
LTX-Video 2B distilled, 8 steps5 seconds at 768×5122.3 min0.57 Wh
Wan 2.1 1.3B, 30 steps5 seconds at 832×4803 h 25 min50.5 Wh
Real-ESRGAN 2× upscale5 seconds, to 1536×10242.9 min0.70 Wh

16 GB of memory is pretty tight. LTX-Video and a larger version of Z-Image ran short of memory and had to spill over onto the disk, so the image tests use a smaller, compressed version of Z-Image instead. Wan fit in memory, but each of its 30 refinement passes took almost 7 minutes on this chip, compared to about 5 seconds on an RTX 4090.

Conventional production

Our benchmarks come from the Sustainable Entertainment Alliance (US studios, 2023–24), BAFTA albert (UK), AdGreen (advertising), and Ecoprod's Carbon'Clap (France). They all measure different things. The US studio benchmark counts fuel, utilities, flights without their high-altitude warming effect, and hotels, while the others also count food, materials, waste, and post-production. So we raise the US studio figures by 1.3 times to close that gap. For indie features, we show two figures: a general estimate of 150 metric tons, and the French indie average of 271 metric tons, which comes from films with bigger budgets than most indies.

Industry benchmarks leave out VFX rendering, and it isn't small. Weta's cloud rendering for Avatar: The Way of Water alone came to roughly 8–28 GWh, which is about what 750–2,600 American homes use in a year. Rendering a single VFX shot takes about 0.15–5 MWh, or what one home uses in 5 days to 6 months.

Avatar also shows how much the circumstances change the answer. Weta is in Wellington, and 87% of New Zealand's electricity came from renewables in 2022. However, Amazon didn't have any cloud data centers in New Zealand then, so the cloud rendering ran in Australia. On Australia's grid, that comes out to roughly 4,800–16,200 metric tons of CO2e. The same work on New Zealand's grid would've been about 850–2,900 metric tons. Amazon says it bought enough renewable energy to match 90% of its electricity use in 2022, which lowers the figure if you count those purchases, but the servers still ran on Australia's grid.

The production side of those same films was unusually clean. Lightstorm ran the sequels' two Manhattan Beach stages entirely on its own 960 kW rooftop solar array and served 55,000 plant-based meals. That doesn't cancel out the rendering, but it shows how far a single production's choices can move its footprint, and how little an industry average tells you about any one film.

Factors
FactorCentralRangeSource
Numbers to stop quoting
  • MIT Technology Review's figure of 3.4 megajoules (about 940 Wh) for one five-second video is about 5 times ML.ENERGY's measurement of what appears to be the same model on the same kind of chip.
  • "Native 4K" for Seedance 2.5 shows up in press coverage, but ByteDance's own service tops out at 1080p.
  • Weta's 3.3 billion hours of processor time for Avatar: The Way of Water only covers the rendering done in Amazon's cloud, not the whole film.
  • Several "AI statistics" sites publish made-up production-carbon figures. None of them appear here.
Sources