Carbon, water, and cost

Prompt vs. camera

Compare making a scene with Seedance 2.5 or MiniMax H3 against shooting it for real, using published production data and the best estimates available for AI video.

One generated second of Seedance 2.5 at 720p
33Wh
22 gCO2e
70 mLwater
$0.23list price

That's about two full phone charges. The catch is that nobody generates a second only once.

Calculator

Make it both ways

Pick what you're making, how you'd generate it, and how you'd shoot it. The comparison updates as you go, and every number shows its likely range.

01

What you're making

Runtime in seconds

30 s

02

The AI version

How you work

80×

Every take you throw away still used a GPU. This ratio drives the AI footprint more than any other setting.

03

The conventional version

What counts

Counts flights with their high-altitude warming effect, food, materials, waste, and post. This is the fair match for AI's all-in compute number.

Scenes

Scene by scene

Some scenes are expensive to shoot, and some cost almost nothing extra. Each comparison counts only what the scene adds to a shoot that's already happening, and uses the AI model you picked above. Badges compare AI with the lower of the two other options.

Access

Anyone can generate. Few people can shoot.

1,000,000
60 s

Annual CO2e on a log scale. Each gridline is 10×. Teal is AI generation, amber is production, and violet is one person's whole year.

In practice

How AI filmmaking actually works

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

How people workTakes per kept shotGenerated seconds per finished secondExample

Professionals mostly start from 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 used.

Seedance's own benchmark rates image-to-video clips lower, clip for clip, than text-to-video. Prepared images help with continuity across shots, which benchmarks don't measure, so the calculator leaves the take ratio up to you.

"Unlimited" plans deserve suspicion. Artlist, Runway, Higgsfield, and Krea have all run unlimited Seedance windows, which make one more take free to you while its energy stays exactly the same. This site counts generated seconds, never spend.

Money is its own 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 only the keepers at full resolution. A Seedance 2.5 draft costs 18% of a 1080p render. Cheap drafts can invite more drafts, though.

Use carefully

Generate lower, upscale locally

<1 Wh

to upscale a 5-second clip with a lightweight model on a laptop. A diffusion upscaler such as SeedVR2 can use about 125 Wh for the same clip, which wipes out the saving.

Counterintuitive

Local isn't greener by itself

3 Wh

per FLUX image on a MacBook Pro M4 Max, against about 1.3 Wh on an RTX 4090 and under 1 Wh on a data-center GPU. Local tools help only when they replace cloud video generations.

Guided platforms like Krea

Krea hosts every Seedance and MiniMax tier alongside cheap exploration tools: a realtime image costs about 1/2,000 of a Seedance clip in Krea's own units, and its node editor shows the cost of each step and skips work that hasn't changed. If you use that layer to settle on a shot before paying for premium video, it plausibly saves energy. Nobody has measured it, though, and realtime video and agents keep GPUs busy for the whole session.

Voices and music

Local text-to-speech and music models use under 2 Wh per minute of audio. Turning off a video model's own audio saves only what the model spends on it, which is about 21–34% of compute for LTX-2. Seedance 2.5 and MiniMax H3 generate audio in the same pass and don't discount silent clips.

Method

How the numbers work

No video company publishes energy per clip, so the AI side is always an estimate. Here's what each estimate rests on.

AI energy, carbon, and water

Energy per generated second starts from GPU time. For Seedance 2.5, three methods agree at 17–24 Wh per second on the GPUs at 720p: OpenRouter's latency benchmark, a calibration against Jegham, Gamazaychikov, and Luccioni's 2026 estimate for Seedance 1.5, and ByteDance's reported 70% gross margin. We use 18 Wh. MiniMax H3 is anchored on MiniMax's own benchmark for the open-weight model.

We then add servers, idle capacity, and data-center cooling (Google's 2025 serving study found the full stack uses about 1.7× the GPU energy), apply the grid where the model runs, and add hardware manufacturing and a 5% share for training. Water counts cooling at the data center plus the water used to generate its electricity.

Ranges combine each factor's own uncertainty. The takes ratio isn't part of the range because you set it.

We timed MiniMax H3 through MiniMax's own API, which reports when each job starts running. A 2K clip took about 1.7× as long as 768p, close to the 1.5× we assume, and 15-second clips cost about 15% less per second than 5-second clips. Timing Seedance 2.5 through resellers measured mostly queueing, so it can't pin down Seedance's compute.

Conventional production

Benchmarks come from the Sustainable Entertainment Alliance (US studios, 2023–24), BAFTA albert (UK), AdGreen (advertising), and Ecoprod's Carbon'Clap (France). They 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. The scope switch adjusts for that gap.

Industry benchmarks leave out VFX rendering, and it isn't small. Weta's AWS rendering for Avatar: The Way of Water alone comes to roughly 8–28 GWh. The custom shoot builder lets you add VFX shots at 0.15–5 MWh each.

Factors
FactorCentralRangeSource
Numbers to stop quoting
  • MIT Technology Review's 3.4 MJ for one five-second video is about 5× higher than ML.ENERGY's measurement of what appears to be the same model on the same GPU.
  • "Native 4K" for Seedance 2.5 appears in press coverage, but BytePlus's API tops out at 1080p.
  • Weta's 3.3 billion thread-hours for Avatar: The Way of Water covers only the rendering done on AWS, not the whole film.
  • Several "AI statistics" sites publish invented production-carbon figures. None of them appear here.
Sources