What this page cannot do
No company has published the energy cost of generating an image or a video. Not one. So every
figure here that isn't about text is something I worked out myself, from academic measurements of
open models.
It also has to guess what you're asking for from the words you type, which is why it shows you
the words it used.
Batching is the bigger hole. Every measurement behind this was taken one request at a time; real
services run hundreds at once, and the single study that measured the difference found a factor of
thirty-six. Your video figure is probably too high. I can't tell you by how much.
Whose water is it, though? Your prompt went to a machine that was already switched on and already
being cooled. Someone divided that building's water bill by the number of requests and called your
share yours. That's bookkeeping.
And it rots fast. Google says it cut the energy of a single prompt thirty-three fold in twelve
months. These numbers were checked again in August 2026 and won't survive the year.
Where the number comes from
There is only one sum in this field and everybody uses it. Take the electricity the task used.
Multiply it by how much water gets consumed per unit of electricity. That is the whole method.
Two things can go into that second number. There is the water evaporated inside the building to
cool the servers, and there is the water consumed at the power station making the electricity. Which
of the two you count is part of what the four figures above are arguing about, though it turns out to
explain less of their disagreement than you would expect.
For text, the length of the answer is what costs. A longer request usually earns a longer answer,
so the figure climbs as you write. For an image or a video it barely moves, because the cost is in
the pixels and the frames, not in how nicely you asked.
One last thing worth knowing. Water consumed means water that evaporated and never came back.
Water withdrawn means water borrowed from a river and returned warm. The two differ by roughly
fourteen times and get mixed up constantly. Everything here is the first kind.