Sustainability reports increasingly ask about AI, and most organisations answer with adjectives. "Minimal impact." "Significant footprint." Neither survives an auditor's follow-up question: measured how?
For AI-generated images, there is a measured baseline, and a simple model lets you turn it into a number for your own operations.
The Measured Baseline
Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell measured inference energy across AI tasks in Power Hungry Processing, published at ACM FAccT 2024. Their headline finding: multi-purpose generative architectures are orders of magnitude more expensive per inference than task-specific systems. The full paper puts image generation at a mean of 2.907 kWh per 1,000 images, with the least efficient model at 11.49 kWh per 1,000, about half a smartphone charge per image.
Those measurements ran on a shared node of eight A100 GPUs, with the idle power of the unused GPUs included. That is a fair picture of a cloud service. It is not a picture of a single workstation.
A Model You Can Run in a Spreadsheet
Energy is power multiplied by time. For a team generating locally, the inputs are the GPU's rated power, the average seconds per image and the electricity price. NVIDIA rates the RTX 4090 at 450 W, and the U.S. EIA put the average residential rate at 18.31 cents per kWh in July 2026.
| Seconds per image | kWh per 1,000 images at 450 W | Cost at 18.31 cents per kWh |
|---|---|---|
| 5 | 0.63 | about $0.11 |
| 10 | 1.25 | about $0.23 |
| 30 | 3.75 | about $0.69 |
Rated power is a ceiling, and the rest of the machine adds load, so treat these as estimates to replace with measurement. Commercial tariffs differ from residential ones; use your own.
What the Model Tells a Leadership Team
- The energy cost of AI images is real but small per unit. At these rates, a heavy year of 73,000 images at ten seconds each is roughly 91 kWh on a single card, about $17.
- Volume is the lever. Energy scales directly with images generated and seconds per image. Wasted renders are the largest controllable cost.
- Location changes the overhead, not the physics. Shared infrastructure carries idle and cooling overhead that a single workstation does not, and a workstation carries its own idle draw when unused.
Reduce Waste, Not Just Watts
The most effective energy policy for creative AI is a workflow policy: fewer discarded generations. That means sensible defaults rather than inflated quality settings, controls that fix composition before rendering, and records that let a good result be reused instead of rediscovered. Those principles shape Vision Studio, which starts every model at 25 steps, guides layout with ControlNet, and keeps an iteration tree of every render's settings.
Cloud Versus Local Is the Wrong First Question
Energy debates about AI often start with where the computation happens. That matters, but it is the second question. The first is how much computation the work requires, and that is set by workflow far more than by venue.
Consider two teams producing the same 1,000 approved images a month. One accepts roughly one in three generations. The other, with better defaults and better control tools, accepts one in ten after hours of re-rolling. The second team uses more than three times the energy for the same output, wherever it runs. No hosting decision closes a gap that size.
Venue still matters at the margins. Shared infrastructure carries overheads, which is part of why the measured cloud mean includes idle GPUs. A workstation carries its own overheads: it draws power while idle and is often left on. Neither is free, and the honest comparison measures both rather than assuming one.
What Belongs in a Sustainability Report
If AI content production is material enough to report on, report it like any other operational energy use:
- Activity data: images and video clips generated, and the share that were used.
- Energy per unit: measured on representative hardware, with the method stated.
- Total energy: activity multiplied by energy per unit, by venue where both cloud and local are used.
- Efficiency actions: the workflow changes made to reduce discarded generations, and their measured effect.
The last line is the one most reports leave out and the one that shows management, rather than just measurement.
Where Procurement Fits
Hardware choices do affect energy, mostly through time. A card that renders an image in five seconds uses less energy per image than one that takes thirty at similar power, which is why the cheapest workstation is not always the lowest-energy one. When the model you need does not fit in a card's memory, offloading can multiply render time, and energy with it. Specify hardware for the workload, and the energy profile follows.
Three Objections, Answered
"Our volumes are too small to matter." Possibly, and the model will tell you so in an afternoon. A measured small number is a better answer to a stakeholder than an assumed one.
"Cloud providers handle sustainability for us." They manage their facilities. They do not manage how many generations your team discards. Workflow efficiency is yours wherever the work runs.
"Rated power overstates real draw." Often true, which is exactly why the method ends with a wall meter. Use rated power to bound the estimate, and measurement to replace it.
The Number to Bring to the Board
Boards and sustainability committees rarely need precision to the watt. They need a defensible order of magnitude, a method, and evidence that someone is managing it. For most organisations generating images locally, that will be a small figure, tens of kilowatt-hours a year per heavy workstation on the estimates above, with a clear lever in workflow efficiency. Present it that way: the number, how it was measured, and what is being done to reduce waste. That is a stronger answer than either a reassuring adjective or an alarming one.
Make It Auditable
- Meter a representative workstation at the wall during a batch, and subtract idle.
- Divide by images produced to get your own energy per image.
- Multiply by annual volume and your actual tariff.
- Report the method alongside the number. A figure with a method survives scrutiny; an adjective does not.
The full creator-level breakdown, with the worked arithmetic is on the Vision Studio blog at vision-studio-x.com/blog/energy-bill-of-an-ai-image. Vision Studio is a free, open-source, local-first AI generation app built by Strategia-X. If you want a second pair of eyes on how your organisation adopts AI tooling, talk to us at strategia-x.com.
-Rocky
#AIEnergy #Sustainability #ESG #GenerativeAI #CostModel #LocalAI #StrategiaX #VisionStudio