It’s bothersome to research this every time I want to explain to someone the ecological costs of training and using AI in perspective, so I’m putting it here.
Training and using generative AI tools do have electrical and ecological costs, and they should not be ignored. However, when people use this fact to support a moral panic about using these tools, it is frequently ignored how these costs compare to things we already do with much higher costs. For example, training GPT 3 (the original Chat GPT) produced about as much CO₂ as the annual emissions of 123 cars. A small neighborhood’s worth of cars for a whole year to equal this one time cost for a model that changed the world.
It’s also frequently mentioned that the electricity used to generate 4 images is as much as charging a cell phone. But cell phone batteries are tiny. The electricity an average person might use to run their home’s AC for a summer is sufficient for producing 1 million such images. I’ve generated maybe 25,000 such images in the 2 years I’ve had access to them.
Every time someone eats a quarter pound hamburger, the production of that beef produced as much CO₂ (about 3kg) as producing roughly 2,500 images. In water cost, I couldn’t find a source for image diffusion, but that same quarter pound hamburger requires 1,700 liters of water to produce, enough to do thousands of Chat GPT sessions of 20-50 prompts per session.
CO₂ Emissions (metric tons)
| Activity | CO₂ Emissions (metric tons) |
|---|---|
| One-Time Training | |
| GPT-3 Training | ~552 |
| Inference | |
| LLM Queries (20–50 per session) | <0.0001 |
| Stable Diffusion (1,000 images) | ~0.0012 |
| Other Activities | |
| One Year of Home AC (US Avg.) | ~1.2 |
| Producing ~46 kg of Beef | ~1–1.5 |
Water Use (liters)
| Activity | Water Use (liters) |
|---|---|
| One-Time Training | |
| GPT-3 Training | ~700,000 |
| Inference | |
| LLM Queries (20–50 per session) | ~0.5 |
| Stable Diffusion (1,000 images) | Negligible |
| Other Activities | |
| One Year of Home AC (US Avg.) | Minimal |
| Producing ~46 kg of Beef | ~700,000 |
Electricity Use (kWh)
| Activity | Electricity Use (kWh) |
|---|---|
| One-Time Training | |
| GPT-3 Training | ~1,380,000 |
| Inference | |
| LLM Queries (20–50 per session) | ~0.1 |
| Stable Diffusion (1,000 images) | ~2.9 |
| Other Activities | |
| One Year of Home AC (US Avg.) | ~3,000 |
| Producing ~46 kg of Beef | N/A |
Notes/Assumptions:
- Electricity-to-CO₂ conversion assumes ~0.4 kg CO₂/kWh (typical US grid average; actual mixes vary).
- GPT-3 training emissions (552 metric tons CO₂) come from published estimates, and electricity usage is a back-calculation based on the assumed carbon intensity of electricity.
- “Minimal” or “Negligible” indicate comparatively very small or not well-reported values.
- Beef production water figure (~15,000 liters/kg) is a widely cited estimate and includes water for feed crops, not just direct consumption.
- Inference energy costs for LLMs and exact water usage per query are rough estimates.
Estimates (Using Previously Stated Figures)
CO₂ Comparison:
- Quarter Pound of Beef:
- Approximate beef usage: 46 kg beef = ~1–1.5 metric tons CO₂
- Per kg: ~21.7–32.6 kg CO₂
- Quarter pound (≈0.113 kg): ~2.5–3.7 kg CO₂ (let’s say ~3 kg CO₂ for a rough midpoint)
- Stable Diffusion Image Generation:
- 1,000 images ≈ 0.0012 metric tons CO₂ = 1.2 kg CO₂
- Per image: 1.2 kg CO₂ / 1,000 ≈ 0.0012 kg CO₂
- With ~3 kg CO₂ (from quarter pound of beef), you can produce:
3 kg CO₂ ÷ 0.0012 kg CO₂/image ≈ 2,500 images.
Water Comparison:
- Quarter Pound of Beef:
- 46 kg beef ≈ 700,000 liters
- Per kg: ~15,000 liters/kg
- Quarter pound (0.113 kg): ~15,000 × 0.113 ≈ 1,700 liters
- Stable Diffusion Image Generation:
- Water usage for stable diffusion was noted as “negligible” per 1,000 images. Even if we assume some minimal amount, it’s orders of magnitude less than beef production.
Given that stable diffusion’s water use is near zero per image, you could produce Thousands to Millions of images for the same ~1,700 liters of water it takes to produce a quarter pound of beef. In practical terms, the difference is so large that water usage for stable diffusion is effectively negligible by comparison.