AI Carbon Footprint Statistics You Should Know
On this page
Good AI carbon footprint statistics are hard to find because so many circulating numbers are either exaggerated or stripped of context. The most useful figures are estimates that vary by model and region, not single headline claims. Treated that way, they are genuinely informative.
This roundup collects the canonical numbers we use to reason about AI impact: energy per token by model class, carbon per query, data-center overhead, water intensity, and the cost of generating images. Every figure below is an estimate meant for order-of-magnitude thinking, not a precise measurement.
Energy per token by model class
The starting point for any AI carbon footprint statistics is energy per token, because tokens are the unit of work. Small, fast models spend roughly 0.0008 Wh per token. Mid-size models, a reasonable baseline, use around 0.0015 Wh. Frontier or large models climb to about 0.0038 Wh, and reasoning-heavy models sit near 0.0040 Wh as they generate long internal chains of thought.
The headline pattern is that model choice changes the footprint of an identical task by roughly five times. That single fact is more actionable than any absolute number, because it means switching to a right-sized model is often the biggest lever you have. Our breakdown of AI energy usage by model expands on these classes.
- Small or fast model: about 0.0008 Wh per token
- Mid-size baseline: about 0.0015 Wh per token
- Frontier or large model: about 0.0038 Wh per token
- Reasoning-heavy model: about 0.0040 Wh per token
Carbon per query
To turn energy into carbon, you multiply electricity used by the grid's emissions factor. The US average grid sits around 0.395 kg CO2 per kWh, though a clean regional grid can be a fraction of that and a coal-heavy one much higher. Applying this to a normal answer gives a footprint that is well under a gram to a few grams of carbon dioxide.
That range surprises people who expect a dramatic figure. The reason it stays small is that a single answer is only thousands of tokens, and each token is a tiny fraction of a watt-hour. For a fuller derivation see how much CO2 does ChatGPT produce, and try your own inputs in the AI carbon footprint calculator.
Data-center overhead and water
Chips are not the whole cost. Power usage effectiveness, or PUE, captures the extra energy spent on cooling, power conversion and networking, and averages around 1.56 across the industry. Multiplying raw compute energy by this figure gives a more honest total for a facility.
Water is the other systemic cost. Between on-site cooling and the water used at off-site power plants, data centers consume roughly 3.4 liters per kWh. Like carbon, this varies widely with climate and cooling design. We unpack the mechanics in what is PUE in data centers and how much water do data centers use.
Image generation statistics
Images deserve their own numbers because they are far more expensive than text. A standard 1024x1024 generation costs roughly 2.0 Wh, and a higher-quality render about 3.5 Wh. In token terms, one image is comparable to hundreds or low-thousands of text tokens, so a handful of images can outweigh a long conversation.
This is why any realistic footprint estimate has to weight image and video generation heavily. If your workflow leans on visuals, that is where your statistics should focus. See how much energy does AI image generation use for the detail.
How to read these numbers responsibly
The most common mistake with AI carbon footprint statistics is treating an estimate as a fact. Every figure here depends on the specific model, the region's grid, the hardware generation and the facility's efficiency. Ranges are honest; single precise decimals usually are not.
Used well, these numbers let you compare choices rather than pass judgment. They show that model selection matters about five times over, that grid cleanliness can swing carbon dramatically, and that images cost more than chats. For the wider evidence base, AI carbon footprint statistics is complemented by environmental impact of AI as a pillar overview.
The headline: The most useful AI carbon footprint statistics are per-token energy by class, grams of CO2 per query, PUE around 1.56 and water near 3.4 liters per kWh, all read as estimates that vary by model and region.
FAQ
How many grams of CO2 does one AI query produce?
A typical text answer produces well under a gram to a few grams of carbon dioxide on an average grid. The exact figure depends on the model class, the length of the response and how clean the local electricity is. Images cost far more per item.
What is a reasonable energy figure per token?
As a baseline, a mid-size model uses about 0.0015 Wh per token. Small models are closer to 0.0008 Wh and frontier or reasoning models around 0.0038 to 0.0040 Wh. Model choice changes the total by roughly five times.
Why do these statistics use ranges instead of exact numbers?
Because real energy and carbon depend on the model, hardware, region and facility efficiency, no single exact number applies everywhere. Ranges and estimates reflect that variance honestly. Anyone quoting a precise decimal as universal is oversimplifying.
What is PUE and why does it matter?
PUE, or power usage effectiveness, measures how much total energy a data center uses relative to the energy that reaches the chips, averaging around 1.56. It captures cooling and power-delivery overhead. Multiplying compute energy by PUE gives a truer facility total.