The Environmental Impact of AI: A Complete Guide
On this page
The environmental impact of AI is easy to misjudge in either direction. A single chat reply is genuinely tiny, well under a gram to a few grams of carbon dioxide, yet billions of requests a day add up to a footprint worth measuring. The honest picture sits between panic and dismissal.
This guide walks through the four channels that matter most: electricity, carbon emissions, water for cooling and power generation, and the hardware that eventually becomes e-waste. We stick to grounded estimates that vary by model and region, and we end with practical steps that individuals and teams can take today.
How AI turns your prompt into energy use
Every request follows the same chain: your words become tokens, tokens drive computation on specialized chips, and that computation draws electricity. A useful baseline is energy per token. Small, fast models spend roughly 0.0008 Wh per token, mid-size models around 0.0015 Wh, and frontier or reasoning-heavy models in the region of 0.0038 to 0.0040 Wh. Model choice alone can swing the footprint of the same task by about five times.
That chip-level draw is only part of the story. Data centers add overhead for power delivery, networking and cooling, captured by a metric called power usage effectiveness, or PUE, which averages around 1.56. In other words, for every unit of energy the chips use, roughly half a unit more is spent keeping the facility running. If you want the underlying mechanics, our how it works page traces the full path from token to offset.
Carbon emissions and water, not just electricity
Electricity becomes carbon when the local grid burns fossil fuels. On the US average grid at about 0.395 kg CO2 per kWh, a typical text answer lands well under a gram to a few grams of carbon dioxide. The same request run in a coal-heavy region emits far more than one run on hydro or solar, which is why location matters as much as model size.
Water is the quieter cost. Data centers use it directly for cooling and indirectly through the power plants that supply them, adding up to roughly 3.4 liters per kWh across on-site and off-site use. It is small per query but real at scale. We cover this in depth in how much water does ChatGPT use and how much water do data centers use.
Images and hardware add to the total
Text is cheap compared with pixels. Image generation is billed per image at roughly 2.0 Wh for a standard 1024x1024 output and about 3.5 Wh for higher quality. A single image is comparable to hundreds or even low-thousands of text tokens, so generating a batch of them dwarfs a normal conversation.
Beyond running costs, the accelerators and servers that power AI have finite lifespans. As models grow and hardware is refreshed, retired equipment becomes electronic waste that must be recycled responsibly. Manufacturing those chips also carries an upfront carbon and water cost that never appears in a per-query number but is part of the full lifecycle.
- Standard image: about 2.0 Wh per generation
- Higher-quality image: about 3.5 Wh per generation
- One image roughly equals hundreds to low-thousands of text tokens
- Hardware refresh cycles drive e-waste and embodied emissions
Putting the scale in context
Comparisons help. Bitcoin mining consumes somewhere around 100 to 150 TWh per year. AI inference is smaller than that today, but it is growing faster, so the trend line deserves attention even if the current absolute number is modest. Framing AI as either negligible or catastrophic both miss the point.
The practical takeaway is that per-request impact is small and highly measurable, while aggregate impact depends on how many people use these tools and how the electricity is generated. Because each step in the chain can be estimated, the footprint can be tracked rather than guessed. You can explore the arithmetic yourself with our AI carbon footprint calculator.
What individuals and teams can do
You do not need to stop using AI to shrink its footprint. The biggest lever is matching the model to the task: use a small or mid-size model for routine work and reserve frontier or reasoning models for problems that genuinely need them. Batching requests, writing tighter prompts, and reusing outputs instead of regenerating them all help. Our guide on how to reduce your AI carbon footprint collects the highest-impact habits.
Teams can go further by tracking usage, choosing providers that measure and report impact, and preferring platforms that offset beyond what they emit. Ecoia measures the energy, carbon and water of every request and retires verified carbon and water offsets past 200 percent of measured impact, making the service net negative. If you are evaluating options for an organization, green AI for business lays out how that works in practice.
The headline: AI's per-request footprint is small but real, spread across electricity, carbon, water and e-waste, and it can be measured, reduced and offset rather than merely guessed at.
FAQ
Is using AI bad for the environment?
A single request has a small footprint, typically well under a gram to a few grams of carbon dioxide for text. The concern is aggregate scale as usage grows, not any one query. Choosing efficient models and offset-backed providers keeps individual impact low.
Which uses more energy, AI or Bitcoin?
Bitcoin mining uses roughly 100 to 150 TWh per year, which is larger than AI inference today. However, AI energy use is growing more quickly, so the comparison is likely to shift over time. Both are worth measuring rather than assuming.
Does generating images use more energy than chatting?
Yes, substantially. A standard image costs around 2.0 Wh and a higher-quality one about 3.5 Wh, comparable to hundreds or thousands of text tokens. If you generate many images, that becomes the dominant part of your footprint.
Can AI's environmental impact actually be offset?
Yes, when the impact is measured accurately first. Ecoia estimates the energy, carbon and water of each request and retires verified offsets beyond 200 percent of that impact. Measurement is what makes credible offsetting possible.