The Carbon Footprint of Training AI Models
On this page
Every large language model starts with training: a compute-heavy phase where the model learns patterns from enormous datasets across thousands of processors running for weeks or months. This upfront burst of electricity is what most headlines mean by the training AI carbon footprint, and for a frontier model it can reach into the hundreds of tonnes of CO2.
But training happens once. After that, the model answers questions billions of times, and each request carries its own smaller energy cost. Understanding how these two phases compare is the key to judging AI's real environmental impact, and to knowing where your own choices actually matter.
What drives the training AI carbon footprint
Training a model means repeatedly adjusting billions of internal parameters until predictions improve. Doing that requires specialized chips running near full power, often for weeks, inside a data center that also spends energy on cooling and power delivery. That overhead is captured by power usage effectiveness, or PUE, which averages around 1.56 in modern facilities: for every watt-hour the chips draw, roughly half a watt-hour more goes to cooling and losses.
The bigger the model and the more data it sees, the more energy the run consumes. Multiply the electricity by the local grid intensity, about 0.395 kg CO2 per kWh on the US grid, and you get the carbon cost. Frontier training runs are large enough that companies increasingly report them in sustainability disclosures. For the full picture of where that power goes, see our guide on how much electricity AI uses.
- Model size: more parameters mean more calculations per step
- Dataset size: more tokens mean more steps overall
- Hardware efficiency: newer chips do more work per watt
- Grid mix: the same run emits far less on clean power
Training is one-time, inference repeats forever
The crucial distinction is frequency. Training is a single event. Once it finishes, the model is copied and served, and it never needs to be trained again unless the team builds a new version. Inference, the act of running the finished model to answer a prompt, happens every single time someone uses it.
A typical text answer costs well under a gram to a few grams of CO2, tiny next to a training run. But that small number is paid over and over. A popular model fields hundreds of millions of requests a day, and those grams accumulate quickly. We break the split down further in training vs inference energy use.
How the two costs amortize
To compare fairly, you can spread the one-time training cost across every request the model ever serves. This is amortization. Divide total training emissions by the number of inferences over the model's lifetime, and the per-request share of training shrinks dramatically.
Imagine a training run that emitted a few hundred tonnes of CO2. Spread across a single day of heavy usage it would look enormous. Spread across billions of lifetime requests, the training share of any one answer falls to a rounding error. The longer a model stays in service and the more it is used, the smaller its amortized training footprint per query becomes.
Why inference dominates the lifetime footprint
Because inference repeats indefinitely, its cumulative energy eventually overtakes the training spike for any widely adopted model. Industry estimates commonly attribute the majority of a deployed model's lifetime energy to serving users rather than to the original training. The exact split depends on popularity: a niche model used rarely may never repay its training cost, while a mass-market assistant crosses that line quickly.
For everyday users, this is the practical takeaway. You cannot change how a model was trained, but you influence inference every time you send a prompt. Choosing a right-sized model instead of a frontier one can swing the energy of a single answer by roughly five times. That is why per-request measurement, not training statistics, matters most for individuals. Our how it works page shows how each request is measured.
Reducing the footprint you control
You will rarely train a model yourself, so your leverage is on the inference side. Pick a smaller model when the task is simple, avoid regenerating answers you do not need, and keep prompts focused. To see the numbers for your own usage, try the AI carbon footprint calculator.
Platforms can also close the remaining gap. Ecoia measures the energy, carbon and water of every request and retires verified offsets beyond 200 percent of the measured impact, making each answer net negative rather than merely neutral. That covers the amortized slice of training as well as the inference you trigger directly.
The headline: Training AI is a huge one-time cost, but because inference repeats billions of times, serving users dominates a popular model's lifetime carbon footprint.
FAQ
Is training or inference worse for the environment?
Training is far more energy-intensive as a single event, but it only happens once. Inference is much smaller per request, yet it repeats billions of times for a popular model. Over a widely used model's lifetime, inference usually accounts for the larger share of total energy.
How much CO2 does training a large model produce?
Estimates vary widely by model size, hardware and grid, but frontier training runs can reach into the hundreds of tonnes of CO2. Smaller or more efficient models cost far less. Because training happens once, this figure is best understood spread across the billions of requests the model later serves.
Can I reduce the training footprint of the models I use?
Not directly, since training is done before you ever touch the model. What you can influence is inference: choosing right-sized models and prompting efficiently. On Ecoia, offsets beyond 200 percent of measured impact also cover the amortized share of training tied to your usage.
Why do headlines focus on training instead of inference?
Training produces a single dramatic number that is easy to report. Inference is spread across countless small requests, so its cumulative cost is less visible even though it often ends up larger. Looking at both gives a more honest picture.