How we estimate training emissions
Most labs do not say how much carbon it took to train their models. When a lab does, we print its figure and label it as theirs. When it does not, we estimate. An estimate is always a range, never a single number, and every input links to where we found it.
The formula
Emissions = compute ÷ (chip speed × utilization) × chip power × PUE × grid intensity.
Compute is the total arithmetic done in training, counted in FLOP. We take it from the lab when it discloses it, and from Epoch AI's published estimates when it does not. Dividing it by how fast one chip runs gives the chip-hours the training took.
Chip speed is the chip's rated peak from its maker. For NVIDIA's H100 we use thedense FP8 peak, 1,979 trillion FLOP per second, the same figure Epoch uses.
Utilization is how much of that peak a real training run gets. Epoch puts thetypical range at 30% to 50%. A run at 30% takes longer, so it is our high end.
Chip power is the chip's rated power from its maker: 700 watts for an H100.
PUE is the data center's overhead: cooling, power conversion and lighting on top of the computers. We use 1.1, about Google's fleet, at the low end, and 1.4, the U.S. average in 2023 per Lawrence Berkeley National Laboratory, at the high end.
Grid intensity is how much carbon the electricity carried. At the low end we use the lab's own market-based figure, which counts the clean power it buys, when it publishes one. At the high end we use the average of the grid, 384.4 grams per kWh for the U.S. in 2025 (Ember, via Our World in Data). A lab that publishes no figure gets the grid average at both ends.
The range
The low end takes every assumption at its most favorable: 50% utilization, PUE 1.1 and the market-based grid. The high end takes the least: 30%, 1.4 and the grid average. We round both to two significant figures, because the inputs are not more precise than that.
What it leaves out
We count the chips' power only, not the servers' processors, memory and networks around them, and training only, not the experiments before it, nor the building and chips themselves. Real runs often reach less than 30% of the chip's FP8 peak. Where a lab disclosed its actual chip-hours, the model page compares them with ours. For these reasons the true figure is more likely above our range than below it.
Comparisons
To make tons readable we compare them with gasoline cars driven for a year, at 4.29tons each, from the EPA's equivalencies calculator.