Data Center CO₂ Projections

Interactive projection

Pick a state and year. The model evaluates both fits at that year and returns the product.

Projected total emissions

Log scale. Dashed lines are the fitted model. Markers are observed MW multiplied by a hybrid emission rate.

MW capacity

Data-center nameplate capacity by state.

Emission intensity

Grid CO₂ output rate, in pounds per MWh.

Methodology

Total emissions depend on two things: installed data-center capacity, and the emission intensity of the grid that supplies it. Both look roughly exponential over the years we have data for, so we fit each separately and multiply them.

1. Capacity fit

MW nameplate data for 2021 through 2024, together with a published 2030 industry projection, is fit to the model:

MW(t) = a · e^b(t−2021)

We use SciPy's curve_fit for this, which runs a nonlinear least-squares fit. Anchoring the exponent at 2021 makes a equal to the 2021 capacity, so the parameters are easier to read directly.

2. Emission-intensity fit

Emission intensity uses the same exponential form, anchored at 2018 to match the start of the data. EPA eGRID state values from 2018 through 2023 are fit to:

E(t) = a · e^b(t−2018)

Because emission intensity is generally falling over time, the solver needs a negative starting guess for b, such as p0 = (E₂₀₁₈, −0.1). A non-negative starting value can cause the fit to flatten out for states whose grids are decarbonizing quickly. Illinois is the clearest example.

3. Annualizing

MW is a rate, not an energy quantity. To get annual energy we multiply by 8,760, the number of hours in a year. Data-center load runs close to flat around the clock, so this is a reasonable first-order estimate:

MWh(t) = MW(t) · 8760

4. Total emissions

The total CO₂ in a given year is the product of annual energy and the corresponding emission intensity:

Total CO₂(t) = a_MW · e^{b_MW(t−2021)} · 8760 · a_E · e^{b_E(t−2018)}

Both factors are exponential, so the product is too, with a growth rate equal to the sum of the two component rates. Virginia's combined rate is large and positive. In the other four states, the falling emission intensity partly cancels out the capacity growth.

5. Fitted parameters

Fitted coefficients from curve_fit for each state:

State	a_MW	b_MW	a_E	b_E

6. Caveats

We only have five years of MW data and six years of emission data per state, so the fitted growth rate has a wide margin of error.
The 2030 MW number comes from an industry forecast, not a real measurement, which means it has an outsized effect on where the projection lands.