An identical model serving an identical query can differ in carbon emissions by roughly an order of magnitude depending on which grid it runs on. Region, not model efficiency, is usually the largest single factor in the emissions of an AI workload.
Grid Carbon Intensity Varies Enormously
Carbon intensity is measured in grams of CO2 equivalent per kilowatt-hour. Grids dominated by hydro, nuclear, or wind sit in the tens of grams. Grids dominated by coal sit in the high hundreds. The spread across regions hosting significant data center capacity is roughly 10 to 20 times.
| Grid type | Typical carbon intensity | Example regions |
|---|---|---|
| Hydro or nuclear dominated | Very low, tens of gCO2e/kWh | Nordics, Quebec, France |
| High-renewable mixed | Low to moderate | Pacific Northwest, parts of Iberia |
| Gas dominated | Moderate to high | Much of Texas, parts of the UK |
| Coal dominated | High, several hundred gCO2e/kWh | Parts of Asia, some US regions |
Values move hour to hour with the generation mix, which is the crux of the accounting problem below.
Matched Versus Hourly Clean Energy
Most corporate clean-energy claims use annual matching: total renewable energy purchased over a year equals or exceeds total consumption. This is a real procurement commitment and it is not the same as running on clean power. A facility can be annually 100 percent matched while drawing coal-fired electricity at 2am on a still winter night, because the wind generation that balanced the books happened at a different time.
Hourly or 24/7 carbon-free energy accounting requires matching consumption to clean generation in every hour, which is far harder and much less commonly achieved. The gap between the two standards is where most of the disagreement about AI emissions lives.
Why AI Makes This Harder
Three reasons. AI load is large and relatively inflexible, so it cannot easily be shifted to hours when clean generation is abundant. Interconnection constraints mean capacity gets built where power is available rather than where power is clean, and the fastest-connecting regions are not the lowest-carbon ones. And rapid load growth in a region can pull dispatchable fossil generation back onto the grid that would otherwise have retired, making the marginal emissions of new AI load higher than the grid average suggests.
That last point matters for accounting: average grid intensity understates the emissions of new load, because new load is served at the margin.
Brand Visibility Implications
Emissions claims are contested territory where corporate statements and journalistic analysis frequently disagree, and AI assistants retrieve both. A company relying on annual matching will find its "100 percent renewable" claim answered alongside coverage explaining what annual matching does and does not mean. Precision helps here: organisations publishing hourly figures and marginal-emissions methodology tend to be represented more accurately than those publishing only the headline claim. See energy claims in AI answers.
Methodology
Figures compiled from Gartner, IDC, LBNL, grid-operator filings, utility rate cases, and press reporting through July 2026. Forecasts are cited to the forecaster because independent projections in this area diverge widely, and several of the underlying quantities are estimates rather than measurements. Where sources disagree, ranges are given rather than a single number. Updated quarterly.
How Presenc AI Helps
Presenc AI measures how a company's sustainability claims are represented in AI answers and which counter-sources are retrieved alongside them.