Consumption forecasts are demand statements. Generation has to be built. This report follows the supply side of the data centre buildout to 2035 across three cases and shows why the near-term fuel mix is dirtier than the headline renewable share suggests.
Key findings
- Generation to serve data centres rises from 460 terawatt hours in 2024 to over 1,000 by 2030 and 1,300 by 2035 in the base case.
- Renewables meet nearly half of additional demand to 2030 globally, but in China coal remains the largest single source of additional supply across that period.
- Europe is the cleanest major region, with renewables and nuclear reaching about 85 percent of data centre supply by 2030, and also the slowest growing.
- The spread between the high efficiency and headwinds cases is more than 40 percent by 2035. Planning against a single point from that range discards most of the information.
- Direct emissions grow from about 180 to 300 million tonnes by 2035 in the base case. The IEA notes AI adoption could reduce emissions elsewhere by more than that, which we treat as an upper bound rather than a forecast.
Generation, not capital, is the constraint
QR-03 established how much electricity data centres will consume. This report asks the question that follows and gets much less attention: where the electricity comes from. Consumption forecasts are demand statements. Generation has to be built, permitted, interconnected and paid for, and the mix that gets built determines both the cost and the emissions of the whole exercise.
Source: IEA, Energy and AI: energy supply for AI. Generation exceeds consumption because of network losses and own use. Renewables meet nearly half of the additional demand over the five years to 2030.
Generation to supply data centres is projected to grow from 460 terawatt hours in 2024 to over 1,000 by 2030 and 1,300 by 2035 in the base case. Over the next five years renewables meet nearly half of the additional demand. That headline is accurate and it conceals a regional picture that is considerably less clean.
The mix is a regional question, not a global one
China
Between 2024 and 2030, coal remains the largest single source of additional electricity for Chinese data centres, with annual generation from coal increasing by nearly 90 terawatt hours. Renewables, mostly solar and wind, add nearly 90 terawatt hours over the same period, supported by a rising renewable share in the grid mix, provincial co-location mandates and policies steering data centre construction toward the renewables-rich west of the country.
After 2030 the introduction of small modular reactors significantly lifts the nuclear share, and between 2030 and 2035 the combined rise in renewables and nuclear pushes coal into decline. By 2035 renewables and nuclear together make up nearly 60 percent of Chinese data centre electricity supply.
Europe
Renewables and nuclear are set to supply most of the additional electricity required, with their combined share rising to around 85 percent by 2030. Europe has the cleanest projected data centre supply of any major region, and also the smallest projected load growth, which is not a coincidence.
Japan and Korea
Together they account for about 5 percent of global data centre electricity demand today and are expected to retain roughly that share.
Renewables meet most of the additional demand to 2035. Fossil fuels meet a great deal of the near-term surge to 2030. Both statements are true and only one of them tends to get quoted.
How wide the uncertainty actually is
Source: IEA, Energy and AI: energy supply for AI. The high efficiency case is more than 15 percent below base; the headwinds case more than 40 percent below. Renewables meet 55 percent or more of the increase to 2030 in both.
The spread across cases is the most useful thing in this dataset. In the high efficiency case, generation reaches about 1,100 terawatt hours by 2035, more than 15 percent below the base case. In the headwinds case it reaches 790, more than 40 percent below. Across all cases renewables meet 55 percent or more of the increase to 2030.
A 40 percent spread at a ten-year horizon is honest forecasting rather than weak forecasting. It reflects genuine uncertainty in three places: efficiency improvement, AI adoption, and the physical delivery of generation and grid. Anyone planning against a single number from this range is discarding most of the information.
The emissions arithmetic, both directions
Source: IEA, Energy and AI. The IEA also notes that widespread adoption of existing AI applications could produce emissions reductions elsewhere in the economy that outweigh the direct emissions of the data centre sector.
Emissions from data centre electricity use grow from around 180 million tonnes today to 300 million tonnes in the base case by 2035, and up to 500 million tonnes in the lift-off case. Those are not large numbers in the context of global energy emissions, and treating them as an existential climate question overstates the direct effect.
The IEA also makes the countervailing point explicitly: widespread adoption of existing AI applications could deliver emissions reductions elsewhere in the economy large enough to outweigh the sector's own emissions. We would treat that as a plausible upper bound rather than a forecast, because it requires adoption in exactly the industrial and logistics settings where diffusion has historically been slowest. But it belongs in the calculation, and it is usually left out.
Who bears the cost
The commercially important question is not how much generation gets built but who pays for it. Generation, transmission and interconnection are largely recovered through regulated tariffs across a rate base. If new supply is procured specifically for data centre load but recovered generally, the cost is socialised across all consumers on that network.
That is an allocation choice rather than an economic necessity, and it is the variable we would watch. Three mechanisms determine it: long-term capacity contracting requirements placed on large loads, cost causation rules in rate cases, and the extent to which operators self-supply. Each shifts the burden between operator and ratepayer without changing the total.
Our expectation is that this becomes a political question before it becomes an economic one, and that it does so first in the jurisdictions with the highest regional concentration rather than the highest national load.
Source register
| Series or claim | Issuing body and vintage | Source link |
|---|---|---|
| Generation of 460 TWh in 2024 rising above 1,000 TWh in 2030 and 1,300 TWh in 2035 | IEA, Energy and AI, energy supply chapter | iea.org |
| Renewables meet nearly half of additional demand over five years; 55 percent or more of the increase to 2030 across cases | IEA, energy supply for AI | iea.org |
| China: coal adds nearly 90 TWh to 2030, renewables nearly 90 TWh, SMRs after 2030, renewables and nuclear near 60 percent by 2035 | IEA, energy supply for AI | iea.org |
| Europe: renewables and nuclear combined share rising to 85 percent by 2030; Japan and Korea at about 5 percent of global demand | IEA, energy supply for AI | iea.org |
| High efficiency case about 1,100 TWh by 2035; headwinds case 790 TWh | IEA, energy supply for AI | iea.org |
| Emissions of 180 Mt today, 300 Mt base case by 2035, up to 500 Mt lift-off; offsetting reductions from AI adoption | IEA, Energy and AI | iea.org |