A Badly Framed Question: How Much, Rather Than What The public debate on artificial intelligence and energy focuses on volumes: how many terawatt-hours data centers will absorb, how many power plants

A Badly Framed Question: How Much, Rather Than What

The public debate on artificial intelligence and energy focuses on volumes: how many terawatt-hours data centers will absorb, how many power plants will need to be built. This reading overlooks a decisive variable. The climate footprint of a digital infrastructure depends first and foremost on the electrical grid to which it is connected. An identical server, running the same model, does not have the same carbon footprint depending on whether it runs on a system dominated by coal and gas or on a mix rich in hydropower, wind, and solar.

The data available for August 2026 do not directly measure data center consumption. They do, however, describe with precision the electrical systems of three very different countries: China, Turkey, and Argentina. They therefore make it possible to answer an upstream question: which grids would a new demand, digital or otherwise, connect to, and at what carbon cost?

A Method Based on Real Flows

The analysis relies on four families of monthly indicators: electricity demand, generation broken down by source, power sector emissions by source, and the average carbon intensity of the kilowatt-hour. The percentages and ratios presented here are calculated from these raw values. The datasets provide neither the share of data centers in demand nor projections linked to artificial intelligence. Any extrapolation therefore remains qualitative.

China: A Grid on the Order of a Terawatt-Hour per Month

In August 2026, Chinese electricity demand reached 1,020.1 TWh in a single month. Generation totaled 1,021.99 TWh, a slight surplus of about 1.89 TWh. At this scale, China consumes roughly 29 times more electricity than Turkey and nearly 80 times more than Argentina over the same period.

This gigantism has a direct consequence for any new energy-hungry infrastructure. A demand addition that would destabilize a small grid represents, at China's scale, a marginal variation. The absorption capacity exists. The question concerns the nature of the electrons that will serve this demand.

A Mix Still Dominated by Fossil Fuels

Of the 1,021.99 TWh generated, 592.83 TWh came from fossil sources, about 58% of the total. Renewables supplied 382.46 TWh, roughly 37.4%. Within this block, wind and solar accounted for 219.32 TWh, nearly 21.5% of generation. Hydropower, bioenergy, and other renewables added 163.13 TWh, about 16%.

The sum of fossil and renewables leaves a remainder of about 47 TWh, not broken down in the categories provided. Emissions confirm the existence of this low-carbon generation outside renewables: the "clean" category shows 14.18 MtCO2, against 13.91 MtCO2 for renewables alone.

More than one-fifth of Chinese electricity now comes from wind and solar. That is considerable in absolute terms: 219.32 TWh in one month exceeds Turkey's total demand several times over. The structure nonetheless remains marked by a majority fossil dependence.

The Carbon Weight of the Chinese Kilowatt-Hour

China's power sector emitted 562.2 MtCO2 in August 2026. Fossil sources accounted for 548.02 MtCO2, about 97.5% of the total. Average carbon intensity stood at 550.1 gCO2/kWh, the highest of the three countries studied.

Relative to their generation, Chinese fossil plants emit about 924 gCO2 per kilowatt-hour. Renewables run at around 36 gCO2/kWh, a level consistent with life-cycle accounting. The gap between the two categories exceeds a factor of twenty-five. The entire challenge of digital electrification lies in this gap: depending on whether additional demand is covered by fossil margin or by new clean capacity, its footprint varies by a factor of twenty-five.

Turkey: A Largely Renewable Grid

The Turkish profile contrasts sharply. Demand reached 35.19 TWh in August 2026, for generation of 35.85 TWh. The country thus ran a surplus of about 0.66 TWh for the month.

More Than Half of Electricity Is Renewable

Renewables generated 20.29 TWh, about 56.6% of Turkish electricity. The "clean" category shows exactly the same value. Wind and solar contributed 9.78 TWh, about 27.3% of the total, a share higher than that observed in China. Hydropower, bioenergy, and other renewables provided 10.51 TWh, nearly 29.3%. Fossil fuels supplied only 15.57 TWh, about 43.4% of generation.

Turkey thus presents a mix inverse to China's: majority renewable versus majority fossil. The share of wind and solar, relative to total generation, is nearly six points higher.

An Intermediate Carbon Intensity

Turkish power sector emissions reached 14.26 MtCO2 for the month. Fossil sources concentrated 13.6 MtCO2, about 95.4%. Renewables generated only 0.66 MtCO2: 0.28 MtCO2 for wind and solar, 0.37 MtCO2 for hydropower and the like. Average carbon intensity came in at 397.88 gCO2/kWh, about 28% below the Chinese average.

The paradox is instructive. With majority renewable generation, Turkey maintains an average intensity close to 400 grams. The reason is arithmetic: its fossil plants emit about 873 gCO2 per kilowatt-hour generated. Even in the minority, they account for the bulk of the balance. A mix at 56% renewable does not guarantee a low-carbon kilowatt-hour as long as the fossil share relies on highly emitting technologies.

Argentina: The Least Carbon-Intensive Kilowatt-Hour in the Panel

Argentina operates at an even smaller scale. Its demand stood at 12.81 TWh in August 2026. The breakdown of generation by source is not available for this month in the data consulted, but emissions outline a clear profile.

Argentina's power sector emitted 3.33 MtCO2. Fossil sources accounted for 3.17 MtCO2, about 95.2%. Renewables weighed in at 0.16 MtCO2: 0.04 MtCO2 for wind and solar, 0.12 MtCO2 for hydropower, bioenergy, and other sources. Average carbon intensity fell to 278.68 gCO2/kWh. That is roughly half the Chinese level and 30% below the Turkish level.

In all three countries, fossil fuels concentrate more than 95% of power emissions. This constant recalls an obvious fact too often forgotten in discussions about the digital footprint. The carbon in a kilowatt-hour comes almost entirely from the fossil fraction of the mix. Reducing the footprint of a data infrastructure therefore means less "buying green" than avoiding, at the margin, calling on thermal plants.

What These Grids Imply for Data Infrastructures

One Terawatt-Hour, Three Carbon Balances

A simple calculation illustrates the stakes. An additional terawatt-hour consumed at the average intensity of August 2026 would represent about 0.55 MtCO2 in China, 0.40 MtCO2 in Turkey, and 0.28 MtCO2 in Argentina. The ratio between the most carbon-intensive and the least carbon-intensive grid in the panel approaches two to one.

This average-based reasoning probably underestimates the real impact of new demand. When a grid must serve an additional load, it is often dispatchable, frequently fossil, capacity that adjusts its output. At the fossil plant emission rate calculated above, the same marginal terawatt-hour would weigh nearly 0.92 MtCO2 in China and about 0.87 MtCO2 in Turkey. At the margin, Turkey's advantage therefore shrinks sharply. For a data center, the grid's average intensity is a misleading indicator if not supplemented by an analysis of the marginal source.

Grid Size Changes the Nature of the Risk

The scale of electrical systems also changes the nature of the risk. In China, with more than 1,000 TWh monthly, the arrival of large computing sites does not threaten grid balance. It above all raises a climate trajectory question, since each terawatt-hour adds to a system already emitting 562.2 MtCO2 per month.

In Turkey, with about 35 TWh monthly and a surplus of 0.66 TWh in August, the margin is narrower. A large data infrastructure project there would quickly become visible in the national balance. It could then compete with other uses for access to renewable capacity. Argentina, with 12.81 TWh of monthly demand, is in an even more sensitive position: any concentrated industrial consumption weighs proportionally heavy there.

Seasonality: The Blind Spot of Monthly Analysis

The figures analyzed cover a single month, August, in the height of summer in the northern hemisphere and winter in the southern hemisphere. Wind, solar, and hydropower generation varies sharply by season. A data infrastructure, by contrast, operates almost continuously. A site connected to a highly renewable grid in summer may therefore depend more on fossil fuels at other times of the year. Serious evaluation of a project requires complete series, not a snapshot.

Three Levers for Climate-Compatible Digital Growth

Choose the Grid Before Choosing the Site

The measured gaps between 278.68 and 550.1 gCO2/kWh show that location is a first-order lever. For infrastructure operators, the grid's carbon intensity should rank among siting criteria, alongside land prices, water availability, or latency. For states, attracting data centers to a highly emitting grid amounts to importing an emissions increase, regardless of how efficient the machines are.

Add Rather Than Reassign

The Turkish case shows that a majority renewable mix can coexist with high intensity when the fossil fraction is highly emitting. If a new infrastructure merely contracts for renewable electricity already generated, it reassigns green electrons without reducing the fossil share called upon by the rest of the system. Only additional clean capacity, brought online in step with new demand, prevents digital growth from translating into more thermal generation.

Steer Consumption Over Time

The third lever is flexibility. Some computing tasks, such as model training or batch processing, can be deferred. Shifting them to hours when wind and solar are abundant would reduce the call on marginal plants. In China, where 219.32 TWh of wind and solar were generated in August 2026, this synchronization could mobilize a considerable resource. It nonetheless requires price signals and steering tools that the available data do not allow us to assess.

Conclusion: The Electrical Geography of Artificial Intelligence

The rise of artificial intelligence will inevitably be electric. The figures for August 2026 nonetheless remind us that electricity is not a homogeneous resource. Between China, where each kilowatt-hour emits an average of 550.1 grams of CO2, Turkey at 397.88 grams, and Argentina at 278.68 grams, the same calculation does not leave the same trace in the atmosphere.

In all three countries, the fossil share concentrates more than 95% of emissions. It therefore sets the bulk of the climate cost of any new demand. The decisive question is not the volume of electricity that future data infrastructures will consume. It is the share of that consumption that will be covered by genuinely new clean capacity. Without this condition, the digital revolution risks extending the lifespan of the plants it claims to render obsolete.

This article was written with the assistance of artificial intelligence and verified by our editorial team.