How Much Electricity Does AI Use? — Why Every Source Gives a Different Number

Same Model, Same Training, Two Numbers

24.7 tons and 50.5 tons. These two numbers are values attached to the same single subject. This is the amount of greenhouse gas emitted to make a language model released in 2022, which has 176 billion internal numbers—parameters—adjusted as it learns. The unit is tons of carbon dioxide equivalent, which is an amount combining several greenhouse gases into a single total based on the equivalent effect of so many tons of carbon dioxide.

Training is the process of making a model, and inference is using a fully made model once. This is just like how teaching a chef is different from that chef plating a single dish. Both of the preceding values are attached to a single training run.

The reason there are two numbers for the same training is that the authors of the paper drew the counting boundary twice. If you only count the variable power that equipment drew to calculate while doing the final training—meaning the power that goes up and down according to the work done—it is about 24.7 tons. If you include every process from making the equipment to running it with energy, it is about 50.5 tons. The same paper writes both values side by side, and adds in its conclusion that precisely estimating such emissions is itself difficult (arXiv:2211.02001, preprint).

The name given to this difference is the key term of this article. The accounting boundary is the line set to decide how far to count as part of this activity. This is like deciding whether to include only ingredient costs when calculating a store's expenses, or to also include delivery gas money and the price of the refrigerator. 24.7 and 50.5 are values counting the same fact with different lines.

The numbers attached behind sentences like "AI is eating electricity" also disagree with each other by several times depending on the source. If you read those discrepancies as the fault of ignorance or lying, nothing is left. If you take apart the choices that made the numbers one by one, the places where numbers split are three: how far did they count, what did they count, and what did they divide by.

Before reading, I clarify three things. First, this article does not produce a total correct answer. Authoritative figures on how much electricity a country or the world uses usually come from international organizations, national laboratories, and corporate reports, but such announcements are not in a format that can be verified here, so they are not cited. What this article asks is why the numbers split rather than what the correct answer is. Second, the evidence grades of the cited studies are uneven. Because it is a new field, there are many public versions that have not undergone peer review, and they are marked as such each time. Third, all the figures appearing here are values that those papers observed or assumed at that time, and figures in this field go stale in a few years. Therefore, I spend space on the choices that make the values rather than the values themselves.

Our Guesses Are Off in the Upward Direction

What percentage of US electricity do data centers use right now? When 303 online respondents, recruited to mimic the US demographic composition, were asked this question, the median response was 20 percent. The value the survey used as its grading benchmark was 4.7 percent, reported by Lawrence Berkeley National Laboratory for 2024, and 92 percent of respondents guessed higher than that benchmark. 63 percent placed the current share equal to or greater than 11.8 percent, the same agency's 2030 projection scenario. They saw the present as larger even than the forecast for a few years down the line (DOI: 10.2139/ssrn.7227338, preprint). This is a preregistered study, meaning the questions and analysis plan were registered in advance.

Earlier, we mentioned that institutional report figures would not be cited, but there is a reason 4.7 percent appeared. This value is not the total quantity this article puts forward, but the baseline chosen by the survey to grade responses. That baseline, too, stands upon the selection of accounting boundaries that this article will examine going forward.

The pattern of the missed guesses also had notable features. The group least confident in their own guess was the most accurate, and the four respondents who answered that they were very confident missed the median by 65.8 percentage points. In the authors' phrasing, confidence ran backward. The gender difference was also large. The median for women was 30 percent and for men 11 percent, and the gap remained even after controlling for age, political orientation, race, sources of information, and self-reported confidence. No clear association was caught in this sample between political orientation and guesses about the current share. However, the authors themselves noted that the two patterns regarding gender and confidence were exploratory discoveries that caught their eye while looking at the data outside their preregistered hypotheses, meaning they require replication.

Reading this result as "so it is all an exaggeration" goes further than what the survey said. The survey showed one thing. This sample's sense of the total quantity is way off, and the direction it is off is upward. And all the survey asked about was the proportion data centers take up of the nation's entire electricity. The harder question in this subject lies outside that proportion—namely, when and where demand bunches up—and we will return to that question after looking at all three branches.

The First Axis — Where the Counting Stops

In Dhaka, Bangladesh, there is a colocation data center where multiple companies rent space and bring their own servers. It is a hot, humid place with tight power grid conditions. A power grid is a system that connects power plants and places that use electricity with wires so they work together as one, which the industry calls the grid. Facilities in this kind of location almost never appear in published research records. The researchers collected and analyzed 8,736 remaining records taken every 15 minutes over three months at this facility (DOI: 10.2139/ssrn.7169838, preprint).

To read this facility, a single indicator is needed. Power Usage Effectiveness is the total electricity drawn by the entire facility divided by the electricity that actually went into the computing equipment. It goes by the English abbreviation PUE, and computing equipment refers to hardware like servers that does the actual calculating. If this value is 1.5, it means the building pulled in 1.5 to put 1 into the servers, and the remaining 0.5 was mainly used for cooling and carrying the electricity. The smaller this value, the better the efficiency. It is the basic efficiency metric of a data center, and this value is also what decides whether a given estimate counted only the computing equipment or the entire facility.

At the Dhaka facility, cooling behaved differently than expected. The electricity used by chillers, the large cooling machines, was explained much better by the outside temperature than by the electricity the computing equipment used. The coefficient of determination, which shows the share of each factor in explaining the ups and downs of chiller electricity, was 0.44 for the outside temperature and 0.11 for the computing equipment side. The authors write that this result overturns the conventional wisdom from studies of mega-facilities in temperate climates—namely, the narrative that computing equipment drives cooling electricity. And the optimal operation method found by the authors reduced cooling electricity by 23.3 percent during peak hours when electricity is used the most, but the facility's power usage effectiveness improved by only 0.83 percent (DOI: 10.2139/ssrn.7169838, preprint).

The authors explain the reason why cutting peak-hour cooling by 23 percent barely moved the metric by 1 percent not as a failure of the model, but as a structural ceiling. Cooling accounts for 28.6 percent of the total facility power at that site, and the proportion of cooling that can be reduced by changing operations is at most 5.8 percent. Because the extent to which the metric can improve is the product of these two values, 1.65 percent was the upper limit from the start, and even when scanning the sensitivity range, it sits between 1.2 and 2.1 percent. At this single facility, cooling accounts for only a fraction of facility power and cooling reducible by changing operations is also just a few percent, so even significantly cutting peak-hour cooling barely budged the power usage effectiveness. Therefore, the authors conclude that power usage effectiveness is an incomplete goal in places where cooling accounts for only a fraction of facility power. The task of removing the heat generated by computing itself is followed in "When Air Is No Longer Enough — Where the Heat from Computing Goes".

The boundary problem also exists inside the metric. Power usage effectiveness counts all electricity that went into the computing equipment as a share used usefully. A 201 러시아(2019) paper questioned this very practice. From the data center's perspective, all electricity that went over to the computing equipment is useful, but is that really true? (DOI: 10.25791/pribor.12.2019.1065). This is a question that existed even before AI became a buzzword.

Boundaries also waver at the meter. The Dhaka facility has an onsite generator that does not get picked up by the meter, creating a mismatch between the electricity intake recorded by the utility meter and the electricity the facility actually used. The authors write that this mismatch can silently undermine demand response programs—systems that reward users for cutting usage to match grid conditions—that rely on utility meter data (DOI: 10.2139/ssrn.7169838, preprint). The question "How much was used?" always hides the question "Which meter did you read?"

There are also approaches that nail down the calculation boundary with an equation. One academic paper that built a scenario up to 2030 divides total data center power into "electricity used by computing equipment × infrastructure multiplier". The former is determined by how training and inference are mixed, while the latter is determined by the efficiencies of cooling, power distribution, and power conversion (DOI: 10.63550/iceip.2026.20.71.004). This paper is not a study presenting new observational data, but rather a space synthesizing existing research to propose a conceptual framework. If you multiply the two terms and publish only the single resulting number, nobody can tell which side moved.

The Second Axis — Making It Once Versus Using It Every Day

Training is a one-time event, and inference is something the model repeats every day after it is deployed. Therefore, depending on which side you count, the numerator of the computation changes entirely. A carbon footprint is the total amount of greenhouse gases directly or indirectly emitted by an activity. A study that built a model to predict the carbon footprint of inference notes this as its starting point: across their entire lifetimes, the carbon footprint language models produce in inference is considerably larger than what they produce in training (arXiv:2410.02950, preprint).

The same study adds that estimating the inference side is actually harder than the training side. The batch size processed at once per request, the length of the prompt, and the number of generated tokens (text fragments) all vary. Each cloud provider attaches different types of computing devices in different quantities. Furthermore, the actual utilization rate of computing devices for actual calculation is low and fluctuates widely, and this property makes existing formula-based models inaccurate (arXiv:2410.02950, preprint).

Just changing the way it is run shifts the values significantly. In a study that measured inference energy while varying software, computing device architecture, and service methods across various tasks including conversational AI and code generation, properly applying efficiency techniques reduced total energy by up to 73 percent compared to an unoptimized baseline. The same study left a heavier sentence: simple estimates based only on computation volume or theoretical device utilization significantly underestimate actual energy consumption (arXiv:2504.17674, preprint).

Placing the two studies side by side reveals the character of this axis. Even when given the same task, the value can diverge by more than a factor of two depending on how it is run, and estimates based solely on computation volume or theoretical device utilization fell far short of actual consumption. Existing formula-based models were also inaccurate in inference. As the primer "Watts and Watt-Hours — Reading Electricity's Two Numbers Like a Speedometer and an Odometer" established, watts tell us how hard it is being used right now, and watt-hours tell us how hard it was used for how long to reach a total amount used. A number like "so many watt-hours per query" is not a value but an impression unless the provider, hardware, and batch configuration are disclosed.

Which model it is also determines the value. In a study comparing deployment costs using the energy and carbon required for 1,000 inferences, a generative model that handles various tasks across the board was more expensive by an order of magnitude—meaning ten or a hundred times—than a model tuned to do only one task. This remained true even when controlling for the number of parameters (arXiv:2311.16863, preprint). The denominator of "a single AI run" fixes nothing unless it specifies which AI it is.

The Third Axis — What Are We Dividing By?

There is a paper that compares the resources required when a human works for one minute and when a language model generates answers for one minute in the same units (DOI: 10.21203/rs.3.rs-9962169/v1, preprint). The authors added up the resources it takes for a representative American adult to live across their entire lifecycle, broken down into transportation, aviation, residential energy, food, manufactured goods, and water supply. Converting that energy into a steady intensity used evenly throughout the entire day yields roughly 6,700 watts. This is the average power mentioned in the primer — in other words, the energy used over a period divided by that time — and it combines not only electricity but also the energy contained in fuel and food. Water amounts to about 7,800 liters per day. Dividing this by a 40-hour work week results in 467 watt-hours and 22.7 liters per minute of work.

The AI-side values expressed in the same units are much smaller. Text inference, where a language model generates text, required 1.5 watt-hours and 3.9 to 4.6 milliliters of water per minute of operation based on current hardware, while image generation took about 2.9 watt-hours and 7.5 to 8.9 milliliters per task.

The authors layer a productivity comparison on top of this. If the human-to-AI productivity ratio is conservatively set at 100 to 1, the human's lifecycle energy demand outweighs text inference by roughly 31,000 times and image generation by roughly 16,000 times. For water, the difference ranges from 70,000 times to 1,900,000 times depending on the methodology. The authors write that this result is robust across hardware, grid, and methodological assumptions, and summarize their contribution as a per-unit baseline that was missing when evaluating the net resource effects of adopting AI.

The intuition that "asking AI a question once is committing a major sin against the environment" does not hold up, at least in front of this comparison. This result is worth keeping alive.

At the same time, this value stands upon three conditions. First, the human-side boundary spans the entire lifecycle. The 6,700 watts includes that person's commuting, air travel, meals, and purchased goods. It is a value that charges the cost of existing to one minute of work, whereas the range the abstract discloses for the AI-side 1.5 watt-hours is merely the value during operation on current hardware. Second, the productivity ratio of 100 to 1 is an assumption. The authors call this a conservative value, and write that the gap widens further at ratios suited to observed AI task performance. However, what that ratio actually is does not appear in the abstract, and determining whether the multiplier is 30,000 or 300 rests entirely on this single line. Third, the authors themselves acknowledge a range of uncertainty. The authors attribute the reason why the water multiplier spans a 27-fold difference — from 70,000 to 1,900,000 — to methodological differences. That is a 27-fold spread within the exact same paper.

Therefore, this paper serves as evidence for the side arguing that "AI's resources per task are small," while simultaneously acting as an example that demonstrates the main point of this piece. The better-crafted an estimate is, the more densely conditions are attached on top of it.

When the Two Numbers Do Not Fight

In the comparison between the human's 467 watt-hours and the AI's 1.5 watt-hours, the proposition emerges that the value per task is low. Meanwhile, there is also the proposition that total and localized power demand are the problem. The two propositions do not contradict each other. That is because their levels are different. The former is the value of a single task divided into one, while the latter is what happens when hundreds of millions of those tasks crowd into a single substation at a specific time and region. A substation is a facility that gathers multiple transformers, serving as a place where two paths of different voltages meet. Even if the value per task is low, concentrated demand can cause difficulties, and in practice, the two propositions are largely true together.

The same levels exist for water. When viewed at the national or global scale, a data center's water consumption may appear small compared to other users, but in regions already experiencing water stress or drought, the localized impact can be substantial (DOI: 10.22541/au.176046418.80101496/v1, preprint — a commentary addressing transparency in data center water footprints). Both statements are true in the same resource and the same year, and the only difference is the scale. The water story is continued in "AI Drinks Water — But Saving It Spends Electricity".

The Same Electricity, Two Ways of Calculating It — Bitcoin Mining

That estimates diverge not because of sloppy measurement, but because they are counting different quantities, is most clearly seen in a case outside of AI. Estimates of the electricity used by Bitcoin mining have long been notorious for disagreeing by orders of magnitude across different sources. One preprint dissects that discrepancy head-on, with the authors writing in the first sentence that while these estimates are usually read as uncertainties about a single quantity, they are not actually so. It is the same statement as this article's title.

The estimation methods branch into two. Engineering literature counts electricity by assuming how equipment efficiency is distributed across the observed volume of mining computation (hashpower). Expenditure literature counts electricity backward from mining revenue, fixing electricity prices and cost structures. The two values match only when revenue equals total economic cost; otherwise, they diverge precisely by the ratio of revenue to cost. When the authors dissected the disclosures of a single mining company, the same power expenditure was reproducible with any of five reported cost denominators, but those ratios were scattered from 0.61 to 1.72. The concluding sentence runs like this: The reported spread is not a range of power, but a range of cost concepts (DOI: 10.2139/ssrn.7179538, preprint).

Gathering more data does not bridge this gap. According to the authors, the computation volume and revenue observable from blockchain records merely place one condition on three unknown values, and even as the records grow longer, that shortfall does not shrink.

The way the same paper handles its own numbers is also worth watching. When the coin price is 1 million dollars, the electricity for 2027 calculated backward from revenue is 2,247 terawatt-hours, which amounts to roughly 6.6 percent of the world's projected power generation. A terawatt-hour is a unit for measuring the amount of electricity, where 1 terawatt-hour is 1 billion kilowatt-hours. A kilowatt-hour is the exact unit printed on a household electricity bill, and a terawatt-hours is the scale used when talking about the annual electricity of a whole country. So 2,247 terawatt-hours is a big value. Yet the authors add immediately: this value is not a prediction, but the requirement that coin price demands, and the side that carries information is that requirement. It is an example showing what a sentence looks like when a condition is pinned right next to the number.

Efficiency Slowed the Growth, and Where That Foundation Shakes

The growth in data center energy use has slowed thanks to efficiency improvements. That is the gist of a recalculation published in Science in 2020, which adds that smart policies can help sustain those efficiency improvements in the near term (DOI: 10.1126/science.aba3758). Although it is peer-reviewed print, the only record this article can contrast is that single sentence without methods, samples, or figures. So what we take from it is precisely that much. It means the growth rate has slowed, not that the total amount has stopped.

In 2022, a stronger claim appeared. It is a prediction that the carbon footprint of machine learning training will peak and then decline. The rationale is four best practices, which the authors write can reduce training energy by up to 100 times and carbon dioxide emissions by up to 1,000 times. Following those practices supposedly kept the energy used by machine learning across research, development, and live services below 15 percent of Google's total energy use for three years, which is a self-reported account leaning on the company's own tally. The authors' conclusion is that if the entire field adopts these practices, the total carbon emissions from training will decrease (arXiv:2204.05149, preprint).

The same paper contains one more sentence that summarizes this entire article. For papers that did not state emissions, later estimates by others were off by anywhere from 100 times to 100,000 times. This article reads this as a signal that even without ignorance or dishonesty, working backward from the outside to estimate can diverge by five orders of magnitude. That is why the authors recommend stating emissions directly in papers.

Why such divergence happens is shown item by item in earlier research by the same authors. Depending on where you run the training, the share of carbon-free electricity varies by 5 to 10 times, even within the same country and organization. Cloud data centers are 1.4 to 2 times more energy efficient than typical data centers, and accelerators—the machine learning-specific computing chips inside them—are 2 to 5 times more effective than general-purpose systems. Therefore, the authors write that simply choosing the neural network, data center, and processor can reduce the carbon footprint by up to 100 to 1,000 times (arXiv:2104.10350, preprint).

These 100 to 1,000 times and the 100 to 100,000 times from the previous paragraph are not the same quantity. The former is the margin by which outside backward estimates missed the mark, and the latter is the margin by which choices can reduce emissions. Regarding these large multipliers, what the authors noted in the earlier paper is that they make backward estimation difficult.

The reason the slowdown mentioned by the 2020 recalculation is now shaking is read this way by the conference paper we saw earlier. In this article, that single piece of work is the only evidence supporting that position. In the short term, improvements in market-average efficiency stall and the workload composition shifts toward AI, making power consumption sensitive to service volume. In the long term, algorithmic and engineering improvements lower that sensitivity. The same paper lays out three scenarios side by side for 2030: modest, baseline, and accelerated (DOI: 10.63550/iceip.2026.20.71.004). Putting forward three instead of just one is itself a form of honesty.

Why Efficiency Can't Close the Answer

If you buy a car with better fuel economy, you end up driving it more often and farther. The rebound effect is the phenomenon where the share saved through improved efficiency is filled right back in because people end up using that much more. In a study tracking the aftermath of efficiency improvement shocks in France, Germany, Italy, the UK, and the US, the economy-wide rebound effect two years later was 78 percent to 101 percent (DOI: 10.2139/ssrn.3854876, preprint). This means most, and in some cases all, of what was saved was filled right back in.

A peer-reviewed paper modeling the same relationship using global data from 1900 to 2018 places two futures side by side. On the growth-maximizing path, even if efficiency investments are increased by about 2 times, the growth in the use of primary energy—that is, energy obtained directly from nature before being converted into electricity or fuel, such as oil, coal, natural gas, and hydropower—does not decrease. This is because the economy-wide rebound effect dominates. On the path that minimizes use growth, increasing efficiency investments by about 3.5 times nearly decouples energy and growth, but it leads to a regime where output does not increase in the long run (DOI: 10.1007/s12053-024-10236-7). The authors note that only the latter is compatible with climate targets. Their warning is that as long as the growth-maximizing path continues, reduction promises relying on efficiency risk failing on a global scale.

A review spanning cases of applying AI to transportation, energy, and manufacturing points in the same direction. As a result of closely reading 41 out of 150 papers from the past decade, AI adoption reduced energy use per unit, but frequently increased overall consumption instead (DOI: 10.3389/fenrg.2025.1460586). This is about the side applying AI rather than the data center, but it provides circumstantial evidence that the rebound spares no sector.

Aside from efficiency, the source of electricity also determines emissions. Carbon intensity is the amount of carbon dioxide emitted alongside the process of generating 1 kilowatt-hour of electricity. Even for the same 1 kilowatt-hour, the carbon dioxide that comes attached differs depending on whether it came from coal or hydropower. When claiming green AI, the levers commonly lumped together are four: computation efficiency, facility efficiency, the carbon intensity of electricity, and system-level additionality. Additionality is whether the purchase of that electricity actually brings about new generation that would not have come to be without it. The conclusion of a paper dividing the four this way is straightforward. If AI demand grows faster than the rate at which energy intensity drops, efficiency improvements alone cannot secure sustainability. The same paper adds that while low-carbon electricity can significantly reduce operational emissions, procurement that fails to match it on an hourly basis or generate new power can shift the burden elsewhere on the grid (DOI: 10.2139/ssrn.7336158, preprint — this is a scenario analysis run with a representative 20-megawatt facility).

The Real Issue — Where and How Fast

Even if the total amount used in a day is the same, if the hours of use bunch up into a single time, the result the power grid receives changes. There is a simulation that ran this on the IEEE 30-bus test system, a small trial power grid model that researchers use to compare methods (where a bus is a point where power lines meet). When they ran a high-growth case with a large data center under an emergency condition where both generator and transmission line outputs were cut together, the unsupplied energy—meaning energy that could not be delivered—jumped from 3.203 megawatt-hours in the baseline case that matched energy to 22.891 megawatt-hours. In a supplementary experiment where they kept the total energy used by the data center the same and only overlapped the hours of use into a single time, the total unsupplied energy increased by 34.4 percent further. Under which emergency condition and as an increase compared to what is not detailed in the abstract. The authors' conclusion is this: temporally concentrated data center demand can amplify resilience impacts without increasing total energy consumption (arXiv:2607.06958, preprint).

There is something that must be read alongside these numbers. This is a simulation on a test system, not a blackout amount measured in an actual grid, and because standard test systems are small artificial models made to compare methods, their absolute values should not be read as real-world predictions. What should be read is the direction. Even with the total amount fixed, the results worsened solely from the time distribution.

The term that captures this difference is load. Load is the electric power that the power grid must supply at that very moment. This is the definition established in the primer "Alternating Current and Frequency — A Giant Merry-Go-Round the Whole Country Pushes Together", and in terms of a dashboard, it is a number closer to a speedometer than an odometer. Even if the total amount used in a day is the same, using it little by little throughout the night versus bunching it up at seven in the evening are completely different matters from the power grid's perspective. The preceding sections dealt with total energy, and this section deals with load.

An overview paper reviewing the interaction between AI data centers and the power grid divides this issue into three time axes. They are long-term planning and grid interconnection, short-term operation and the electricity market, and moment-to-moment dynamics and stability (arXiv:2509.07218, preprint). As this article sees it, the question of "how many terawatt-hours the total amount is" does not directly answer any of these axes. Interconnection is followed by "Plugging It In Is Harder Than Building It — What an Interconnection Queue Actually Decides", and momentary stability is followed by "The Few Seconds a Generator's Spinning Mass Gave Away for Free — What Gets Harder as Grid Inertia Falls".

The oldest pressure lies on the neighborhood side. Data centers draw electricity heavily and crowd into a small number of regions. The proposition that these two traits overlap to pressure the available capacity of the operator managing the distribution network—the thin path that finely divides electricity down to neighborhoods and buildings—and cause data centers to compete with other demands such as residential projects or electric vehicle charging was raised in 2023 (DOI: 10.1093/ooenergy/oiad014 — this is a forum raising questions rather than a study presenting data).

Who Bears the Concentrated Results?

An estimate suggests that in the five U.S. states where data centers are most concentrated, the gap between residential and commercial electricity rates widened by 1.27 cents more per kilowatt-hour than in paired comparison states since January 2022. This level exceeds standard statistical significance thresholds (p = 0.025), which works out to 133 dollars per year per household based on average U.S. household consumption. This is the result of applying difference-in-differences to 1,188 state-month observations from 2015 through 2025. Difference-in-differences is a method that measures the changes before and after a certain point in data center-concentrated states and comparison states respectively, and looks at the difference between those changes. The event study pinpoints the causal break not at 2022 but at 2023, which the authors interpret as excluding the Russia-Ukraine shock and reflecting the 12 to 18-month regulatory lag from a data center's interconnection agreement to approved rate adjustments (DOI: 10.2139/ssrn.7158958, preprint).

This result is a single study, a publicly available preprint that has not undergone peer review, and claims causality from observational data. Furthermore, the authors themselves frame this as an issue of rate design structures. It is cross-subsidization, where large operators negotiate discounted commercial rates and recover the shortfall from households, meaning it can be handled through rate-making reforms regardless of the pace of AI deployment. Because the link the authors point to between data center growth and the rate gap is the rate structure rather than AI itself, this article views the reading that "bills go up because AI is growing" as distant from this study's conclusion. What this study measured is also not the level of rates, but the gap between residential and commercial rates.

The Same Prescription, Different Values

A data center's flexibility is its ability to shift the time or place it uses electricity, or to reduce use when needed. One study put that value into two grids using capacity expansion models—models that choose what future facilities should be built and when so that costs are lowest (arXiv:2608.19622, preprint). In the US PJM grid, which is market-driven and fossil-heavy, the value of flexibility came out spatially. Shifting work between regions reduces system costs by 6 percent in 2028 and 19 percent in 2038, and avoids having to build 4.4 gigawatts and 8.9 gigawatts of gas and nuclear generation. On the other hand, in South Korea, which has a carbon cap and is centrally coordinated, the value came out temporally. Shifting load to midday solar hours gives value to building 0.5 gigawatts more solar in 2028, and avoids 1.2 gigawatts of gas and 0.3 gigawatts of battery in 2038. These are values produced by a model, not observed events. The authors also note that in both grids, realistic event shape constraints—the practical limits on the length and shape of a load-reduction event—reduce the value on the load-reduction side. The same prescription works along different axes in the two countries. The details of shifting are continued in "Is a Data Center Really an Inflexible Load? — What Can Move and What Cannot".

There is also a review that lowers expectations. One review summarizing flexibility research notes that while various techniques are proposed, the scale and cost of their potential are relatively little known. And using actual energy use data, it shows that data centers are unlikely to be sensitive to electricity prices except during extreme events (DOI: 10.1007/s40518-025-00258-9). What is technically possible and what actually happens are different.

There are also boundaries on the carbon side. A working paper dealing with the greenhouse gases of electricity bought from the power grid examines three widely proposed conditions: that the electricity must come from nearby, at the same time, and from newly built power plants. The conclusion is that electricity satisfying those conditions is cleaner than average, but not carbon-free. The additional emissions generated by that purchase—the incremental emissions per megawatt-hour—are 30 to 43 percent lower than the unconstrained average. The authors estimate that among existing or planned generation, about 10 percent of US electricity in 2030 will satisfy all three conditions. However, the same authors also note that if clean power demand exceeds limited supply, the three conditions can produce additional benefits of reducing total electricity demand or driving the construction of new clean generation (DOI: 10.2139/ssrn.5172090, working paper).

The Range Is This Wide Even Inside a Single Paper

A single paper analyzing China's power generation and data center consumption together compresses the core argument of this article. In 2018, China's data center power consumption was 161 terawatt-hours, accounting for 2.35 percent of the total. The 2030 projection is not a single value but a range. It is 277 to 500 terawatt-hours, depending on the speed of AI adoption and the degree of efficiency improvement.

The reduction potential side has an even wider spread. The same abstract states the value in one sentence as 30,000,000 to 50,000,000 tons per year by 2030, and in the next sentence as "reported 20,800,000 to 45,000,000 tons." It adds that even that range is uncertain due to time-varying emission factors and core assumptions, and under a load-shifting scenario, it widens to 14,000,000 to 52,000,000 tons depending on how well renewable generation and timing align (DOI: 10.3389/fenrg.2026.1774740). Even within a single paper, the values do not converge to one.

With this much spread within one country, one year, and one paper, discrepancies between data sources are less a flaw in the data than a property of the subject itself.

The Next Time You Meet a Number

The next time you meet a sentence saying "AI uses this much electricity," you should ask three things.

First, where is the boundary? Ask whether it includes just the computing gear, the whole facility, the factory that built the gear, or even the water that evaporated to make the electricity. The distance between 24.7 and 50.5 comes from here.

Second, what was counted, and when? Ask whether it is training or inference, which model was run and how, and which year's hardware, which year's grid, and which year's model it is. In a field where a few choices can move a figure by up to 100 times to 1,000 times, a number without a year is hard to use as a value.

Third, what was it divided by? Ask whether it is per single query, per minute of labor, for an entire country, or for a single substation. If the denominators are different, the two numbers are not answering the same question. And one more thing, look to see if conditions are attached. The reason 2,247 terawatt-hours was an honest figure was not its size, but because "when the coin price is 1 million dollars" and "a requirement, not a prediction" were written right beside it.

Right now, the most certain conclusion in this subject is not any total-amount figure, but the fact that reporting is lacking. Different camps rarely agree on this conclusion. The side that expects training emissions to peak and fall recommended stating emissions clearly in papers to create a way to compete beyond just model quality (arXiv:2204.05149, preprint). The side that built 2030 scenarios placed the standardization of measurement and mandatory reporting at the very top of their practical implications (DOI: 10.63550/iceip.2026.20.71.004). A commentary covering water footprints wrote that without comprehensive monitoring and public disclosure, neither regulation, innovation, nor regional planning will work properly (DOI: 10.22541/au.176046418.80101496/v1, preprint). The side that dissected Bitcoin estimates wrote that no disclosure states how long mining equipment can actually be used profitably—in other words, its effective economic life—and that this would be the most valuable disclosure of all (DOI: 10.2139/ssrn.7179538, preprint).

Finally, let us shift the axis of our questions. Questions about total amounts have multiple answers. Local questions remain not because there is only a single answer, but because the entity to answer and the deadline are fixed. Will that region's distribution grid make it through this summer? Can that substation take on new loads? Who pays the increased costs? These questions remain exactly the same whether data centers account for 4.7 percent or 20 percent of U.S. electricity.

So, in the sentence "AI is devouring electricity," the part that holds real information is not the verb, but the three blank spaces: how far the counting went, what was counted and when, and what it was divided by. Once those blanks are filled, the debate usually vanishes, and two different questions remain. And those two questions are usually both true.

Source Papers

Papers published on arXiv, SSRN, and Research Square are marked as "preprint" and whether they were subsequently published in academic journals was not separately verified. Working papers from institutional research series (NBER) are marked as such. Simulations and single-facility results are indicated in the text each time.

• "Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model" (2022, preprint) — arXiv:2211.02001 — The same final training yields 24.7 tons and 50.5 tons depending on the boundary. The conclusion explicitly states that precise estimation is difficult.
• "Grid Expectations: Americans' Overestimation of US Data Center Electricity Use" (2026, preprint) — DOI: 10.2139/ssrn.7227338 — An online sample of 303 people mimicking the US demographic composition. Median of 20 percent versus a baseline of 4.7 percent. Gender and confidence patterns are exploratory.
• "Data center cooling in a hot and humid climate" (2026, preprint) — DOI: 10.2139/ssrn.7169838 — 8,736 Dhaka facility instances. 23.3 percent cooling reduction with a PUE of 0.83 percent (upper bound 1.65 percent). Results described by the authors as overturning the load-centric narrative.
• "Power Usage Effectiveness of the data center as a combination of engineering and IT infrastructure" (2019) — DOI: 10.25791/pribor.12.2019.1065 — Raises questions about the practice of counting all power going to IT equipment as useful share.
• "AI Energy Elasticity and Data Center Energy Consumption Growth Scenarios to 2030" (2026, conference paper) — DOI: 10.63550/iceip.2026.20.71.004 — Total power = IT component × infrastructure multiplier. High short-term elasticity and long-term decline. A conceptual framework synthesizing other studies.
• "Energy Considerations of Large Language Model Inference and Efficiency Optimizations" (2025, preprint) — arXiv:2504.17674 — FLOP-based estimation significantly underestimates actual consumption. Total energy reduced by up to 73 percent with optimization.
• "LLMCO2: Advancing Accurate Carbon Footprint Prediction for LLM Inferences" (2024, preprint) — arXiv:2410.02950 — Inference emissions exceed training over the lifetime (assuming deployment). Formula-based models are inaccurate.
• "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" (2023, preprint) — arXiv:2311.16863 — Based on 1,000 inferences. Multi-purpose generative models are orders of magnitude more expensive than task-specific ones.
• "To AI or not to AI? Per-task energy and water demands of AI inference are orders of magnitude lower than human labor" (2026, preprint) — DOI: 10.21203/rs.3.rs-9962169/v1 — Representative adult roughly 6,700 watts and 467 watt-hours per minute of work. 1.5 watt-hours per minute of text inference operation under current hardware. Roughly 31,000 times at a 100-to-1 assumption.
• "Data Centers Water Footprint: The Need for More Transparency" (2025, preprint) — DOI: 10.22541/au.176046418.80101496/v1 — Even if small at the national or global scale, local impacts in water-stressed regions can be substantial.
• "Two Estimands, One Identity" (2026, preprint) — DOI: 10.2139/ssrn.7179538 — The order-of-magnitude difference is not uncertainty about a single quantity. Ratio of 0.61 to 1.72 in disclosures. 2,247 terawatt-hours is a requirement, not a forecast.
• "Recalibrating global data center energy-use estimates" (2020) — DOI: 10.1126/science.aba3758 — Efficiency gains have slowed growth, and smart policies can maintain this in the near term (the entire abstract is this single sentence).
• "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink" (2022, preprint) — arXiv:2204.05149 — Best practices reduce training energy by up to 100 times. ML energy less than 15 percent of Google's total over 3 years (company's own figures). Retrospective estimates were off by 100 to 100,000 times.
• "Carbon Emissions and Large Neural Network Training" (2021, preprint) — arXiv:2104.10350 — Region 5 to 10 times, cloud 1.4 to 2 times, accelerator 2 to 5 times. Up to 100 to 1,000 times reduction simply by choosing the neural network, data center, and processor.
• "Do Energy Efficiency Improvements Reduce Energy Use?" (2021, preprint) — DOI: 10.2139/ssrn.3854876 — S-FAVAR estimation for 5 countries. Economy-wide rebound effect of 78 to 101 percent after 2 years.
• "Economy-wide rebound and the returns on investment in energy efficiency" (2024) — DOI: 10.1007/s12053-024-10236-7 — The maximum growth path fails to reduce energy growth even with a 2x efficiency investment, while the minimization path requires a long-term zero-growth regime.
• "AI beyond efficiency, navigating the rebound effect in AI-driven sustainable development" (2025) — DOI: 10.3389/fenrg.2025.1460586 — Read closely 41 out of 150 papers applying AI to transport, energy, and manufacturing. Per-unit drops, but total consumption rises.
• "Governing the Energy Footprint of Artificial Intelligence for Sustainable Development" (2026, preprint) — DOI: 10.2139/ssrn.7336158 — Four levers (compute efficiency, facility efficiency, carbon intensity, additionality). Scenario analysis for a 20-megawatt facility.
• "Evaluating Grid Resilience in the Era of Ever-Increasing Data Centers" (2026, preprint) — arXiv:2607.06958 — IEEE 30-bus test system simulation. Base case 3.203 to 22.891 megawatt-hours, with an additional 34.4 percent increase under simultaneous loading.
• "Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects" (2025, preprint) — arXiv:2509.07218 — Three time axes: long-term planning and interconnection, short-term operation and markets, and real-time dynamics and stability.
• "Stretched grid? Managing data center energy demand and grid capacity" (2023) — DOI: 10.1093/ooenergy/oiad014 — A page raising the problem that heavy loads and local concentration strain distribution capacity (ends with a question).
• "Who Pays for AI? Causal Evidence That Data Center Growth Widens the Residential-Commercial Electricity Price Gap" (2026, preprint) — DOI: 10.2139/ssrn.7158958 — 1,188 state-month differences-in-differences. Gap of 1.27 cents per kilowatt-hour and 133 dollars per household annually. Authors characterize this as cross-subsidization addressable through rate reform regardless of AI deployment speed.
• "Shift or curtail? How much data-center flexibility is worth depends on the host power grid" (2026, preprint) — arXiv:2608.19622 — Capacity expansion model. Spatial type for PJM (6 and 19 percent), temporal type for South Korea. Realistic constraints lower curtailment value.
• "The Potential of Data Center Energy Demand To Provide Grid Flexibility" (2025) — DOI: 10.1007/s40518-025-00258-9 — The scale and cost of flexibility are largely unknown, and it is unlikely to be sensitive to electricity prices except during extreme events.
• "Greenhouse Gases Resulting from Grid-Connected Electricity Demand: Three Pillars and Scope Two" (2025, working paper) — DOI: 10.2139/ssrn.5172090 — Not carbon-free even if three conditions are met, with incremental emissions 30 to 43 percent lower than average. Also notes potential additional benefits.
• "Analysis of China's power development and the impact of data center energy consumption" (2026) — DOI: 10.3389/fenrg.2026.1774740 — 161 terawatt-hours in 2018, 277 to 500 terawatt-hours in 2030. The same abstract records reduction potential as both 30,000,000 to 50,000,000 tons and "reported 20,800,000 to 45,000,000 tons".