How Big Is AI's Climate Impact? This Is How Much Energy and CO₂ an AI Query Requires

In this article, we cover:

• How much energy and CO₂ a single AI query actually requires — and why the numbers vary so much between studies.

• The difference between a simple text query and heavier AI tasks like reasoning, image generation, and video.

• Why AI is quickly becoming more energy efficient per query — yet still uses more electricity overall.

• How much of the world's electricity use and CO₂ emissions data centers and AI actually account for today, and how that's expected to develop by 2030.

• What you as an individual can keep in mind — and where the biggest responsibility really lies.

The short answer is that a typical text query to a modern AI model today often appears to require around 0.2–0.4 Wh of electricity (roughly as much energy as watching TV for 7–15 seconds), with climate emissions of a few hundredths of a gram of CO₂e per query. The variation is large, though: a long AI conversation, an advanced reasoning model, image generation, or video generation can require significantly more energy. And above all — when billions of people and businesses use AI, the combined energy use becomes substantial.

There is therefore reason both to add nuance to, and to take seriously, AI's climate impact.

How much energy does an AI query require?

Until recently, almost all estimates of the energy consumption of services like ChatGPT and Gemini were based on indirect calculations, since AI companies didn't publish enough information about how their systems actually ran.

That changed somewhat in 2025, when Google published a detailed study of actual energy use in its production environment for Gemini. The study found that a median text prompt in Gemini used 0.24 Wh of electricity, caused an estimated 0.03 grams of CO₂e, and consumed 0.26 milliliters of water. [1] It's particularly interesting because it counts not just the electricity used by the AI chip itself, but also the CPU, memory, standby capacity, and the data center's other energy use, such as cooling.

Another 2025 study estimated the median consumption of a query to a so-called frontier model — a very large, advanced language model — at around 0.31 Wh, with a range of 0.16–0.60 Wh. [2]

Several independent estimates point in roughly the same direction: a fairly typical text query to today's large AI models often falls somewhere around a few tenths of a watt-hour. But it's important not to turn this into a universal emissions factor, for reasons that become clear below.

How much CO₂ does that add up to?

Energy use and climate impact are two different things. If an AI query requires 0.24 Wh of electricity, the resulting greenhouse gas emissions depend on what electricity the data center uses. Electricity produced with coal power can have a very high climate impact, while wind, hydro, and nuclear power have significantly lower lifecycle emissions. On top of that, big tech companies use different kinds of power purchase agreements, renewable electricity production, and climate measures, which makes it even harder to set a general emissions factor for AI.

Google's calculation for Gemini came out at 0.03 grams of CO₂e per median prompt. [1] Using those figures specifically, that works out to roughly:

• 1 AI query: 0.03 g CO₂e

• 100 AI queries: 3 g CO₂e

• 1,000 AI queries: 30 g CO₂e

• 1 million AI queries: 30 kg CO₂e

A million such queries would at the same time use about 240 kWh of electricity.

But these numbers shouldn't be read as emissions factors for all AI. A query to ChatGPT can't automatically be assumed to have the same climate impact as a query to Gemini — different models, hardware, data centers, and tasks can produce completely different results.

To put the numbers in perspective:

• An AI query (0.24 Wh) requires roughly as much energy as watching TV for less than nine seconds — a comparison Google itself makes in its study. [3]

• The water use for a query (0.26 ml) is roughly equivalent to five drops of water. [3]

• The CO₂ emissions (0.03 g CO₂e) correspond to roughly driving an average new car in Europe just under 30 centimeters (based on the EU's average emissions of 107 g CO₂/km for new cars in 2023). [4]

• The energy for an AI query is roughly equivalent to about 1 percent of a full phone charge — it would take around 75 AI queries to use as much energy as a single phone charge. [5]

• According to the water footprint study on GPT-3, the model "drinks" a 500 ml bottle of water for every 10–50 medium-length responses it gives. [6]

Zooming out from the individual query to the industry as a whole, the picture gets bigger. Goldman Sachs estimates that AI and data centers globally could account for around 3–4 percent of the world's electricity use by 2030, up from today's 1–2 percent. [7] In the US specifically, a study in Nature Sustainability, co-authored by KTH researcher Francesco Fuso Nerini, points to AI servers alone potentially consuming up to ten percent of the country's current electricity use by 2030 — roughly as much as Germany's total annual electricity consumption. The same study estimates that, if current growth rates continue, AI expansion in the US could cause 24–44 million additional tonnes of CO₂ emissions per year by 2030 (equivalent to the emissions from 5–10 million additional cars on US roads) and draw 731–1,125 million cubic meters of water annually (on par with the annual household water use of 6–10 million Americans). The researchers also note that best practices — such as siting servers where the grid is already fossil-free — could cut emissions by up to 73 percent and water use by up to 86 percent. [8]

Not all AI queries are equal

Perhaps the biggest problem with the question "how much CO₂ does an AI query emit?" is that an AI query isn't a standardized activity. Compare, for example, the question "What is the capital of Sweden?" with the request "Read this 300-page document, analyze the material, run ten alternative calculations, and write a 5,000-word report." Both can technically be described as a prompt, but the computational demand is obviously completely different.

The number of tokens the AI needs to read and generate affects energy use. So does the size of the model, the hardware it runs on, how many users can be processed at once, and how efficient the software is. Researchers who have reviewed the research on AI inference have found that estimates of climate impact can vary widely depending on the methodological choices and boundaries used. [9]

In a study of so-called test-time compute, where models are allowed to use significantly more computation to reason through an answer, the estimated energy use rose from about 0.31 Wh to 3.91 Wh per query — more than ten times as much. [2]

The International Energy Agency (IEA), in its 2026 report Key Questions on Energy and AI, notes that energy use per simple AI task has dropped dramatically in recent years, even as AI is increasingly used for much heavier workloads. The IEA specifically highlights reasoning AI, AI agents, and video generation — tasks that, according to the agency, can use hundreds or thousands of times more energy than simple text generation. [10]

Saying that "an AI prompt uses X Wh" is therefore about as misleading as saying "a car trip uses X liters of gasoline" without specifying whether the trip is one kilometer or a thousand.

AI is quickly becoming much more energy efficient

There's an important development here that often gets lost in the debate: AI models are becoming much more efficient. Google reports that the energy use of their median Gemini prompt dropped 33-fold in a year, while the climate footprint dropped 44-fold. [1] The IEA makes a similar observation, describing the efficiency gains per AI task as exceptionally fast. [10]

The explanation is a combination of better AI models, algorithms, chips, data centers, software, and batching of user queries — that is, processing several requests together to use the hardware more efficiently.

Research has also long shown how much of a difference the choice of model, hardware, and location can make. A study by Google researchers estimated that the combination of model architecture, processor, and data center can in some cases change the climate impact of machine learning by several orders of magnitude. [11] A later review of the research on AI inference confirms the same pattern: workload, hardware choice, and software optimizations can dramatically change energy use, even for the same model. [12]

This is an important reason why older calculations of AI's energy consumption can quickly become misleading.

But overall, AI is using more and more electricity

This creates a paradox: each AI query becomes cheaper in energy terms, while AI as a whole uses more and more electricity. The reason is scale. More people use AI, companies build AI into their products, models are used more continuously, and new use cases like video, reasoning, and AI agents require far greater computing resources.

The IEA estimates that the world's data centers used about 485 TWh of electricity in 2025, and that consumption is expected to roughly double by 2030 to around 950 TWh per year — equivalent to about 3 percent of the world's expected electricity consumption. [10]

Data centers used primarily for AI are growing even faster: according to the IEA, their electricity use rose by about 50 percent in 2025 alone, and is expected to roughly triple between 2025 and 2030. [10] That doesn't mean all of this electricity is used by ChatGPT queries — data centers power everything from cloud services and databases to streaming, search engines, and traditional IT. But AI is a major part of the rapid growth.

What's happening with data center emissions?

How much climate impact this increased electricity use has depends largely on how the electricity is produced. In 2025, the IEA estimated that the world's data centers accounted for about 1.5 percent of global electricity use in 2024, and that the share could approach 3 percent by around 2030. [13]

Renewable electricity is expected to account for a large share of the new power generation, but not all of it. According to the IEA, natural gas and coal together are expected to account for more than 40 percent of the increased electricity generation for data centers through 2030. The IEA therefore calculates that emissions from the electricity supplying data centers could reach around 320 million tonnes of CO₂ per year by around 2030, before declining somewhat thereafter in the agency's main scenario. [14]

Globally, that still amounts to less than one percent of total climate-warming CO₂ emissions. That makes AI and data centers an important climate issue — but not the dominant one.

Training AI models also requires energy

When a new AI model is created, it first has to be trained on huge amounts of data, which can require enormous computation. A well-known example is the 176-billion-parameter language model BLOOM. A lifecycle analysis estimated the emissions from the training itself at about 25 tonnes of CO₂e, or around 51 tonnes of CO₂e when including things like equipment manufacturing and other energy use. [15]

Other large models may have significantly bigger climate footprints, but the problem is that the information is often missing. For many commercial models, we don't know how many GPUs or AI accelerators were used, how long training took, how much electricity was consumed, where the data centers were located, what the electricity mix was, or how large the climate impact of hardware manufacturing was.

Researchers have repeatedly called for better reporting from AI companies. [11][16] It's therefore considerably easier to find an exact number than a correct one when trying to calculate the climate impact of an AI model.

AI's climate impact is bigger than just electricity

A complete environmental analysis shouldn't look at electricity alone. Data centers and the manufacturing of AI hardware require, among other things, water for cooling and power generation, metals and minerals, semiconductors, construction materials, and servers and networking equipment.

A 2023 study drew attention to AI systems' water use and estimated that training GPT-3 in US data centers could be linked to about 700,000 liters of direct water consumption. [6]

There's also a growing materials issue. A study published in Nature Computational Science estimated that, depending on how fast the technology develops, generative AI could contribute to a total of 1.2–5 million tonnes of electronic waste between 2020 and 2030. [17]

These estimates are uncertain too, but they illustrate why AI's climate impact can't be reduced to the energy use of a single prompt.

So how bad for the climate is it to use ChatGPT?

For an individual, the answer is likely: not much, per ordinary text query. If today's estimates of a few tenths of a watt-hour per simple text query are roughly right, you'd need to ask an awful lot of questions before your own AI use becomes a significant part of your personal climate footprint.

That doesn't mean AI has no climate impact. The real climate issue instead arises when a small impact is multiplied by billions of uses, while AI is built into more and more products and increasingly used for far more computationally intensive tasks. It's the same phenomenon we've seen in many other areas: more efficient technology doesn't automatically lead to lower total energy use if usage grows even faster at the same time.

How can AI emissions be reduced?

For an everyday user, there's no reason to feel guilty about every question you ask an AI. But a few reasonable principles can serve as guidance.

• Use AI where it creates value. Generating large amounts of content that's never used is obviously harder to justify than using AI to make actual work more efficient.
• Don't use more computing power than you need. A simple question doesn't always need the world's largest reasoning model.
• Be especially mindful with images and video. Generating large images, and especially video, can require significantly more computation than text.

That said, the biggest opportunities lie with the AI companies themselves. They can reduce climate impact through more efficient models and hardware, by locating computation where electricity has a low climate impact, by extending the lifespan of equipment, and — not least — by publishing much better data on their actual energy and resource use.

A reasonable answer in 2026

If someone asks "how much CO₂ does an AI query emit?", today's best answer is roughly this: for an ordinary text query to a modern AI service, energy use often runs around a few tenths of a watt-hour. Google's measured median for Gemini is 0.24 Wh and 0.03 grams of CO₂e per prompt. But the variation between models and tasks is enormous, and more advanced AI — especially reasoning, agents, images, and video — can require many times more energy. That's why there's no single emissions factor that accurately describes all AI use.

In other words, AI's big climate question is probably not the individual prompt. It's the scale. When hundreds of millions of people use the technology every day, AI is integrated into nearly every digital service, and data centers are being built at a rapid pace, the combined demand for electricity, hardware, and resources becomes substantial. At the same time, efficiency gains are happening extremely fast.

That's why we need to be cautious both of alarmist claims that every AI query has an enormous climate impact, and of the argument that AI's climate impact is too small to matter. Both oversimplify a question that is still evolving very quickly.