AI resource consumption

Energy per query

The figure depends directly on token length. In 2025, Epoch AI published the cleanest token-based model:

Source: Epoch AI, 2025. Calculation model: GPT-4o, ~100 billion active parameters, NVIDIA H100.

The 10x myth: origin and context

The widespread claim that an AI query uses 10 times as much energy as a Google search goes back to a single estimate by Alphabet chairman John Hennessy (Reuters interview, February 2023). No measurement, no study, a casual “likely”.

From there the chain ran on: de Vries quoted it in a Joule paper, the IEA adopted it, the UN Environment Programme too. Since then it has circulated as an established figure.

The second problem: the reference value of “0.3 Wh per Google search” comes from a Google blog post from 2009. Current estimates put a search at ~0.04 Wh. Both sides of the calculation were overestimated by a factor of 10.

Source: Engineering Prompts, 2024

The comparison is losing its second half

The comparison assumes an AI query on one side and a classic keyword search on the other. Google is currently dissolving that distinction itself. AI Overviews have appeared automatically above the result list since 2024, without users selecting them. According to Google, AI Mode has more than one billion monthly active users, with queries more than doubling every quarter. On 19 May 2026 Google made Gemini 3.5 Flash the default model in AI Mode for everyone globally.

So anyone saying today that an AI query costs ten times as much as a Google search is comparing an AI answer with a product Google is in the process of replacing with AI answers.

This is where the evidence stops, however. Google publishes no figure for what an AI Overview adds to a search, and no current figure for a classic keyword search. The 0.24 Wh measured in the next section applies to a median text prompt in the Gemini apps, not to an AI Overview in Search. Equating the two would repeat the exact error behind the original myth. Estimates of how widespread AI Overviews are come only from SEO vendors and vary considerably: Ahrefs arrived at roughly 13 to 25 per cent of searches in 2025 depending on the month. Google itself gives no share.

Source: Google, Search I/O 2026 (19 May 2026). Prevalence estimates: Ahrefs, 2025.

Search engines that let you switch the AI off

Once the AI answer becomes the default in search, the most obvious lever is not your own prompting behaviour but your choice of search engine. Several providers leave the decision with the user.

• Ecosia. Since 2 July 2026, one click under Profile, Settings, "AI-free searching" turns off every AI feature, both the overviews and the AI chat. The setting persists.
• DuckDuckGo. AI Assist can be disabled permanently in the search settings, as can the Duck.ai panel. In image search, AI-generated images can be hidden.
• Brave Search. The AI answer block can be suppressed. The separate "Ask Brave" chat interface is not affected by that setting.
• Qwant. AI features require a free account, so they stay inactive for signed-out users.
• Startpage. Adds no AI layer of its own on top of the results.
• Google. The parameter &udm=14 returns the classic result list. It can be saved as a custom search engine in the browser, but it does not apply in AI Mode.

Why Ecosia is more than a tool tip for impact organisations

Ecosia is itself a social enterprise. Since 2018 it has been held in steward-ownership via the Purpose Foundation: shares cannot be sold at a profit and profits cannot be taken out of the company.

More relevant to this topic is the energy side. Ecosia does not buy certificates; it started building its own solar plants in 2018 and states that it now feeds roughly twice as much clean electricity into the grid as its own searches consume. The title of its blog post on this says, in effect, that carbon neutral is not enough. That is exactly the distinction made further down in the certificate trap section, drawn here by a provider itself: additional capacity actually built, rather than balance-sheet compensation.

For organisations that have to justify their choice of tools, that is a defensible argument. It rests on a physical installation, not on a label.

Sources: Ecosia: You can now choose AI-free search (2 July 2026); Ecosia: Why carbon neutral is not enough; overview of the opt-out routes: PPC Land, 6 June 2026. The figures on Ecosia's own electricity generation come from Ecosia and are not independently audited.

Efficiency leap 2024 to 2025

Google instrumented and measured its production Gemini infrastructure directly: a median text prompt consumes 0.24 Wh of energy and 0.26 mL of water, which the authors compare to less energy than watching nine seconds of television. The measurement covers the full stack including host system, idle capacity and data-centre overhead, not just the active chip.

Within one year, energy per prompt fell by a factor of 33 and the carbon footprint by a factor of 44. Studies based on GPT-3/GPT-4 data from 2022/2023 cannot capture this leap.

Important when quoting this: the 44 is not pure efficiency. The authors name two drivers, software efficiency and clean energy procurement. The procurement share is subject to exactly the certificate question discussed below. The factor of 33 for energy is the harder number, because it does not depend on electricity accounting.

Source: Elsworth et al. (Google), arXiv 2508.15734, August 2025. Direct measurement in the production Gemini infrastructure.

Rebound: why efficiency alone is not enough

The efficiency gain is real, but it does not reduce total consumption. It makes use cheaper and therefore more frequent. This pattern is known as the Jevons paradox, after the economist William Stanley Jevons: using a resource more efficiently ends up consuming more of it, not less.

The evidence: the computing effort required to train AI models doubles every five months. The number of ChatGPT users more than doubled within a year. The share of AI-specific hardware in data-centre energy consumption is rising from 14 per cent (2023) to an estimated 47 per cent (2030).

The honest statement is therefore: per query things improve quickly, in total they still grow quickly. Anyone quoting only the first half is arguing incompletely.

Source: Öko-Institut Consult for Greenpeace: Environmental impacts of artificial intelligence (May 2025), drawing there on Epoch AI.

Three footprints that do not always move in the same direction

The UN report (UNU-INWEH, June 2026) introduces a central finding that is almost entirely absent from public debate: carbon, water and land use are three independent dimensions. A decision that lowers one of these footprints can raise another.

Practical consequence: “carbon-neutral” is not the same as “ecologically harmless”. A data centre running on nuclear power has a low carbon footprint but high water use. A solar-powered data centre in a water-scarce region can cause a land conflict despite very low carbon emissions.

Source: UNU-INWEH: Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints (June 2026). Authors: Aczel, Chamanara, Matin, Farsi, Marwala, Madani.

Server location as a decision criterion

Users can influence the energy mix underlying their AI use through their choice of provider and region. The major providers now allow European server regions to be selected.

• Anthropic (Claude): API access runs via AWS regions, EU regions are available.
• OpenAI (GPT): EU data processing is possible via Azure regions. European customers can configure data residency in the EU.
• European electricity mix: On average considerably cleaner than the US mix, but not everywhere. Germany sits above the US average, see the next section. Anyone using region as a sustainability criterion has to look at the specific country, not at "EU".

Organisations that want to make their AI use transparent in their sustainability communication can document server location and energy mix as concrete criteria in their choice of provider.

The certificate trap

An important distinction: the carbon neutrality announced by the large providers for 2030 so far rests largely on purchased green electricity certificates. In reality, coal and gas power stations fill the gaps, because there are no direct lines from renewable generators to the data centres.

The Öko-Institut study therefore deliberately calculates without certificates. It considers them insufficient for a guaranteed clean supply and instead measures the real emissions of the electricity that physically reaches the data centres through the grid. On-site generation and PPAs are excluded too, because providers publish no reliable data on them.

For provider selection this means: a European server location is a real lever, because it sits on a measurably cleaner grid. A carbon-neutrality label is not, as long as it rests on certificates rather than on additional, directly connected generation. The robust question to ask a provider is therefore not "are you carbon neutral", but "which electricity mix supplies this region, and was the renewable capacity for it built additionally".

Source: Öko-Institut for Greenpeace: Regional Climate Divide of Data Centers (June 2026). The study considers the carbon neutrality pledged in the "Climate Neutral Data Centre Pact" for 2030 unattainable.

The regional climate divide

Commissioned by Greenpeace, the Öko-Institut analysed the electricity sources of around 6,700 data centres worldwide, based on the BloombergNEF database plus regional emission factors from Ember Energy and the US EPA. The metric is CUE, the carbon intensity of computing work, in grams of CO2 equivalent per kilowatt-hour of IT electricity.

The most important finding is in the fourth row. Germany, at 430 g, sits above the US average of 414 g. The good European average comes from countries with plenty of hydro and nuclear power, not from Germany. Within Europe the spread is wide, from Serbia at 911 g down to Scandinavian levels.

In practice this means: a server in the EU is first of all a data protection criterion. As a sustainability criterion it only works if it is a clean European location.

485 TWh is around 2 per cent of global electricity use, roughly Germany's entire demand. The 170 million tonnes exceed the emissions of Germany's entire transport sector.

By 2030 the gap widens: Europe falls to 195 g, the US rises to 424 g. Dedicated gas power stations built solely to run data centres push individual values as high as 727 g.

Between the hyperscalers themselves there is no meaningful difference. AWS, Google Cloud and Microsoft are practically level on emission intensity. Purely European providers do better. Switching from one US hyperscaler to the next therefore achieves nothing; the lever is the region and the type of provider.

Source: Öko-Institut for Greenpeace: Regional Climate Divide of Data Centers (June 2026). The study explicitly reports all absolute values as lower bounds: the dataset captures an estimated half of installed capacity, because providers treat their consumption data as a trade secret.

AI-specific projections to 2030

The following figures come from a different, earlier study by the same institute. The distinction between all data centres and AI-specific ones is what matters here.

AI data centres overtake traditional ones on emissions by 2030, a factor of 5.7. Traditional data centres even decline slightly, because falling grid emission factors more than offset their moderate growth. AI data-centre electricity demand in 2030 is eleven times the 2023 level. On top of that come up to 5 million tonnes of additional e-waste.

Source: Öko-Institut Consult for Greenpeace: Environmental impacts of artificial intelligence (May 2025). The projection explicitly does not account for individual operators purchasing lower-emission electricity. The figures are drawn differently from those in the Climate Divide study and are not directly comparable with them.

Global scale

The IEA figures are drawn differently from the Öko-Institut numbers in the two sections above. Do not read them as one series.

Water use: context

Water does not physically “disappear”. What matters: is it still available where it was taken from? Data centres deliberately evaporate water in cooling towers. For the region it is gone for good, which is especially problematic in drought areas.

Data centres draw 80 to 90 per cent on high-purity “blue water”: directly from rivers, lakes or the drinking-water network, the same water as in the local water supply. Avocados and grain, by contrast, live largely on rainwater (“green water”) within the natural cycle.

This means: the global water demand of AI is minimal compared to industrial agriculture (approx. 70 % of all freshwater withdrawals worldwide). The real problem is local and qualitative, not global and quantitative.

The volume is growing fast, however. Data centres worldwide used around 175 billion litres of water in 2023, projected to reach 664 billion litres by 2030, more than a threefold increase. AI data centres account for most of the growth: from around 30 to 338 billion litres. That intensifies exactly the local distribution conflict described here.

Sources: UNU-INWEH, June 2026; GI study, June 2025; Öko-Institut Consult for Greenpeace, May 2025

Why the figures diverge so widely

The range in the press is not a measurement error but a methodological problem.

System boundaries: is only the active GPU chip measured (underestimates real consumption by a factor of 2.4), or the entire site including upstream electricity generation? Both approaches yield valid figures, but for different questions.

Outdated efficiency assumptions: Many studies use hardware nameplate values from 2022/2023 and ignore algorithmic optimisations of the past two years.

Lack of transparency: No provider publishes AI-specific environmental metrics. Amazon (AWS) does not publish aggregated water-consumption data at all.

Methodological basis: Frontiers in Communication, March 2025 (14 models, 7B–72B parameters, direct measurement on NVIDIA A100).

Practical consequences

Based on current research, concrete behaviours can be derived that reduce resource consumption:

• Prompt discipline. Shorter prompts save energy proportionally. Pleasantries towards AI systems have measurable, if small, costs.
• Use AI images and videos sparingly. A single high-resolution AI image uses as much energy as a full smartphone charge.
• Limit agentic loops. Autonomous agents without a stop condition multiply consumption exponentially. Build in clear limits and manual approval steps.
• Small model first. For classification, extraction and structured tasks, a smaller model needs about 30 times less energy than a large one.
• Opt out of training data. Anyone using the OpenAI API should actively object to their input data being used for training purposes.

Not just consumption: who controls the cloud

Strictly speaking this does not belong to resource consumption, but it comes up in almost the same breath in most conversations. It needs different answers: consumption can be steered through usage discipline and region choice, market power cannot.

• AWS customer base. According to the Greenpeace report, AWS does business with at least 100 of 263 companies classified as non-investable under recognised financial-industry exclusion lists. 70 of them are excluded on ethical grounds by Norway's sovereign wealth fund, among them JBS, Shell, Palantir and Anduril.
• EU Cloud and AI Development Act. European Commission proposal of 3 June 2026: stricter criteria for public cloud contracts in banking, energy and health, including where software and hardware are developed and whether third countries control the provider.
• German EnEfG reform 2026. According to the Greenpeace response, the draft weakens PUE requirements, the obligation to use waste heat, and transparency duties for data centres.
• Ethics policy. Greenpeace has published a ready-to-sign voluntary commitment for providers of critical digital infrastructure: usage limits, supply-chain and environmental transparency, lobbying disclosure, and an independent ethics council.

Sources: Greenpeace: Amazon's Toxic Web Services (May 2026); Greenpeace Ethics Policy (May 2026)