Jevons paradox: why more efficient AI doesn't necessarily mean savings

Logic dictates that the more efficient a technology is, the fewer resources it consumes. But economist William Stanley Jevons showed as early as the 19th century that this isn't how it works. When steam engines became more efficient and used less coal, operating costs fell — and demand for coal skyrocketed instead. This exact scenario now threatens the field of artificial intelligence as well.

The new UN report titled The Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints, highlighted by 100+1 zahraniční zajímavost, warns that the growing efficiency of AI models will actually lead to even greater consumption, not savings. Cheaper operation means wider deployment, new use cases, and a higher overall burden.

"When technology gets cheaper, people start using it everywhere. This is a principle that has held true since the invention of the wheel," comments one of the report's authors. And with AI, it seems, it will be no different.

Chilling data: electricity, water, and emissions

The UN report estimates that by 2030, AI-related energy consumption could double and reach a level equivalent to 3% of global electricity production. For comparison — that's roughly the same share consumed today by the entire country of Japan, the world's fifth-largest economy.

No less alarming is the view on emissions. According to the report, AI's carbon footprint could by 2030 match the total emissions of the United Kingdom — a country of 67 million people, including its industry, transport, and agriculture combined.

The biggest shock, however, comes from the water figure: cooling AI data centers could consume more water than the total annual drinking water consumption of all humanity. Data centers require enormous amounts of water for server cooling — and with each new, more powerful chip, this need continues to grow.

Specific historical figures

As early as 2021, Google's AI division consumed 10–15% of the total 18.3 TWh the company drew from the grid that year. Training the GPT-3 model required 1,287 MWh — roughly the annual consumption of 121 American households. And that was a model from 2020. Today's models like GPT-5 or Gemini 3 are an order of magnitude more demanding.

The International Energy Agency (IEA) noted in its 2024 analysis that the capital investments of Google, Microsoft, and Amazon in data centers in 2023 for the first time exceeded the investments of the entire US oil and gas industry — together representing roughly 0.5% of US GDP.

What the tech giants are doing about it

The major players are aware of the problem — and are investing massively in renewable energy. Amazon, Microsoft, Meta, and Google are the world's four largest corporate buyers of green energy and have so far contracted nearly 50 GW of renewable sources through PPAs (Power Purchase Agreements), equivalent to the generation capacity of the entire country of Sweden.

Both Google and Microsoft have also committed to running their data centers on exclusively carbon-free energy 24 hours a day, 7 days a week by 2030. This is a significantly more ambitious goal than merely "annually matching" consumption by purchasing green certificates — it means genuinely covering every megawatt-hour in real time.

The problem, however, is that these commitments are failing to keep pace with the rate of growth. According to the IEA, data center electricity consumption in some US states has already climbed above 10% of total consumption. In Ireland, it exceeds 20% — and the country had to temporarily suspend construction of new centers due to grid overload.

European context: regulation that can help

For the Czech Republic and the European Union, the topic is particularly timely. As early as 2024, the EU introduced mandatory reporting of data center energy performance (under the revised Energy Efficiency Directive, EED). Every data center with an installed capacity above 500 kW must, since May 2024, report total energy consumption, the share of renewable sources, water consumption, and the rate of waste heat reuse.

Furthermore, the CSRD (Corporate Sustainability Reporting Directive) has, since 2024, required large technology companies to report their environmental impact in detail. This creates pressure for transparency — and at the same time gives regulators data on which to base minimum performance standards.

The Czech Republic is a member of the IEA, and these European regulations fully apply to it. While the country has not yet seen such dramatic growth in data centers as Ireland or the Netherlands, with the increasing digitalization of businesses and public administration, the question of energy intensity will become ever more pressing.

What readers should know — and what they can do

For the average user of ChatGPT, Gemini, or Midjourney, it is practically impossible to estimate how much energy a single query consumes. Expert estimates vary significantly — a single chat query can, depending on complexity, represent 0.001 to 0.01 kWh, roughly the equivalent of running an LED light bulb for several minutes to several tens of minutes.

The key factor, however, is scale: while one family has one light bulb at home, ChatGPT is used by hundreds of millions of people daily. And with the rise of AI agents that autonomously perform tens or hundreds of queries in sequence, the ecological footprint multiplies.

The UN report therefore calls for the adoption of principles of transparency, equity, and sustainable resource use. And as with other environmental challenges, the rule holds true: technology alone won't solve the problem. It will take a combination of regulation, innovation, and — most importantly — conscious use.