There is no free lunch, says an old adage of economists. And in the case of Artificial Intelligence (AI), the menu is not only salty, but astronomically expensive.
Every economic enterprise consumes energy, but technological megaprojects devour enormous amounts of it, measured in “Terawatt hours” (1 “Tera” means 10 to the power of 12). To put it into perspective: in 2023, the planet consumed 27,400 TWh of electricity.
We have identified some energy guzzlers in the following enterprises.
The European Center for Nuclear Research (CERN) in Switzerland, which has not yet been bombed by Trump, spends 1.3 TWh a year, the equivalent of the annual consumption of 200,000 Swiss and 1.3 million Bolivians. We are still far from Switzerland, Mr. Evo Morales!
In turn, the Bitcoin cryptocurrency devours energy on a scale 100 times greater, between 120 and 180 TWh/year in 2024, similar to the consumption of Argentina and ten times that of Bolivia. On the other hand, the Visa card administrator is a moderate diner: 0.74 TWh/year, almost a candidate for the “Social Tariff”!
But artificial intelligence surpasses everyone. Between training models such as GPT, Grok or DeepSeek, data storage and server cooling, its current consumption is around 200 TWh/year… and continues to grow unstoppable.
Who pays the bill?
The energy to power AI comes mainly from fossil sources (with their carbon footprint) and, to a lesser extent, from renewable sources. That said, technology companies are looking with interest at nuclear energy: Microsoft has already entered into contracts with dedicated plants. Although it does not emit CO2 like fossil fuels, nuclear energy is subject to criticism for its risks and, due to increasingly stringent safety protocols, contributes to increasing the user’s energy bill. Other companies opt for carbon credits, a questionable patch, from several points of view.
The future is even more voracious: if Trump manages to invest the promised $500 billion in AI, consumption by 2028 could skyrocket to 325 TWh (the equivalent of 300 CERN and/or all of Italy!), requiring 0.1 TW, 10% of the energy capacity of the United States.
Artificial intelligence doesn’t just eat… Drink. It swallows water to cool the servers and evaporates up to 50% in the process. In 2023, the water consumption of AI companies was 66,000 million liters, and by 2028, this figure could double.
Are there any discounts?
While artificial intelligence swallows energy like a black hole, according to a recent report by the Spanish newspaper El País, the human brain runs on just 175 kWh per year (although some politicians seem to spend much less). Per year, the 7,000 million human beings would consume about a thousand TWh. The difference reveals the current inefficiency of artificial intelligence, compounded by misuse and trivial prompts such as “What color was Napoleon’s white horse?” (Answer: gray, because of the sweat to explain it so much).
Therefore, improving the efficiency of AI is not only a technological problem, but fundamentally a problem of user education, as is the case with the efficient use of electric lighting.
Is there any chance of a tip for Bolivia?
In this context, there is an opportunity for Bolivia: if it could host data centers, it could monetize its potential solar and hydroelectric power and get a good chunk of this technological feast. Of course, politicians should first convince themselves that AI is not an offspring of the devil, nor an invention of imperialism, but a source of income.
Acknowledgments: the original version of this column has been revised, corrected and compressed by AI (DeepSeek, precisely), to which the author thanks for the valuable collaboration, apologizing for the unnecessary use of the energy consumed.