Ask ChatGPT how much energy it takes to answer a user’s query, and it will tell you it’s as much as one kilowatt-hour (kWh) of electricity. That's roughly equivalent to ten hours of light from a single, 100-watt light bulb.
With more than 200 million active weekly users, that's a lot of lightbulbs. In less than two years, the generative AI chatbot has turned the world of large language models (LLMs), which most artificial intelligence is currently built on, from a niche tool for techies into an everyday service for the wider public.
From translation tasks to dinner ideas, AI platforms are spitting out answers quicker than any Google search – and taking more energy to do it. Then there are the higher order demands like vaccine development, policing oversight and medical diagnostics, all of which AI promises to take on and improve beyond any human capability.
According to the International Energy Agency (IEA), energy demands for data centres and networks – essential for keeping the world digitally connected and operating AI platforms – already account for at least 1% of both global electricity use and greenhouse gas emissions. That's a GHG footprint roughly equivalent to that of the aviation sector.
With the European Union committed to reaching net-zero emissions by 2050, a debate is growing over the future of the “energy hog,” as an investment analysis called generative AI last year. Large language models, which have become a standard method for building AI systems, are particularly energy intensive, especially during their training phase.
“We have to take a sufficiency approach, meaning that we have data centres for the things that we need – not for every single thing under the sun,” Anastasia Tsougka, a programme manager at ECOS, an environmental standards coalition, told The Parliament.
Digital defenders point out that energy consumption of data centres trails the internet usage that it serves, thanks to advances in efficiency. Data centres’ energy needs have remained "modest,” according to the IEA, compared to a 25-fold increase in web traffic since 2010.
“The sector walks its talk. Everything is far more efficient,” a spokesperson for DigitalEurope, a technology industry association, told The Parliament. "Profitability, competitiveness and sustainability, in our world, are all tied together.”
Limits of efficiency
Data centres and networks consume only a fraction of the electricity used by the residential and industrial sectors. Tech advocates argue that AI could play a role in making these sectors more efficient, thereby partly offsetting the impact of making buildings “smart.”
Still, the IEA rates the global data centre ecosystem as "more efforts needed” – indicating that while it’s making progress towards decarbonisation and curbing energy use, the sector is set to miss net-zero targets by the middle of the century. Despite the push to use renewables, natural gas remains a trusted source of energy.
As AI applications add to the demand on data centres and the hive of semiconductors that run hot to power them, keeping energy consumption to a minimum is likely to become an increasing challenge.
There may be limits to how much the consumption curve can be flattened. A recent Goldman Sachs analysis of AI's impact on data centres noted that efficiency gains have "dwindled” since 2020, while power consumption has soared alongside the growth of AI-powered applications. Tech companies have seen a surge in their greenhouse gas emissions, largely due to AI developments.
"The energy efficiency of data centres will hit a wall at some point,” Tsougka, from ECOS, said.
According to a recent IEA forecast, the growing use of AI, alongside cryptocurrencies, could add to global electricity demand the equivalent of Sweden or Germany's annual consumption by 2026. Ireland, a European base for tech giants, may see up to one-third of its electricity going towards data centre tasks by then. The sector there already consumes at least as much electricity annually as the country's urban residential buildings.
Twin transitions, partially linked
This presents the European Union with a conundrum. AI puts additional pressure on climate targets, and the EU wants to aggressively pursue both. The bloc is home to 16% of the world's 8,000 data centres, with the European Commission estimating that 3% of EU electricity powers these facilities. Meanwhile, the EU's regional energy grids are among the oldest in the world and may need hundreds of billions of euros in upgrades so they can handle the AI demand.
That scale of investment gets to the heart of the competitiveness report presented by Mario Draghi in early September. The former European Central Bank president, who coined the phrase "whatever it takes” to save the euro more than a decade ago, now wants to save the EU from "slow agony” in a global struggle for socioeconomic and geopolitical relevance. He has framed the struggle as existential.
For the EU to survive, Draghi is pushing decarbonisation and AI development as top priorities for the bloc to pursue. That builds on the EU’s existing framework of “twin” transitions – digital and green – outlined during Ursula von der Leyen's first term as European Commission president.
The Commission has said that these twins can complement each other, but Green MEP Alexandra Geese told The Parliament she sees a "missed opportunity” for the last Commission to not explicitly link them.
“Among the many good things that the former Commission did, the very big mistake was not to intertwine the twin transition,” says Geese, who sits on the European Parliament’s Industry, Research and Energy Committee. “We could have achieved [putting] the digital transition in the service of the green transition and vice versa.”
The Commission declined an interview request, pointing instead to policies already in place that aim to regulate data centres for their environmental, energy and climate impacts.
Efficiency label for data centres
More than 500 data centres have signed up to a voluntary EU code of conduct for efficiency standards since 2008. Environmental and energy regulations are also included in the AI Act, passed earlier this year. The Energy Efficiency Directive, updated last year, requires some data centres to disclose their power consumption and performance indicators.
That information is intended to feed a Commission database to eventually develop an efficiency rating system for data centres – similar to what applies to household appliances, but more complex. However, the 15 September deadline for a first round of reporting has come and gone.
Geese was unaware of the delay, and both a Commission spokesperson and the DigitalEurope spokesperson could only speculate as to the reason for it. It is unclear when companies running EU-based data centres might be able to fulfil this requirement.
Given the web of legislative and regulatory requirements kicking into gear, Geese sees little additional appetite for more in the next term, saying now is the time for “enforcing what we did in the past term.”
However, the data centres themselves may be only part of the energy equation. The training and running of AI are another. A broader debate is heating up between proponents of the promise of AI and those more sceptical of its far-reaching benefits.
OpenAI’s CEO Sam Altman, in the US, is seeking to actively push an exponential expansion of data centres that would bring AI into every corner of modern life. Californian Governor Gavin Newsom recently vetoed an AI safety bill, concerned it would limit innovation. Major tech companies, some based in his state, opposed the bill.
From this side of the Atlantic, lawmakers like Geese are closely watching these efforts as they consider the best approach to pursuing the technology. How the EU can compete with the US and China in this digital realm, as the Draghi report calls for, while setting its own rules on resource consumption may be the forthcoming Commission’s big balancing act.
“The more energy something consumes, the more you want to see what that consumption is giving you,” Geese says.