During a golden sunset in Memphis in May, Sharon Wilson pointed a thermal imaging camera at Elon Musk\u2019s flagship datacentre to reveal a planetary threat her eyes could not. Free from pollution controls, the gas-fired turbines that power the world\u2019s biggest AI supercomputer were pumping invisible fumes into the Tennessee sky.<\/p>\n
The Guardian\u2019s journalism is independent. We will earn a commission if you buy something through an affiliate link.\u00a0Learn more<\/a>.<\/p>\n \u201cIt was jaw-dropping,\u201d said Wilson, a former oil and gas worker from Texas who has documented methane releases for more than a decade and estimates xAI\u2019s Colossus datacentre was spewing more of the planet-heating gas than a large power plant. \u201cJust an unbelievable amount of pollution.\u201d<\/p>\n That same week, the facility\u2019s core product was running riot on news feeds. Musk\u2019s maverick chatbot, Grok, repeated a conspiracy theory<\/a> that \u201cwhite genocide\u201d was taking place in South Africa when asked about topics as unrelated as baseball and scaffolding. The posts were quickly deleted but Grok has gone on to praise Hitler<\/a>, push far-right ideologies<\/a> and make false claims.<\/p>\n Grok, Elon Musk\u2019s AI chatbot, has previously praised Hitler and repeated conspiracy theories about \u2018white genocide\u2019. Photograph: Algi Febri Sugita\/ZUMA Press Wire\/Shutterstock<\/p>\n \u201cIt\u2019s a horrible, horrible waste,\u201d said Wilson, the director of campaign group Oilfield Witness, pointing to Grok-generated images of Nazi Mickey Mouse<\/a> as an example of what fossil gas was being burned to produce. \u201cWhat useful purpose does this serve?\u201d<\/p>\n Wilson is not alone in asking this question. Scientists are watching the AI boom with unease as it pollutes the natural world with carbon and the digital world with dangers ranging from dodgy health myths<\/a> to deepfake pornography targeting children<\/a>.<\/p>\n Some experts fear datacentres may derail the shift to a clean economy, adding an unnecessary hurdle to the quixotic task of keeping the planet from heating 1.5C (2.7F). Others are sanguine about the energy costs, arguing they pale in comparison not just to polluting industries, but also to the technology\u2019s power to reshape society.<\/p>\n How big a threat is AI to the climate? And could it help more than it harms?<\/p>\n When Hannah Daly ran models for the International Energy Agency (IEA) in Paris and then became a professor of sustainable energy at University College Cork in Ireland, she spent little time worrying about the carbon cost of computers. Cars, cows and home heating were among the knotty problems draining the carbon budget; emissions from digital services were small and relatively flat.<\/p>\n But in Ireland, the computational hunger for energy has reached levels that are impossible to ignore. Datacentres guzzle one-fifth of the country\u2019s electricity and are projected to devour nearly one-third of it within a few years. The rapid expansion of chip-filled warehouses, which have sprung up faster than the electricity grid can handle, led to them being effectively banned from grid connections in 2021.<\/p>\n The trajectory of \u201cenormous, exponential growth\u201d is what is so worrying, said Daly. \u201cI\u2019m not sure whether Ireland is a real outlier or a harbinger of what\u2019s to come. But it\u2019s definitely a cautionary tale.\u201d<\/p>\n Datacentres consume just 1% of the world\u2019s electricity but may soon demand much more. Their share of US electricity is projected to more than double to 8.6% by 2035, according to BloombergNEF, while the IEA projects datacentres will account for at least 20% of the rich world\u2019s growth in electricity demand to the end of the decade.<\/p>\n Some of the demand is being met through long-term agreements to buy renewable power \u2013 supporting the expansion of clean energy even when the electricity powering the facility is dirty \u2013 while some tech firms have signed deals to use nuclear energy.<\/p>\n But for the near future, fossil fuels are set to dominate supply. China\u2019s datacentres are clustered in its coal-heavy east. In the US, where natural gas is expected to generate most of the electricity in datacentres over the next decade, the Trump administration used them to justify burning more coal. \u201cBeautiful, clean coal will be essential to \u2026 winning the AI race,\u201d energy secretary Chris Wright said in September when announcing a $625m (\u00a3467m) investment package.<\/p>\n An Amazon datacentre in Didcot, Oxfordshire. Photograph: Horst Friedrichs\/Alamy<\/p>\n In Ireland, which is building terminals to import liquefied natural gas (LNG) and power plants to burn it, the datacentre boom has offset climate gains from renewable expansion in the power sector, according to an analysis<\/a> Daly conducted for Friends of the Earth Ireland last year. Poorer countries may not be immune, either. Cheap solar has begun to displace coal at a remarkable rate in Pakistan, but datacentres are poised to pick up the spare capacity of idling power plants after the government announced it would devote 2GW of power to AI and bitcoin.<\/p>\n \u201cThis idea that the lower cost of renewables alone will drive decarbonisation \u2013 it\u2019s not enough,\u201d said Daly. \u201cBecause if there\u2019s a huge source of energy demand that wants to grow, it will land on these stranded fossil fuel assets.\u201d<\/p>\n Does this mean using chatbots to draft emails, write essays and plan holidays endangers the planet? Tech companies have resisted pressure to provide detailed data on their AI energy footprints, but popular guesstimates hover about 0.2 to 3 Watt-hours (Wh) for a simple text query and rise considerably for \u201cdeep research\u201d and video production. In a blogpost in July, OpenAI\u2019s CEO, Sam Altman, claimed a ChatGPT query uses as little energy as running a lightbulb for a couple of minutes, in line with a recent report from Google on the median text prompt to its AI assistant, Gemini.<\/p>\n The figures are negligible compared to activities such as flying, eating meat, or driving a car. Still, sceptics worry about the sheer scale of the technology \u2013 ChatGPT claims to have several hundred million weekly users just three years after its release \u2013 and the fervour with which companies have shoved it into every aspect of digital life. Google controls an estimated 90% of the global search engine market and has forced generative AI on to its results page. The rise of AI agents and behind-the-scenes services is expected to add to the background hum.<\/p>\n \u201cWhat I\u2019m worried about is that we\u2019re deploying AI in such a way that we don\u2019t have a good idea of the energy use,\u201d said Sasha Luccioni, climate lead at AI company Hugging Face, who has grown frustrated by \u201cselective disclosures\u201d from big companies that obscure the real climate impact of their products. \u201cWe\u2019re essentially operating on the hypothesis that it\u2019s not a problem \u2013 or that if it is a problem it will somehow be solved \u2013 instead of getting ahead of it.\u201d<\/p>\n What if AI could pay off its energy debts by saving carbon in other parts of the economy? That is the thesis put forward in an IEA report in April, which argued existing AI applications could cut emissions by far more than datacentres produced. A paper<\/a> by researchers at the London School of Economics and Systemiq came to a similar conclusion in June after modelling scenarios in which AI would help to integrate solar and wind into the electricity grid, identify alternative proteins that mimic meat, improve battery composition in electric cars and nudge people to make climate-friendly choices.<\/p>\n \u201cAI can accelerate the deployment of those clean technologies by basically accelerating their position along the curve of innovation and adoption,\u201d said co-author Roberta Pierfederici, a policy fellow at the LSE\u2019s Grantham Institute.<\/p>\n The projected carbon savings carry large uncertainties \u2013 greater efficiency can lead to more use, the IEA warns, and rebound effects may undercut the gains, such as self-driving cars undermining public transport \u2013 but examples<\/a> already exist. Google says AI has helped it cut cooling in datacentres by 40%. Spain\u2019s Iberdrola says AI has optimised wind turbine maintenance and performance, increasing operational efficiency by 25%. France\u2019s Engie says it has reduced downtime on solar farms by using AI to detect faults.<\/p>\n Because other sectors are so polluting, the researchers say, AI would only need to cut their emissions by a small fraction to cover the carbon cost of its computing, which a recent study<\/a> estimated to be 0.1-0.2% of global emissions, and rising. \u201cIn the power sector we are already seeing the results,\u201d said Pierfederici. \u201cWhereas the meat sector is not there yet.\u201d<\/p>\n Clean tech advocates are not the only ones to have noticed AI\u2019s transformative potential.<\/p>\n When Holly and Will Alpine decided to quit their jobs at Microsoft last year, they knew they were throwing away a good deal. The married millennial couple enjoyed US tech salaries through their positions in the company\u2019s responsible AI and sustainability teams, with close colleagues and work that gave them a sense of purpose. Will had been among the early voices pushing to tackle the energy cost of datacentres.<\/p>\n But Microsoft\u2019s work for oil and gas clients troubled the pair, and they started to grow more concerned about the emissions it enabled than the ones it produced. In 2019, the company announced a partnership with ExxonMobil with the potential to expand production by as much as 50,000 barrels a day. That same year, it began a digital project with Chevron that the oil company says has cut 30 days off its deepwater well planning process. As more deals emerged, the Alpines began to push their employer for answers.<\/p>\n \u201cThe response from the company was often pointing back to their own operational footprint, which is not relevant,\u201d said Holly Alpine, who left the company with Will to campaign for the tech industry to tackle its enabled emissions. \u201cAfter a four-year internal advocacy campaign, where we got a lot of promises but most were unfulfilled, we realised that internal pressure was not enough.\u201d<\/p>\n The chief executive of oil company Saudi Aramco recently said the company had embedded AI \u2018in everything\u2019. Photograph: Bloomberg\/Getty Images<\/p>\n The IEA estimates that AI could boost technically recoverable oil and gas reserves by 5% and cut the cost of a deepwater offshore project by 10%. Big oil is even more bullish. \u201cArtificial intelligence is, ultimately, within the industry, going to be the next fracking boom,\u201d Mike Sommers, head of the American Petroleum Institute, told Axios<\/a>. Amin Nasser, chief executive of Saudi Aramco, said the company had embedded AI \u201cin everything\u201d in an interview with Bloomberg Television<\/a> earlier this year. The world\u2019s biggest oil company doubled its technology spending from 2023 to 2024, according to Nasser, and the widespread adoption of AI has \u201cincreased productivity, and with that, the number of wells\u201d.<\/p>\n At the same time, the oil and gas industry says AI can cut its carbon intensity, for instance by analysing satellite data to spot methane leaks. But even here, critics say there is a gap between digital insights and corporate actions. Wilson, who saw \u201cgiant clouds of gas escaping everywhere\u201d during a recent field trip to the Permian Basin, said the industry\u2019s sophisticated network of satellites have achieved little because leaks are only a small problem compared to intentional releases of methane.<\/p>\n \u201cThey are using this as an excuse to delay action,\u201d said Wilson. \u201cWatching methane from space is not stopping methane.\u201d<\/p>\n Perhaps even more concerning than the expansion of fossil fuel supply is the effect on consumption. Generative AI adverts outperform human ones, a study<\/a> found in October, and the ease with which they can be made slashes the cost of encouraging consumption. The marketing industry, already familiar with hyperpersonalised adverts and streamlined shopping, is preparing for AI agents that could buy presents and book flights on a customer\u2019s behalf. Tui, Europe\u2019s biggest travel operator, says it is investing heavily<\/a> in AI as people turn to ChatGPT to book their holidays.<\/p>\n \u201cThe narrative is really focused on this false comparison between the energy used to run the technology and the positive use cases,\u201d said Alpine. \u201cBut it is dangerous to omit the negative use cases.\u201d<\/p>\n Some voices are calling to hit pause, at least until better rules are in place. In October, the UN\u2019s special rapporteur on the human right to safe drinking water called for a moratorium<\/a> on the development of new datacentres, citing the damaging environmental impacts. In December, a coalition of more than 230 environmental groups in the US demanded a national moratorium<\/a> until they were regulated. Ireland\u2019s utility commission has lifted its de facto ban on grid connections, but said 80% of a datacentre\u2019s annual electricity consumption must eventually come from new renewable projects.<\/p>\n