The rise of artificial intelligence (AI) over the past two years has sparked a frenzy of building data centres, the infrastructure that makes it possible for this technology to work. Regardless of the effect of China's DeepSeek, which has shown that models can be developed with fewer resources, US President Donald Trump announced in his first week in office an ambitious plan that will mobilise $500 billion over four years to build data centres and ensure sufficient energy supply. The growth of this market in the US is beginning to put the country's energy grid on the ropes, which is looking for imaginative formulas (including the authorisation of pocket nuclear reactors) to add power to the system. But the dark side of AI production goes beyond its voracious consumption of energy and water, used to cool the equipment, or the generation of more and more electronic waste. It can also affect the health of citizens, as the White House recently acknowledged for the first time.
One of the latest executive orders signed by former president Joe Biden (EO 14141, 14 January), entitled Advancing US Leadership in AI Infrastructure, implicitly acknowledges that data centres are harmful to health. The document, which has not been revoked by Trump (unlike another executive order that sought to reduce the risks posed by AI ‘to the citizenry and national security’, which was overturned), guarantees the transfer of federal land for the construction of this type of infrastructure, which it considers ‘strategic’ for the country's economy and security, as well as the development of sufficient energy capacities to power them, including nuclear energy if necessary.
One of the sections of the executive order sets out the requirements to be met by the locations where these new data centres are to be built. Topographically suitable land, respect for surrounding wildlife and cultural resources, proximity to high-voltage power grids and good connections... And another: ‘Location within geographic areas that are not at risk of persistent nonattainment of the National Ambient Air Quality Standards, and where the overall risk of cancer due to air pollution is at or below the national average according to the EPA's AirToxScreen 2020 tool’.
‘The executive order recognises for the first time the public health impact of data centres. While the order is likely to be modified or rescinded, the explicit recognition of decreased air quality and increased cancer rates highlights the immediate risk of AI data centres,’ says Shaolei Ren, associate professor of electrical and computer engineering at the University of California, Riverside, and a specialist in AI sustainability.
Ren has co-authored a scientific paper, still under review, on the economic and human cost of illnesses directly attributable to pollution associated with data centres. Basically, the problems are caused by toxic gases, such as nitric oxide or PM2.5 particles, expelled during the process of generating the electricity that powers the plants. These gases come both from the power plants that feed the data centres, which are usually only a few kilometres away, and from the back-up generators they have in case the regular supply is cut off. Ren's figures are lumpy: the health bill for AI was between $17 billion and $29 billion between 2019 and 2023. During that period, according to his calculations, a minimum of 1,100 premature deaths were caused.
Data centres are no more or less polluting than other industries. The difference, according to Ren, lies in the growth rate of this sector and the apparent ignorance about the pollution it generates. ‘It is well known that cars pollute, which is precisely why there are strict regulations to control and reduce the gases they emit. But data centres are growing so fast that by 2028 they will exceed the emissions of the entire California vehicle fleet even if we add 35 million vehicles, according to the Department of Energy's recent projection of data centre energy demand’.
What has the Trump administration said about this? Nothing directly, but it has given some signals. The first, very significant, is that it has not revoked Biden's executive order, which means that it does not reject its contents. The second came just this week: the new director of the Environmental Protection Agency (EPA), Lee Zeldin, presented on Tuesday an initiative entitled Driving America's Great Comeback in which, while not even mentioning CO₂ or greenhouse gases, he places ‘clean air’ as the first pillar to ‘protect the health of people and the environment’. It also emphasises ensuring that AI data centres ‘can be powered and run cleanly with American-made energy’. ‘In this context, the term ‘clean’ seems to refer to low emissions of air pollutants,’ Ren notes.
Where to build what type of facility
The report provides another clue: more than half of the data centres in the US by 2028 will be colocation. This is the industry jargon for facilities that house third-party data and, unlike those developed for big tech, known as superscalar, ‘are typically located in densely populated urban areas’, Ren stresses. ‘These data centres, as well as those specialising in AI, which have a much higher energy demand than the rest, and facilities with downwind populations, are the ones that will have the greatest impact on health,’ he says.
Biden's executive order proposes to minimise the impact on the health of data centres by deciding where to do what. For example, model training, which is the most energy-intensive and therefore the most polluting, could be done in facilities located in desert areas, and less energy-intensive tasks could be left for data centres located in densely populated areas. ‘The proposal is to prioritise data centre projects in locations where air quality is good and cancer rates due to air pollutants do not exceed the national average. In other words, where public health inequalities are not further exacerbated,’ says Ren.
A recent study carried out by the scientist's team used the locations of Meta's data centres in the USA as a case study to analyse the impact on health of following this decision scheme (training models in locations with a low health impact, i.e. sparsely populated). The results are that, taking as a reference the 2023 activity reported by the company itself, the health bill (i.e. the cost of doctors associated with respiratory diseases and cancer caused by pollution) could be reduced by 30%.
Source: El País