|
||
|
||
AI’s water woes Apart from its prodigious consumption of energy, AI’s water usage is adding to its environmental costs. Lean Ka-Min WHILE it’s well known that artificial intelligence is power-hungry, less widely recognised is the fact that AI is also extremely ‘thirsty’. The same server computers that consume massive amounts of energy in running AI models require lots of water to cool down. Any discussion of AI’s environmental impact will therefore have to look not only at the carbon footprint it generates from its energy use but also at its ‘water footprint’. And it’s shaping up to be one large, wet imprint – global demand for AI could end up extracting more water in 2027 than the equivalent of 4–6 Denmarks. This is cause for concern at a time when over 2 billion people live in countries with inadequate water supply and 4 billion experience severe water scarcity for at least one month each year.1 Water is gulped down in huge quantities by data centres housing the servers that run AI applications and other online cloud services that have become part and parcel of everyday life for many, such as email and Internet search engines. The considerable heat generated by the energy-intensive servers has to be expelled from the data centres, usually through either cooling towers or outside air cooling – both of which need water, plenty of it. In a cooling tower, some of the water that had been heated up after being circulated through the server room is evaporated to dissipate the heat into the outside environment, and the remaining, cooled water is pumped back into the system to absorb further heat. This water can only be recycled a few times, however, before the accumulated salts and minerals render it unsuitable for use in the pristine setting of a data centre. Fresh supplies of clean water – often the same potable water needed for human consumption – must therefore be constantly added to make up for the evaporated water and discharged water.2 Instead of a cooling tower, outside air can be blown through the servers to cool them down. Nevertheless, water is still required for evaporation when the outside air is too hot, and for humidity control when the air is too dry.3 This factor becomes all the more important in scorching tropical countries earmarked as potential data centre sites. Besides their direct use of water, AI data centres’ thirst extends to the water consumed by the power plants generating the electricity on which they run. As in the data centres themselves, much water is needed to cool down thermal power (e.g., coal and natural gas) and nuclear plants, while water is lost through the expedited evaporation from hydropower generation.4 A further indirect form of water usage is embedded in the chips and servers doing the data-crunching for AI applications. Ultrapure water is needed for wafer fabrication, and water’s cooling properties are called upon by semiconductor plants too.5 What all these add up to is some gargantuan guzzling. Google’s, Microsoft’s and Meta’s data centres worldwide together extracted an estimated 2.2 billion cubic metres of water in 2022, roughly equivalent to the total annual water withdrawal of two Denmarks. Not all of this can be attributed to AI, of course, since data centres run a gamut of other cloud applications. But with AI contributing one of the fastest-expanding workloads in these centres, their need for water is set to soar. Already, Google’s data centre water usage shot up by 20% from 2021 to 2022, and Microsoft’s by 34% in the same period. It has been estimated that the combined water withdrawal of global AI could amount to 4.2–6.6 billion cubic metres in 2027 – greater than the annual water withdrawal of 4–6 Denmarks.6 Take GPT-3 for example, the ‘large language model’ on which the popular ChatGPT AI chatbot runs. A 2023 study by four US-based researchers found that training GPT-3 in Microsoft’s US data centres can consume a whopping 5.4 million litres of water. On a more quotidian scale, engaging GPT-3 in a question-and-answer session consisting of 10–50 responses will see the model ‘drink’ a half-litre bottle’s worth of water. These figures will likely increase with the newer, larger GPT-4 iteration.7 As water scarcity intensifies, more people are voicing concern about AI’s ‘drinking problem’. Residents of West Des Moines in the US’ Iowa state – home to a data-centre cluster running GPT-4 – filed a lawsuit which revealed that the cluster used some 6% of the district’s water in July 2022, the month before training for the model ended.8 In drought-stricken Chile, Google’s plan to build a data centre in Cerrillos, Santiago, has faced local pushback, culminating in a court ruling calling on the firm to take into account the effects of climate change on the Central Santiago Aquifer and to revise the design of the centre’s cooling system.9 Such design changes – which may entail a greater reliance on air cooling or purifying non-potable water, for instance – are among the measures proposed to shrink AI’s water footprint. In addition, where and when AI is trained and deployed also matter, since its water usage differs according to temperatures outside the data centre and the energy sources employed by the local electricity grid (thermoelectric plants require far more water than their solar- and wind-powered counterparts). ‘For example,’ says Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside, ‘AI consumes 1.8–12 litres of water for each kWh of energy usage across Microsoft’s global data centres, with Ireland and the [US] state of Washington being the most and least water-efficient locations, respectively.’ 10 Despite research undertaken by the likes of Ren, AI’s water-guzzling ways have thus far not drawn widespread scrutiny, certainly less so than its energy use. Kate Crawford, a University of Southern California Annenberg professor specialising in the societal impacts of AI, cautions: ‘Without better transparency and more reporting on the issue, it’s impossible to track the real environmental impacts of AI models. And this matters at a time when many parts of the planet are experiencing deep and extended droughts, and fresh drinking water is already a scarce resource.’11 More clarity may be on the way. Under the European Union’s Artificial Intelligence Act, recently approved by the European Parliament, ‘high-risk AI systems’ (including the ‘foundation models’ that power ChatGPT and similar applications) will have to report their energy consumption, resource use and other impacts throughout their lifecycle. In the US, a bill introduced in Congress would require the government to assess and establish a standardised system for reporting AI’s environmental impacts. Meanwhile, the International Organization for Standardization (ISO), which develops international production standards for goods and services, is drawing up criteria for ‘sustainable AI’ that will cover energy efficiency and water consumption, among others.12 A clearer picture of AI’s environmental footprint should in turn raise greater awareness that, while the technology’s potential is said to be sky-high, its use of natural resources has to be grounded in the reality of scarcity. Water that is essential to the survival of humankind must not be sucked dry by clusters of thirsty machines, however ‘intelligent’ they may be. Lean Ka-Min is editor of Third World Resurgence. Notes 1. UNICEF. ‘Water scarcity’. https://www.unicef.org/wash/water-scarcity 2. P. Li, J. Yang, M.A. Islam and S. Ren (2023). ‘Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models’. https://arxiv.org/abs/2304.03271 3. Ibid. 4. Ibid. 5. Ibid. 6. Ibid. 7. Ibid. 8. K. Crawford (2024). ‘Generative AI is guzzling water and energy’. Nature, Vol. 626, 22 February. 9. S. Moss (2024). ‘Chile partially reverses Google data center permit over water use concerns’. Data Centre Dynamics, 28 February, https://www.datacenterdynamics.com/en/news/chile-partially-reverses-google-data-center-permit-over-water-use-concerns/ 10. S. Ren (2023). ‘How much water does AI consume? The public deserves to know’. OECD.AI Policy Observatory, 30 November, https://oecd.ai/en/wonk/how-much-water-does-ai-consume 11. Quoted in: C. Criddle and K. Bryan (2024). ‘AI boom sparks concern over Big Tech’s water consumption’. Financial Times, 25 February, https://www.ft.com/content/6544119e-a511-4cfa-9243-13b8cf855c13 12. D. Berreby (2024). ‘The growing environmental footprint of generative AI’. Undark, 20 February, https://undark.org/2024/02/20/ai-environmental-footprint/ *Third World Resurgence No. 359, 2024/2, pp 21-22 |
||
|