Big U.S. tech companies are surging in water consumption, AI will trigger a "water war"?
If you look at the latest environmental reports of large US technology companies, you can find one thing in common: their water consumption has surged. Google's 2023 environmental report shows that the company's water consumption increased by 20% year-on-year last year to 5.6 billion gallons (about 21.2 billion liters), which is nearly enough to fill 1.5 West Lake. Microsoft's water use for the whole of last year increased by 34 percent year-on-year to nearly 1.7 billion gallons, which is roughly equivalent to the total storage capacity of 2,500 Olympic swimming pools. One of the reasons for the surge in water consumption by tech giants is the rapid development of artificial intelligence (AI) around the world. As for whether high water consumption will limit the development of AI, the Global Times reporter also asked the AI chatbot ChatGPT, which replied: "Water consumption may indirectly pose a challenge to AI development, but it is not the main factor limiting the development of AI." ”
Billowing water vapor billowed from Google's data center in The Dalles, Oregon. (Google Inc. website)
Every 10-50 conversations, ChatGPT "drinks" a 500ml bottle of water
According to CNBC, a study by Ren Shaolei, an associate professor in the Department of Electrical and Computer Engineering at the University of California, Riverside, found that ChatGPT "swallows" 500 milliliters of water every 10 to 50 conversations, and hundreds of millions of users ask questions on this platform every month, which shows how "thirsty" AI models are. Researchers in Ren's team warn that if the problem of AI "thirst" is not addressed, it could become a major obstacle to the sustainable development of AI in the future.
In fact, it is not the AI model itself that uses water, but the cooling system of the data center. It is understood that when the AI model is computing, the data center generates a lot of heat, and in order to keep it running smoothly, water resources need to be consumed to cool it. Most of Google's increased water consumption last year was spent on cooling data centers. The Register, a British technology outlet, reports that data centers vary greatly in the amount of water used to cool down depending on how they are built, where they are located, and the cooling technology used.
Deploying data centers in low-temperature environments is an important means for tech giants to reduce cooling costs. For example, Facebook (now Meta) built its first overseas data center in 2013 in the small Swedish town of Luleå, just about 150 kilometers from the Arctic Circle and where temperatures have been above 30 degrees Celsius for less than 24 hours since 1961. Evaporative cooling technology is also a common cooling solution in warehoused data centers, that is, the evaporation of water absorbs heat from the air. The technology consumes less power and only needs to be used during the hottest months of the year. In addition, Microsoft said this year that it would use "zero water" cooling for two data centers at its Goodyear campus in Phoenix, Arizona, but that it would use more power.
According to the Global Times, compared with traditional air-cooling technologies, such as natural ventilation or air conditioning cooling, liquid cooling technology is currently considered to be better able to reduce the PUE value (i.e., power utilization efficiency, which is an indicator to evaluate the energy efficiency of data centers) and achieve the purpose of energy conservation and emission reduction. According to the different contact methods between the liquid and the heating device, the liquid cooling technology can be roughly divided into cold plate type (indirect contact), spray type and immersion type (direct contact). Immersion liquid cooling technology is widely recognized as more promising and is known as the "next generation of thermal management technology".
Google's data center cooling tower in The Dalles, Oregon. (Google Inc. website)
Is data centers becoming the new front of the "water wars"?
More than 30% of the world's data centers are located in the United States, and according to a study by Virginia Tech, operating data centers is one of the top 10 most water-intensive commercial industries in the United States. It is worth mentioning that some American technology companies "preference" data centers in places where water is scarce, because they value the low electricity prices and abundant clean energy. This move has often caused dissatisfaction among the local people, and some media have bluntly said that this may lead to a "water war".
When Jane Duff, a member of the Mesa City Council, Arizona, learned that Meta was going to build another data center there, she immediately questioned: "My first reaction was to worry about our water." According to the Washington Post, the 500,000-resident "desert city" of Mesa has been loaded with large data centers from tech giants such as Google and Apple.
This is not unique in the drought-prone western United States. In The Dalles, Oregon, Google was forced to publish information that its data centers had "usurped" more than one-quarter of the city's water resources, and in Los Lunas, New Mexico, farmers protested against the city government allowing Meta data centers access. Researchers say a large data center consumes as much water per day as a town with a population of 10,000 to 50,000.
In addition to posing a threat to domestic water use, U.S. multinational Internet companies are also trying to build data centers overseas. Earlier this year, Google's announcement of plans to build a data center in southern Uruguay sparked outrage and led to protests, according to The Guardian. Daniel Pena, a researcher at the University of the Republic of Uruguay, said that the water used in Google's data centers will be taken directly from the country's public drinking water system, and that Uruguay is currently suffering from the worst drought in 74 years, and the capital Montevideo has declared a water emergency.
In Chile, which is also facing water stress, the city of Quilicura, located in the capital region, is favored by tech giants. Google built a data center in the area back in 2015 and announced an expansion in 2018. However, residents of the Chilean suburbs of Santiago held a referendum in February 2020 against the center's construction, which, although non-binding, persuaded Google to adopt a cooling system that consumes less water. Microsoft also announced a data center project in Kilikula in December 2020, and protests are still ongoing.
If data usage continues to grow at an exponential rate, the world may need to build 10 to 20 times as many data centers by 2035, and Latin America, where water and electricity are cheap, is seen by the IT industry as a "cheap place" for such expansion, according to environmental website Mongabay. Some analysts argue that the U.S. "exporting" data centers to the Global Global South is not an economic opportunity, but a new form of exploitation, data colonialism.
Yan Zhanyu, deputy director of the National Security Computing Laboratory at the University of International Business and Economics, told the Global Times that water consumption will become a hidden concern for AI development for three main reasons: first, large amounts of water use and water-saving measures will raise the hidden costs of enterprises; second, enterprises need to bear the risk of international and domestic policies restricting water use; in addition, the concerns of the public, especially environmental activists, have forced companies to spend energy to deal with opposition, which has increased the resistance to the development of AI.
However, Yan Zhanyu believes that even if there are these three hidden concerns, the development of AI is unlikely to slow down, mainly because the technology competition and application competition of AI have become the most cutting-edge track in the field of modern industry and national security. "Water consumption is an important issue, but not an urgent one. Therefore, whether it is the state, enterprises or the scientific community, they will not put environmental protection issues above safety and technical issues, and sacrificing a certain amount of environmental protection to ensure enhanced competitiveness may be the trend of AI development for a considerable period of time. ”
The generational gap between AI and water-saving technologies makes it challenging to achieve "water positive benefits".
ChatGPT, the most advanced large language model from the Open Artificial Intelligence Research Center (OpenAI), was born in Iowa, USA, and they pump water from the confluence of the Nguyen Kun River and the Des Moines River in the central part of the state to cool supercomputers, according to Fortune magazine. OpenAI responded in a statement a few days ago that "we recognize that training large models can consume a lot of energy and water" and is "seriously considering" how to make better use of computing power to improve efficiency.
Tech giants that have been criticized for "swallowing too much water" from their big models are trying to salvage their reputations, with companies such as Meta, Google and Microsoft pledging to achieve "water positive" by 2030. According to these tech companies, this means that the water they give back to their communities will exceed the amount of water they use for their direct operations. Microsoft has said it is tackling water use in two ways: by reducing water intensity and by replenishing water supplies in water-scarce areas where it operates. Microsoft's 2022 Environmental Sustainability Report, released this year, shows that the company has signed a series of water offset projects that are expected to provide more than 15.6 million cubic meters of water, bringing Microsoft's total water offset to 35 million cubic meters.
In addition, countries are beginning to roll out responses at the government level. According to the Global Times, relevant policies and assessment documents have been introduced in China to optimize the water efficiency of data centers, such as the Beijing Municipal Development and Reform Commission, which was revised and issued in July this year, which was called "the strictest in history" by the media, "Several Provisions on Further Strengthening the Energy Conservation Review of Data Center Projects". According to Barron's, the Biden administration proposed in April that the federal government be allowed to limit the amount of water pumped from the shrinking Colorado River in Arizona, California and Nevada.
On the other hand, Business Insider commented that there is a huge environmental cost to running a large data center, and that "it is only going to get worse" as the AI race heats up and AI tech companies race to build new data centers.
Yan Zhanyu told the Global Times that at present, the development speed of AI technology is much faster than the development speed of water-saving technology, even if the technology giants propose to achieve "positive water resource benefits", but the generation gap between the two technologies will bring a lot of obstacles. In addition, companies such as Google are moving domestic water scarcity to less visible places like Uruguay, which could obscure the severity of the problem and make it difficult to expose.
China will build the world's first commercial submarine data center near Hainan Island
Liu Liang, dean of the School of Artificial Intelligence of Beijing University of Posts and Telecommunications, told the Global Times that the problem of AI water consumption can be slightly alleviated by optimizing the location of the computing center and optimizing the scheduling of computing power, but the fundamental problem is that AI large model training has a huge demand for computing power and brings huge energy consumption, which has also become a bottleneck restricting the development of AI technology. A common view in the academic community is that it is important to break through the structure, training methods and hardware environment of existing large models, and develop in the direction of simplified models, small-sample learning Xi, and low-power computing.
Yan Zhanyu told the Global Times that it is more important for companies and countries to realize that water conservation is an issue as important as technological development. At the same time, solving this problem requires strengthening international cooperation, and the international community needs to set a "bottom line" for the development of artificial intelligence.
The British media outlet The Register advises against training AI models in hot climates. Ren's team also suggested setting up a battery backup system for the data center, so that solar energy stored during the day could be used when training large models at night when the temperature was cold, so as to solve the conflict between water efficiency and carbon emissions. In addition, there is a need to increase the transparency of data center WUE (Water Efficiency).
The Global Times reporter paid attention to the fact that some progress has been made at home and abroad in solving the problem of AI water consumption. On November 24, China completed the installation of a "submarine data capsule" at a depth of 35 meters near Hainan Island, where the world's first commercial submarine data center will be built. After the completion of construction, compared with the traditional data center on land of the same scale, the data center can save 122 million kilowatt hours of total electricity per year, save 68,000 square meters of construction land area, and save 105,000 tons of fresh water every year. The Spanish newspaper El País quoted Meta as saying that the company plans to develop an electric drying system specifically for AI data centers that does not require water (except for power generation).
In addition, there is also an argument that AI poses not all threats to water resources. According to the official website of the World Future Energy Summit, AI is changing the way we use water. It can analyze the flow of water in pipelines in real time, alert or automatically shut down in the event of anomalies such as water leaks, which is an important tool for water-scarce countries to avoid water waste, and AI can also enable "smart irrigation", which is at the heart of efficient water use on "next-generation farms".
So, will AI bring about a "water war" or a "water future"? Yan Zhanyu said that even though AI can optimize the management of water resources through smart water systems and smart agriculture, the problem of water competition caused by training AI models has not yet been solved. In other words, if a company has good technology to treat water, it often means that it will take up more water. In the long run, the "water future" also faces the question of competition from international standards, that is, to what extent is it acceptable for AI to be used in this water cycle and water reuse, and which country's standards should be referred to?