Edara Sreenivasa Reddy
Comingup with one creation after another, today everyone is an alternate Brahma, with each person using artificial intelligence to create new videos, photos, and reels according to whatever comes to one’s mind and whatever they like. Ironically, these ‘Brahmas’ are inflicted with a strange malady-They are enjoying seeing how they would have looked back in the 1980s. In this category are those born before 1980 and are without any old photographs in their collection.
As those born after 1980 do not have that opportunity, many of them have descended upon Gemini AI, ChatGPT and the like. Each one has started creating hundreds upon hundreds of photos. In India, 70 to 100 million are active AI users every day. Of them, around 35 million (38 per cent) use ChatGPT, Gemini AI and similar tools, while away time in the name of entertainment. The figure crosses 50 million when viral trends rule the roost. Among Gen-Z (ages 16-24) youth, 60 per cent spend hours on reels, face-swaps, and memes alone.
Is there any benefit to society from this? None. But they are unable to realize that these activities are causing considerable harm to society, especially to the environment. It is not just them; those who make reels on social media, those who make short videos, those who create avatars through animation, well just about everyone is swaying to the AI magic; it is like each person is contributing a handful of sesame seeds and causing their share of harm to the environment. Because of them, not only is carbon dioxide increasing in our surroundings, but the water resources are also evaporating.
The reason for this creation upon creation is the generative AI models powered by artificial intelligence. Like the proverbial dwarf growing ever larger, along with their form, their usage is also increasing. Not only that, they are expanding in terms of words, photos, music, and videos that they produce and users and creators. For generative AI models to create new things, they must first be trained. As the days go by, not only will these models be trained, but more models will come along. Both their training and usage involve not only financial expenses but also numerous associated costs. These models emit enormous amounts of carbon dioxide. At the same time, they consume water on a massive scale. The data centres and millions of high-end graphics processing units (GPUs) that operate on these systems must run continuously. To control the enormous heat that is generated, water is continuously circulated through cooling towers. For instance, to run data centres 24x7 hours a day, enormous amounts of water are required. Ultra-pure water is used directly to reduce the heat of servers, in power plants, and in plants that manufacture AI chips and semiconductors.
Data centres that power generative AI models generate electricity, which produces heat that can be cooled by water.
Water use by these can be broadly classified into three types-One is direct water use for on-site data centre cooling, two is off-site water for electricity generation and three is water for server manufacturing. At data centres, cooling towers evaporate water to dissipate the heat. Around 80 per cent of the non-evaporated water can be reused and that too only a few times. All of them are potable water that can prevent pipes from clogging or bacteria from growing. Another method is to use air for cooling without evaporating water in more temperate climates.
It is the training of these models (i.e., the highest level of computational processing that takes place before they are used) that creates the high energy and cooling requirements. Generative AI (Generative AI) has been extremely useful for various tasks across industries, particularly in the media sector. Training and using these models also cause environmental harm. Within this broad scope, water consumption (water footprint) has not been discussed much. As the climate crisis intensifies, clean water is becoming an increasingly limited resource. In both continuously training new models and using them, we all need to pay more attention to the impact these models have on the environment. In a world where clean water is becoming scarce commodity, it is essential to be aware of the water consumption of these models.
During the training process of GPT-3 (an outdated 2020-era model belonging to OpenAI), seven lakh litres of fresh water evaporated. In 2023, Google data centres consumed 2,900 crore litres of fresh water to cool their systems. By 2027, it is estimated that 4.2 to 6.6 trillion litres of water will be required worldwide for AI needs.
For example, the Philadelphia Water Department estimated that an average person’s daily water consumption is 384 litres.
By this calculation, for GPT-3 training, the amount of water required is equivalent to what one person would consume over 1823 days. That is equal to the annual water consumption of five people. The water consumed by Google data centres in 2023 was equivalent to the annual water consumption of 2,06,906 people (75,520,833 days).
A single AI video 5 to 10 seconds long evaporates approximately 4.1 litres of water. By 2027, according to worldwide demand, it is estimated that training these systems will equal the annual water consumption of 30–47 million people.
If everything is added together, it is no exaggeration to say that a few top companies are going to use half the water on Earth. According to United Nations reports, currently more than 220 crore people worldwide do not have access to safe drinking water. Another 350 crore people experience severe water stress for at least one month a year. Our country too is going to become one of the countries facing an extremely severe water crisis, the UN is expressing concern. According to the NITI Aayog Composite Water Management Index. Nearly 60 crore people in the country are facing severe water shortage. Approximately, 25 per cent of all groundwater extracted in the world is reported in India alone. On the other hand, our country is becoming a hub for data centres. Cities such as Mumbai, Chennai, Hyderabad, Noida, Visakhapatnam are rapidly becoming data centre hubs. Experts point out that in these metro cities, which depend on tankers for drinking water in summer, the operation of massive servers is putting severe pressure on local people’s drinking water. They warn that if this excessive trend continues unchecked, humanity will inevitably face difficulties in the future.
Carbon emissions can be reduced by using solar energy, but unfortunately water consumption cannot be reduced. We need more efficient processes today that effectively balance the needs for both carbon and water.
The first step toward changing that is to highlight this issue and create awareness among people. So, if we are responsible in the way we use these models and use them only as much as we need, we will be protecting the environment. If you are someone who uses generative models repeatedly and excessively, try to balance it by reducing your other water consumption, such as taking shorter showers, there is no better way than that.
(The writer is Professor – VITAP; Artificial Intelligence Researcher)