The amount of water required in generating energy and cooling data centres that run AI models is referred to as the water footprint of AI. The water footprint has two components: direct water consumption and indirect water consumption.
The water evaporated or released as waste during the cooling of data centre servers is referred to as direct water consumption. The water needed to generate the electricity that powers the data centre servers is referred to as indirect water consumption.
AI’s water footprint varies depending on factors such as the size and type of AI model, the efficiency and location of the data centre, and the source of energy generation.
Statistics
- A new study titled “Making AI Less ‘Thirsty:’ Uncovering and Addressing the Secret Water Footprint of AI Models” discovered that training large AI models such as GPT-3 can consume up to 700,000 litres of clean freshwater. When it comes to consumption in this context, “water cannot be recycled,” according to academics at US-based universities, which is where the majority of the AI’s water usage is concentrated.
The unpublished research, which was submitted as a pre-print on arXiv, also estimates that a conversation with the AI chatbot in a single system may “drink” a “500ml bottle of water.” - According to Down to Earth research, this amount of water is equivalent to generating 370 BMW cars or 320 Tesla electric vehicles. Furthermore, utilising AI chatbots like ChatGPT in chats might consume up to 500 ml of water for 20-50 queries and responses. Given that ChatGPT has over 100 million active users who are engaged in many discussions, this water use adds up. The future GPT-4, which is planned to be significantly larger, is expected to raise water use even more, while precise estimations are difficult due to restricted data availability.
- Data centres, where AI models are housed, consume a lot of energy and water. In 2021, Google’s US data centres alone used 12.7 billion litres of freshwater for cooling. These facilities use water-intensive cooling systems and demand large volumes of water for power generation.
- According to the study, if the data had been generated in less energy-efficient data centres in Asia, water consumption may have climbed threefold. Another study asserts that the number could rise further with the recently introduced GPT-4 AI system, which has a higher model size. “The water footprint of AI models can no longer be ignored,” the researchers said, adding that the water footprint must be addressed “as a priority as part of the collective efforts to combat global water challenges.”
Challenges
- AI’s water footprint contributes to water shortage by consuming large amounts of freshwater to cool AI infrastructure, putting a burden on limited water resources. This exacerbates the worldwide water scarcity problem.
- Freshwater extraction for AI operations can affect aquatic biodiversity and have a negative influence on the ecosystem. Furthermore, the energy used for water purification and transportation for AI activities contributes to carbon emissions and climate change, increasing environmental issues.
- Water diversion for AI activities may result in unsustainable resource management. This water diversion may impede access to water for human consumption, agriculture, and other important purposes, potentially increasing communities’ water scarcity difficulties.
- AI’s water-intensive nature can have equity and social consequences. Water scarcity disproportionately affects vulnerable groups who rely on restricted water supplies for a living. AI’s water needs could exacerbate existing injustices by diverting water away from communities that need it the most.
- Water is required for many stages of the development and manufacturing of AI hardware components, such as semiconductors, including fabrication and cleaning. The extraction and processing of raw materials, like rare earth minerals utilised in AI systems, can potentially influence the environment.
- AI algorithms can be tuned to reduce their environmental impact. To achieve such optimisation, resource utilisation and efficiency must be carefully considered throughout the AI development lifecycle, including algorithm design, deployment, and system operation.
Solutions
- The use of renewable energy sources should be encouraged to power AI infrastructure, such as solar or wind power. The water footprint associated with traditional energy generation, which frequently relies on water-intensive operations like cooling in power plants, can be lowered by moving to clean energy.
- Improve the efficiency of data centres, which consume a lot of water due to cooling needs. Reduce water usage by implementing modern cooling techniques such as direct-to-chip liquid cooling or closed-loop systems.
- There is a need to optimise data storage and management practises to reduce needless data transfers and storage, hence lowering the amount of energy and water required for data processing.
- There is a need for collaboration among AI industry stakeholders, researchers, and policymakers to share best practices, technologies, and approaches for lowering AI’s water footprint.
- Such AI research and development need to encourage that prioritises water efficiency. This includes creating algorithms and models to reduce water consumption during training and operation.
- Governments can play a critical role in incentivizing water-efficient practices in AI development and operations by enacting legislation and regulations. Water pricing schemes, environmental certifications, or required reporting of water usage by AI corporations are examples of this.
- Raise awareness among AI developers, users, and the general public about AI’s water footprint and the importance of environmentally friendly practices. Education campaigns can urge individuals and organisations to examine the environmental effects of their AI applications and promote responsible AI usage.
As we investigate the potential of AI, we must keep the surroundings in mind. With freshwater scarcity and droughts on the rise, environmentally sound practices are critical. Addressing AI models’ hidden water footprint is critical for a sustainable future.
