AI needs to reduce its water dependency

Artificial intelligence (AI) continues to attract considerable attention in financial markets. While much of the spotlight is on the competitive opportunities offered by AI strategies, soaring demand has caused supply chain challenges. However, water resources are rarely viewed as an important constraint. If the water consumption of data centres is not dramatically reduced, it will be an increasingly important factor in their location.

AI and datacentres

The expansion of data centres is closely linked to the growth of AI. Rising demand for computational power, especially inference (training AI) and deep learning models, requires graphics processing units (GPUs) and custom made tensor processing units (TPUs) which are hosted in data centres. Data intensive AI applications, notably for natural language processing (NLP), need the vast amount of data and data centres provide such solutions. Cloud computing and AI as a Service (AIaaS) optimize workload distribution across servers, and use data centres instead of requiring companies to build their own infrastructure. It is also worth noting that Edge Computing, where data processing happens closer to the source of data generation, has encouraged the development of decentralised data centres, often located close to users, supporting low latency and real-time applications. To help optimise these requirements AI itself is often used to design and achieve operational efficiency.

To meet this demand, considerable investment is required. Estimates vary, often dependent on exactly what is defined as AI spend. Goldman Sachs estimate that “leading tech giants, other companies, and utilities to spend an estimated ~$1tn on capex in coming years, including significant investments in data centres, chips, other AI infrastructure, and the power grid”.ⁱ The International Data Corporation (IDC) estimates that “global spending on artificial intelligence (AI), including AI-enabled applications, infrastructure, and related IT and business services, is projected to reach $632 billion by 2028”.ⁱⁱ After software spend, “spending on AI hardware, including servers, storage, and Infrastructure as a Service (IaaS), will be the next largest category of technology spending”.ⁱⁱ

Supply chain strains

Soaring demand, has led to strains in the AI supply chain. Among the key components in this supply chain is hardware, such as the semiconductors and computer chips that run the AI algorithms and the data centre infrastructure such as the servers, networking equipment, storage and cooling systems.

Grabbing the headlines have been the shortage of computer chips (integrated circuits), demand for experience AI professionals and the rise in demand for electrical power. Some forecasts predict a doubling of data centre power consumption in the US, reaching 35GW by the end of the decade, almost double its 2022 level.ⁱⁱⁱ

However, rising water consumption should be receiving similar attention.

Water usage by data centres

Water requirements can be measured based on withdrawal – where water is taken from a source and later returned to it, or consumption – where water is “lost”, usually through evaporation.

Water (often potable water) is used by data centres for cooling, ensuring that servers do not overheat, but there is also an indirect usage of water, through the water requirements of non-renewable electricity generation

Water can be a key part of power generation, especially in fossil fuel and nuclear plants, and US average water intensity for electricity generation was 2.18 litres per kilowatt hour in 2015. The renewable energy transition is important for both GHG and water intensity. It is estimated that moving to wind and solar energy could reduce water withdrawals by 25% in the US.^iv

Data centre cooling systems use water as it efficiently absorbs and transfers the heat. The water may be circulated through pipes to absorb heat from the equipment. The heated water is then moved to a cooling tower or other cooling mechanisms where it is cooled down and recirculated. The largest data centres often use chilled water systems.

Cooling towers are also evident, using evaporation to cool water, which is then recirculated through the data centre’s cooling system. A steady supply of water is needed to keep this system operating effectively. Such evaporative cooling can be more energy-efficient than mechanical cooling options (e.g. air conditioning),

To guarantee continuous operation, data centres often have redundant cooling systems, Reliable access to water is critical for these backup systems to function if called upon.

To optimise water use, operators will use water use efficiency (WUE) by controlling the amount of water use relative to their energy consumption. In turn this has led to innovations in reducing water usage, in some instances eliminating the need for water. However, we note that less than one-third of data centre operators tracked any water metrics and water conservation was ranked as a low priority.^iv Companies such as Google have been called out for considering water use to be a trade secret and for preventing the release of information about how it works with local utilities.^iv

Improving water management

As water demand increases increased regulation is likely to follow. The Association for Computer Operations Management (AFCOM), which represents data centre and IT infrastructure professionals, focused on this topic at the end of last year, highlighting “the need for innovative cooling technologies and sustainable solutions; the significance of community engagement, including cooperation with local municipalities; and a proactive approach to assessing and planning for future water needs”.^v

As mentioned above, there are alternatives being tested as solutions for managing the heat generated by the IT equipment. Direct air cooling is possible in cold climates – using naturally cool outside filtered air – or passing the air through a refrigerant unit, known as indirect air cooling.

Refrigerants can also be used instead of water as they have a lower boiling point and can absorb more heat energy per unit volume. Although these systems are like air conditioning, they are often more efficient and do not require water.

Some operators of data centres have opted for immersion cooling. Users of single-phase immersion submerge servers in a non-conductive liquid that absorbs heat directly from the components. The heated liquid is then circulated to a heat exchanger, where it is cooled and recirculated. This eliminates the need for traditional water-based cooling in favour of specialized non-conductive liquids (dielectric fluids).

Other non-water cooling technologies being trialled include geothermal cooling by leveraging the relatively stable temperatures underground for cooling, phase-change materials (PCMs): which can store and release large amounts of energy as they change from one state to another; these are particularly useful for handling peak loads or short-term cooling needs.

Global data centres

The US hosts the largest number of data centres in the world, accounting for 39% of the total. (See Figure 1).^vi There are 2,928 data centres in the US and 3,168 in North America, out of the 7,597 globally. By comparison, Western Europe accounts for 1,888 data centres and Asia 788. Note that Figure 1 is not ranking the data centre by size. Alternative rankings use data centres density – i.e. how much power the centre consumes in peak load. Physical space is also used sometimes.

Figure 1: Countries ranked by total number of data centres with water stress category (Sources: Data Centre Map [accessed 21 August 2024] and baseline water stress score, from Aqueduct, WRI^vii)

Among these 15 top ranked countries, 3 are classified as countries with high (Italy & Turkey) or extremely high water stress areas (India). Under WRI forecasts to 2050, Australia is expected to join the high water stress category. It’s worth noting that presently only one country (Switzerland) is viewed as low water stress risk, although a further 6 are classified at low to medium, although one country (France) is forecast to move into the medium-high category by 2025. Clearly the main country to watch is the USA, which remains in the medium to high water stress category. Below, we examine the USA in more detail.

Data centres and water stress in the USA

In Figure 2, we show the location of over 2,500 data centres^vi mapped against water stress data. We used the “baseline water stress” indicator from WRI Aqueduct tool, which measures the ratio of total water demand to available renewable surface and groundwater supplies. It reveals that although 30% of data centres are in areas on low water stress, 32% are in are within areas of high or extremely high water stress. For example, there are 68 data centres in the Los Angeles and a further 32 in the San Jose areas, accounting for 36% of all the data centres in California. Other states with water supply challenges include Arizona (86), Colorado (55), Nebraska (17), Nevada (40), New Mexico (9), and Wyoming (11). The state with the most data centres, by a large margin, is Virginia (474), accounting for 16% of the country’s total.

Figure 2 Data centre locations in the USA mapped against water stress data. The map excludes Alaska. (Sources: DataCentreMap, WRI and Planet Tracker).

The building of new data centres is increasing demand for water resources. Some data centres are presently located in areas of water stress or are likely to be in the future. Developing cooling technologies which minimise or do not require water is becoming increasingly important. Perhaps AI will find a scalable solution to this problem.

REFERENCES