Source: S&P Global Media Portal via S&P Global.
The data center segment has moved from being a largely predictable real estate and infrastructure category to a matter of national politics. Only a few years ago, the global market’s growth appeared steady and manageable: expanding at mid-single-digit rates, addressing a balanced mix of hyperscale and organizational demand, and measuring “large” builds in tens of megawatts. Today, the conversation has shifted to gigawatt-scale pipelines, grid constraints and — increasingly — water usage.
The COVID-19 outbreak compressed years of cloud adoption into months, with hyperscalers rapidly absorbing capacity to meet work-from-home demand. Then, widely accessible generative AI pushed the industry into a new era of computation. Accelerator-focused architecture has fundamentally changed facility design assumptions, spurring higher power density and making energy procurement, transmission planning and load growth central to understanding data center expansion.
The Take
Water consumption has emerged as a primary issue in the data center industry alongside the power bottleneck. As high-density facilities scale, cooling choices increasingly determine not just operating cost, but regional viability. Water stress is rising while disclosure of resilience planning lags. Evaporative systems can lower energy use, but consume water. Mechanical systems reduce water use but raise energy demand. In this environment, water consumption is more than a branding exercise — it is a siting constraint, a permitting risk and a core design variable for the next generation of digital infrastructure. Public vitriol, however, has been pointed in the wrong direction. As such, data centers are not as water consuming as they are being portrayed in the media. The technology deployed to generate power to operate data centers is the real culprit. Nuclear generation — the new face of clean energy — is among the worst.
Why does AI increase demand so dramatically?
Traditionally, data centers were based on CPU computing while AI uses GPU, or more broadly, accelerator-based computing. Accelerators employ something to the order of 10 times the power of traditional CPUs on average. They are also being deployed at unprecedented scale. Going back to 2018, folks were wringing their hands over power densities of 20-30 kilowatts per rack — we are now hearing that hyperscalers are requesting up to 2,200 kW per rack.
One might wonder about efficiency increases. Although new generations of technology are becoming more efficient, the efficiency gains require further increases in power draw. At least for now, things are becoming more efficient, but the power continues to scale up alongside the efficiency gains.
Water stress
This condition occurs when there is more forecast water demand than water supply. Water is a key resource across the majority of sectors — two-fifths of countries’ gross value-added capacity relies on water, with agriculture being the most dependent.
Water stress is projected to be the second-largest contributor to climate financial impacts in the S&P Global 1200 Index, with impacts reaching $265 billion by 2050.
At the same time, water-intensive industries are booming. Global data center power demand is projected to nearly double between 2024 and 2030, with 43% of existing data centers already located in regions of high water stress. Roughly 60% of data center assets in China and 38% of data center assets in the US are projected to be exposed to high water stress this decade.
Although projections indicate that all sectors will experience at least moderate exposure to climate hazards by 2030, only 42% of companies disclosed adaptation and resilience plans to address the looming climate hazards.
As the growth rate in some high-risk markets exceeds planned resilience measures, particularly regarding water use, companies need to take a step back and consider how to deal with the increasing risk of water stress.
The importance of water in data centers
Data centers deploy water in two distinct ways. There is an internal loop that connects to the IT equipment to take heat away from the facility, and an external loop that goes outside the building to chillers and cooling towers to dissipate that heat. It is common to hear operators state that they run a “closed loop” system, meaning that water is neither added nor removed from the loop once it is filled. But what loop are they talking about?
All data centers operate a closed-loop system for the internal water loop, with the only water consumption happening at initial fill-up and marginal losses due to maintenance. The external loop is where water usage can occur at scale. Finally, it is important to note that when organizations talk about “liquid cooling,” they are referring to the internal loop, which uses very little water — sometimes none at all, as the liquid inside the liquid cooling systems could be a dielectric fluid.
The external loop runs to large-scale water-cooling towers. When water is sprayed onto a material, the subsequent evaporation of that water creates a cooling effect on the material, and then water passing through the material is cooled, in a very similar way to how the human body keeps itself cool via perspiration. Water loss occurs from both evaporation and what amounts to “overspray” caused by wind across these systems.
Of course, this is not the only option, although it is commonly used. There are also compressor- and coolant-based cooling systems that operate much like those seen in modern vehicles. These systems do not use large amounts of water like the evaporative systems — however, they employ 10% to 20% more energy compared with the evaporative systems. So there is a trade-off here: either use a lot of water and less energy, or less water but more energy.
The question then becomes, what is the greater sin? This could largely depend on location. If the location has a strong water supply but poor or dirty power generation, evaporative cooling might be the better option. Vice versa, in areas with high water stress but a robust power infrastructure, the non-evaporative system could be a better choice.
Data center water use
When we talk about water usage, we must first talk about water usage effectiveness. WUE measures the amount of water consumed by a data center, for every kilowatt hour of electricity it uses. In other words, a data center with a WUE of 1 would consume 1 liter of water per kWh. A more applicable example would be a 10-MW data center with a WUE of 1.8 that would consume 41.65 million gallons of water per year.
Most of the US hyperscalers do publish water usage metrics. Google comes in at the highest by far, with over 7.7 billion gallons of water usage in 2024. Microsoft Corp. is next at just over 1.5 billion gallons. Meta Platforms Inc. is lowest at 785 million gallons. Amazon Inc. does not report water usage, but it claims a WUE of 0.15, which is the lowest of the group. The US Department of Energy has made some further estimates on this metric, suggesting that the average WUE for all data centers is roughly 1.8 — much higher than the hyperscalers’ estimates.
Water use in other industries
Google has adopted an interesting approach in its filings and compares the water usage of its data centers with the water usage of golf courses. The organization claims that the overall yearly water consumption (7.7 billion gallons) is equivalent to the annual consumption of about 51 golf courses. For reference, the state of Florida has over 1,200 golf courses. Similarly, the growth of corn across the US consumes upwards of 150 billion gallons annually, most of which is used for livestock feed or ethanol production.
More relevant for the current discussion, the energy industry also employs enormous amounts of water. While data centers on average in the US consume 1.8 liter of water for every kWh of energy used, the weighted national average for power production is 7.6 L of water consumed per kWh, for thermoelectric and hydroelectric generation. Hydroelectric plants evaporate as much as 68 L of water per kWh, due to water surface evaporation in reservoirs. While significant, hydroelectric power only accounts for about 6.5%-7% of the generation in the US.
Thermoelectric generation is the leader in the US, with natural gas and nuclear combined generating about 55% to 60% of the total, with higher percentages in certain regions. In PJM, for example, which serves the largest data center market in Northern Virginia, gas-fired power accounts for roughly 40% of the total generation, while nuclear accounts for about 30%.
On average, a nuclear power plant will consume between 1.5 and 3 liters of water per kWh. The US nuclear fleet consumes over 300 billion gallons annually. Recall that the hyperscalers, minus Amazon, consumed just under 10 billion gallons in 2024 globally. This throws a wrench in the sustainability rallying cry of the nuclear renaissance. While nuclear is clean in terms of emissions, it consumes enormous quantities of water and increases the temperatures of natural bodies of water upon its discharge.
The real issue
Given the relative water efficiency of hyperscale data centers, indirect water usage via power generation has a far larger impact. Interestingly, hyperscalers are not addressing the issue of indirect water usage the way they address indirect emissions. It is common for hyperscale operators to offset the emissions released by the generation deployed to power the data centers, even though the facilities themselves are not releasing the emissions. It is also true that the same generation is using significantly more water on average than the data centers themselves. While some organizations are working hard to replenish water consumed by data centers, there is still a double standard in terms of the water used by power generation.
What is an AI Datacenter?
Want insights on datacenter trends delivered to your inbox? Join the 451 Alliance.
