Feature The explosive growth of datacenters that followed ChatGPT's debut in 2022 has shone a spotlight on the environmental impact of these power-hungry facilities.

But it's not just power we have to worry about. These facilities are capable of sucking down prodigious quantities of water.

In the US, datacenters can consume anywhere between 300,000 and four million gallons of water a day to keep the compute housed within them cool, Austin Shelnutt of Texas-based Strategic Thermal Labs explained in a presentation at SC24 in Atlanta this fall.

[1]

We'll get to why some datacenters use more water than others in a bit, but in some regions rates of consumption are as high as [2]25 percent of the municipality's water supply.

[3]

[4]

This level of water consumption, understandably, has led to concerns over water scarcity and desertification, which were already problematic due to climate change, and have only been exacerbated by the proliferation of generative AI. Today, the AI datacenters built to train these models often require tens of thousands of GPUs, each capable of generating 1,200 watts of power and heat.

However, over the next few years, hyperscalers, cloud providers, and model builders plan to deploy millions of GPUs and other AI accelerators requiring gigawatts of energy, and that means even higher rates of water consumption.

[5]

According to researchers at UC Riverside and the University of Texas Arlington, by 2027 global AI demand could [6]account for the withdrawal of 4.2-6.6 billion cubic meters of water annually. That's roughly the equivalent of half the UK's water withdrawal over the course of a year.

However, mitigating datacenter water consumption isn't as simple as ditching evaporative cooling towers for waterless alternatives.

The datacenter water cycle

Datacenters consume water in a couple of ways. The first, and the area we'll focus most of our attention on, is direct water consumption. This is water that's pulled from local sources including water and wastewater treatment plants.

This water is pumped into cooling towers, where it evaporates, transferring heat to the air. If you've ever used a swamp cooler to chill your home or apartment, cooling towers work in a similar manner.

Evaporative cooling has become popular among datacenter operators for a couple of reasons, but the big one is they're really good at getting rid of heat and don't require a ton of electricity to do it.

[7]

According to Shelnutt, evaporating ten gallons a minute is enough to cool roughly 1.5 megawatts of compute.

When we talk about "consumption," we're referring to water that's been evaporated. It isn't actually consumed so much as it's removed from the local watershed by the prevailing winds. This can be problematic given evaporative coolers are most effective in arid climates where water scarcity is commonly a problem.

According to researchers, about 70-80 percent of the water that enters a cooling tower is actually [8]consumed [PDF], the rest is used to flush out mineral deposits similar to those found when cleaning a humidifier. The brine that's left behind is recycled through the system until it exceeds a certain concentration, at which point it's flushed away to a holding pond or treatment plant run onsite or by the local municipality before it's returned to the local watershed.

For this to work, the wastewater treatment plant needs to be sized correctly to handle the volume and concentration of brine generated by datacenters in the region. Things can get complicated pretty quickly when this isn't done, as was the case for Microsoft's [9]campus in Goodyear, Arizona.

Why datacenters' drinking habit is so hard to quit

One of the reasons that datacenter operators have gravitated toward evaporative coolers is because they're so cheap to operate compared to alternative technologies.

"It is always of a higher coefficient of performance (COP), meaning less energy required, to evaporate water, regardless of what cooling medium is being utilized," Shelnutt said.

In fact, COP, which refers to the amount of heat removed for a given amount of power, for evaporative cooling comes in at 1,230 while dry coolers and chillers manage a COP of about 12 and 4, respectively, he explained.

In terms of energy consumption, this makes an evaporatively cooled datacenter far more energy efficient than one that doesn't consume water, and that translates to a lower operating cost.

The challenge is that not every location and climate is well suited to evaporative cooling. In hotter climates where water is either scarce or places with high humidity where evaporative coolers are ineffective, chillers, which function similar to your AC unit, may be used instead.

In cooler climates such as the Nordic regions, datacenters often make use of free cooling and dry coolers, which take advantage of the lower ambient air temperature to eject heat into the atmosphere without consuming any water.

Whether or not evaporating cooling is used is highly dependent on location and climate, Digital Realty CTO Chris Sharp told The Register .

"You have to understand water is a scarce resource. Everybody has to start at that base point," he explained. "You have to be good stewards of that resource just to ensure that you're utilizing it effectively."

The colocation giant operates more than 300 bit barns around the globe, and uses a variety of designs based on predicted capacity requirements and environmental factors. The company's standard datacenter design, Sharp says, doesn't consume any water at all, instead relying on chillers to pull energy from the facility. However, in some locations, evaporative cooling and dry coolers are employed instead.

Most datacenter water isn't consumed onsite

While dry coolers and chillers may not consume water onsite, they aren't without compromise. These technologies consume substantially more power from the local grid and potentially result in higher indirect water consumption.

According to the US Energy Information Administration, the US [10]sources roughly 89 percent of its power from natural gas, nuclear, and coal plants. Many of these plants employ steam turbines to generate power, which consumes a lot of water in the process.

Ironically, while evaporative coolers are why datacenters consume so much water onsite, the same technology is commonly employed to reduce the amount of water lost to steam. Even still the amount of water consumed through energy generation far exceeds that of modern datacenters.

A 2016 [11]study [PDF] by Lawrence Berkeley National Lab (LBL) found that roughly 83 percent of water consumption attributable to datacenters could be attributed to power generation. As a result, reducing onsite water consumption at the expense of higher power draw could lead to an increase in the amount of water consumed.

However, just because power plants may pull more water than datacenters, that doesn't mean they're pulling the same water, Shaolei Ren, associate professor of electrical and computer engineering at UC Riverside, told The Register , adding that many power plants get their water from sources like rivers and lakes that may not be suitable for datacenters.

Ren and his team have been studying the datacenter's environmental impact on water consumption and [12]air quality .

This, again, is highly dependent on location and the grid mix. For example, datacenters located in regions with an abundance of hydroelectric, solar, or wind power will have lower indirect water consumption than one powered by fossil fuels or combustion.

What can be done to curb datacenter water consumption?

Understanding that datacenters are, with few exceptions, always going to use some amount of water, there are still plenty of ways operators are looking to reduce direct and indirect consumption.

One of the most obvious is matching water flow rates to facility load and utilizing free cooling wherever possible. Using a combination of sensors and software automation to monitor pumps and filters at facilities utilizing evaporative cooling, Sharp says Digital Realty has observed a 15 percent reduction in overall water usage.

"That equates to about 126 million gallons of avoided withdrawal from the system because we're just running it more efficiently," he said.

[13]Just how deep is Nvidia's CUDA moat really?

[14]Google thinks the grid can't support AI, so it's spending on solar for future datacenters

[15]Day after nuclear power vow, Meta announces largest-ever datacenter powered by fossil fuels

[16]Cloudy with a chance of GPU bills: AI's energy appetite has CIOs sweating

We're also seeing datacenters built in colder climates that can take advantage of free cooling most of the year. Better yet, in many Nordic countries, large quantities of hydroelectric power mean that even if auxiliary dry coolers or chillers are required, indirect water consumption isn't as much of an issue.

We've also seen heat generated by datacenters used to warm local offices, support district heating grids, or even greenhouses to grow produce year round.

In locations where free cooling and heat reuse aren't practical, shifting to direct-to-chip and immersion liquid cooling (DLC) for AI clusters, which, by the way, is a closed loop that doesn't really consume water, can facilitate the use of dry coolers. While dry coolers are still more energy-intensive than evaporative coolers, the substantially lower and therefore better power use effectiveness (PUE) of liquid cooling could make up the difference.

If you're not familiar, PUE describes how much power consumed by datacenters goes toward compute, storage, or networking equipment – stuff that makes money – versus things like facility cooling, which don't. The closer the PUE is to 1.0, the more efficient the facility.

This is possible because a sizable chunk, upward of 20 percent, of the energy used by air-cooled AI systems goes to chassis fans. On top of that, water is a much better conductor of heat. Shifting to DLC, something that's already happening with Nvidia's top-specced Blackwell parts, has the potential to drop PUE from [17]1.69-1.44 to around 1.1 or lower.

However, as Shelnutt noted in his SC24 presentation, this balancing act depends heavily on the power saved by DLC not being reallocated to support additional compute.

Water-aware computing

While many of these water-saving technologies require changes to facility infrastructure to implement, another approach might be to change the way workloads are distributed across datacenters.

The idea here isn't that different from carbon-aware computing, where workloads are routed to different locations based on the time and carbon-intensity of the grid, Ren explained. "They can do something similar based on the water stress level and real-time water efficiency, because this water evaporation rate does change over time - an hourly noon time versus the night time."

This, he admits, isn't something that the cloud providers and hyperscalers will have an easier time achieving as they maintain a tight grip on the orchestration of their infrastructure. "Colocation providers have more challenges due to limited control over the servers and workloads."

This approach may also not be appropriate for latency-sensitive workloads, like AI inferencing, where proximity to users is imperative for real-time data processing. However, workloads like AI training don't have these same limitations. One can imagine an AI training workload, which might run for weeks or months, could be queued up to run in a far-flung datacenter located in the polar regions that can take advantage of free cooling.

Fine-tuning workloads, which involve changing the behavior of a pre-trained model, are far less computationally intensive. Depending on the size of the base model and the dataset used, a fine-tuning job may only require a few hours to complete. In this case, the job could be scheduled to run at night when temperatures are lower and less water is lost to evaporation.

Is water the new oil?

While datacenter water consumption remains a topic of concern, particularly in drought-prone areas, Shelnutt argues the bigger issue is where the water used by these facilities is coming from.

"Planet Earth has no shortage of water. What planet Earth has a shortage of, in some cases, is regional drinkable water, and there is a water distribution scarcity issue in certain parts of the world," he said.

To address these concerns, Shelnutt suggests datacenter operators should be investing in desalination plants, water distribution networks, on-premises wastewater treatment facilities, and non-potable storage to support broader adoption of evaporative coolers.

While the idea of first desalinating and then shipping water by pipeline or train might sound cost-prohibitive, many hyperscalers have already committed hundreds of millions of dollars to securing onsite nuclear power over the next few years. As such, investing in water desalination and transportation may not be so far fetched.

More importantly, Shelnutt claims that desalinating and shipping water from the coasts is still more efficient than using dry coolers or refrigerant-based cooling tech.

"Desalinate ocean water right now at three kilowatts per cubic meter – that's an average over the last ten years; there are many installations of desalination plants that are down below one kilowatt hour per cubic meter – that's a COP of 222," he said.

Ship that 1,000 miles by pipeline and Shelnutt says the COP drops to 132. Shipped by train, the COP falls yet further to 38, far less than evaporating water sourced from a municipal treatment plant, but still far more efficient than using dry coolers. ®

Get our [18]Tech Resources



[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z3m9jUx1tDYrMVKhYc4LIgAAAQo&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0

[2] https://www.theregister.com/2022/12/19/google_datacenters_dalles/

[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z3m9jUx1tDYrMVKhYc4LIgAAAQo&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z3m9jUx1tDYrMVKhYc4LIgAAAQo&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0

[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z3m9jUx1tDYrMVKhYc4LIgAAAQo&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[6] https://arxiv.org/pdf/2304.03271

[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z3m9jUx1tDYrMVKhYc4LIgAAAQo&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0

[8] https://arxiv.org/pdf/2304.03271

[9] https://www.theregister.com/2023/04/11/microsoft_evaporative_coolers_arizona/

[10] https://www.eia.gov/tools/faqs/faq.php?id=427&t=3

[11] https://eta-publications.lbl.gov/sites/default/files/lbnl-1005775_v2.pdf

[12] https://arxiv.org/abs/2412.06288

[13] https://www.theregister.com/2024/12/17/nvidia_cuda_moat/

[14] https://www.theregister.com/2024/12/12/google_solar_energy_datacenter/

[15] https://www.theregister.com/2024/12/05/meta_largestever_datacenter/

[16] https://www.theregister.com/2024/11/29/public_cloud_ai_alternatives/

[17] https://journal.uptimeinstitute.com/large-data-centers-are-mostly-more-efficient-analysis-confirms/

[18] https://whitepapers.theregister.com/