In Part 1 of this multi-part series, we gave an overview of why wastewater is proliferating in the age of AI and the growing implications of unchecked wastewater discharge. In Part 2, we’ll examine cooling in data center operations. Today, we will discuss conventional vs. AI-focused data center operations and different cooling techniques, plus the tradeoffs data centers will need to balance in order to maintain cooling efficiency.
Cooling in Data Center Operations: Conventional vs. AI-focused
Whether for AI or traditional cloud computing, data centers need robust physical compute, power, and water infrastructures. While the first two are straightforward—comprising CPUs, GPUs, networking equipment, and a continuous electricity supply—the water infrastructure is where significant differences emerge.
Unlike conventional data centers, AI-focused data centers have significantly higher power density and, therefore, higher cooling requirements. Water has a significantly higher heat absorption capacity than air and is favored over other options that remain costlier. This makes it so that, even with a variety of cooling solutions at a data center’s disposal, the higher use of power is leading to an increase in water consumption and wastewater discharge.
Data centers by their nature produce a significant amount of heat, requiring effective cooling to maintain optimal operating temperatures. Conventional data centers typically use air-based cooling systems, which are cost-effective and easy to implement but less efficient for high-density environments. AI data centers, however, necessitate more advanced cooling techniques due to their higher heat output. This has led to a rise in liquid-based and hybrid cooling systems.
Conventional Cooling vs. AI-focused Cooling Systems
- Air-Based Cooling: Circulates cool air through server racks. It’s cost-effective but less efficient for high-density environments and consumes more energy.
- Liquid-Based Cooling: Uses chilled water or other coolants directly over or immersing equipment. It’s highly efficient and space-saving but has higher upfront costs and maintenance challenges.
- Hybrid Cooling: Combines air and water cooling, enhancing efficiency and cooling capacity while being moderately more expensive than air-based systems.
- Advanced Techniques: Include geothermal and phase change cooling, which are highly efficient and sustainable but limited by geography and complexity.
Given the high power density needed for AI data centers (>20 kW/server rack) compared to conventional ones (<10kW/rack), we predict liquid-based and hybrid cooling systems will inevitably become more prevalent. Hybrid cooling, in particular, is favored for its balance of efficiency and cost, making it adaptable to existing data center designs. However, these have higher water requirements and, therefore, will lead to more wastewater discharge.
Remember: the tradeoff for improved power efficiency is higher water consumption. Water is consumed (or lost) irretrievably when evaporated, and the remaining water that isn’t lost is either:
- Recycled after filtration OR
- (Partially) used for blowing down cooling systems to prevent mineral and scale buildup, AND
- Eventually discharged into public waters
Water Consumption and Discharge Trends Due to AI Demand
The principle behind hybrid cooling involves direct heat exchange between cool water and hot air, significantly enhancing cooling capacity but consuming large amounts of water in the process. Water is lost irretrievably through evaporation, while the remaining is either recycled or discharged.
The exact figures on water usage are challenging to quantify due to a lack of detailed public data. However, estimates suggest data centers consume millions of gallons of water daily. KETOS’ work with data center partners estimates that about one-third of the water used in data centers is consumed, with the rest getting either recycled and eventually discharged.
With an absence of quantifiable information, it can be difficult to assess future water usage trends. However, we’ve attempted to begin the process of estimation using the Water Usage Effectiveness (WUE) metric. The WUE metric is the ratio of the annual total water utilization to the annual compute energy consumption (Information Communication Technology, or ICT, equipment energy consumption). This calculation can relate total water usage to energy consumption.
Some hyperscalers have published WUE data for various data centers by region, which helps us manage our internal calculations.
Water Usage Efficiency Metrics published by a Hyperscaler.
The WUE is a site specific metric because water usage varies tremendously by geography, power and water availability. However, a metric such as WUE helps us understand the tradeoff between power and water consumption on average. It also helps us derive water usage trend estimates (which are unavailable) from power usage trends (which are more widely studied and well known).
So, How Much Wastewater Should We Expect?
Under the assumption that:
- Hybrid cooling will continue to dominate cooling solutions for AI data centers.
- Roughly half of the water usage is lost (i.e. consumed).
A factor of roughly half applied to the total water consumption under various WUE scenarios gives us a reasonable estimate of the trends for wastewater discharge based solely on the growth in AI data centers.
According to our calculations, the industry should anticipate an additional 100 billion gallons of wastewater annually originating from the cooling demand of AI data centers (in an optimistic scenario of only WUE = 1.0).
AI data center power demand is expected to equal conventional data center power demand by 2028. Therefore, it’s safe to expect an additional 100 billion gallons of wastewater due to net data center cooling demand from hybrid cooling alone.
Peak Power Demand Projections Create Even More Wastewater
While 100 billion gallons (calculated at WUE = 1.0) seems astronomical, consider this:
A single 1 GW AI data center operating at a (conservative) WUE = 1.5 may release nearly 350 million gallons of wastewater annually by 2026. A hundred such data centers will release wastewater equivalent to about 5-6% of US annual total wastewater discharge. However, if we scale up power consumption in these data centers by two orders of magnitude at the same WUE (i.e. 100 GW clusters), only 30 of these data centers will release the equivalent of all wastewater released in the US annually.
Note: WUE = 1.5 is conservative considering the power-water tradeoff. Given the power usage projected into the future, the expected amount of carbon emissions from this usage and the efficiencies that can be gained from water-based cooling, it is not unreasonable to assume WUEs of 1.5+. Our analysis also discounts the impact of wastewater generation and water usage from power generation. The average WUE across generation sources is approximately 7.6. That means 7.6 liters (2.0 gallons) of water are consumed per kWh used by end consumers. An equivalent amount is reasonably assumed to be discharged back into public waters as wastewater.
Moving Forward
The growth of AI data centers presents a significant challenge for water usage and quality. It also poses this critical question:
Can our wastewater treatment plants handle double the volume of wastewater expected over the next few years from AI applications alone?
Both data center operators and utilities will need to work together to assess existing wastewater monitoring and treatment capacities in the area where data centers are planned to operate and to reflect on the existing public infrastructure capabilities and budgets to preemptively and proactively address the impending wastewater issue.
By understanding and mitigating the impacts of increased water consumption and wastewater discharge, we can ensure that the digital revolution driven by AI remains sustainable and environmentally responsible.
At KETOS, we are committed to providing innovative water quality monitoring and management solutions, helping safeguard our water resources in this new age of AI. Our SHIELD technology makes it possible to minimize a center’s wastewater contaminant footprint while maintaining operational efficiency. If you are a data center operator or a utility that works with a data center, we’d love to help you manage your wastewater discharge.
Want more insights into wastewater discharge trends in data centers? Download our whitepaper.