The Next Hurdle for AI Data Centers: Cooling, Water, and Workers
```
Author: Bao Yilong
Source: Hard AI
The expansion of AI computing power is pushing the underlying infrastructure conflicts of data centers to a new critical point. Beyond electricity, cooling systems, water consumption, and labor shortages are becoming key constraints on the next wave of data center construction.
According to Hard AI, on May 13, the Barclays William Thompson team’s research report pointed out: As server rack power density surges, the importance of cooling systems is now on par with power supply.
NVIDIA GPU rack power density has leaped from 10-25kW/rack in 2020 to 120-150kW/rack with the Blackwell architecture, and is expected to exceed 600kW/rack with the Rubin Ultra architecture after 2027. Cooling systems are now one of the core variables restricting the deployment pace of AI data centers.
(Corresponding server rack power density for different NVIDIA GPU architectures)
Cooling systems account for about 10-15% of capital expenditure for data center building envelopes, and this proportion is still rising. Meanwhile, water availability and labor shortages are jointly reshaping the opportunity and risk landscape for the entire supply chain.
Cooling: The Core Infrastructure of Data Centers
Essentially, data center servers are "electric boilers"—almost all the electrical energy flowing through IT equipment is eventually converted to waste heat. Cooling failure means performance degradation, equipment damage, and even business interruptions.
As AI workloads come to dominate data center operations, the share of cooling systems in capital expenditures is increasing.
Barclays estimates cooling makes up about 10% to 15% of data center infrastructure construction costs, and expects this proportion to keep rising as liquid cooling architectures become more widespread.
The core work of cooling systems is in two steps: First, extracting heat from chips and servers; second, releasing this heat into the external environment.
(Schematic: liquid cooling + water-cooled chiller + wet cooling tower system)
The former can be realized via air, liquid, or phase-change refrigerants; the latter includes direct exhaust, air-cooled condensation, or evaporative cooling, among others. Evaporative cooling is the main source of direct water consumption in data centers.
A common investor misconception is equating liquid cooling with high water usage in data centers.
Barclays points out that liquid cooling systems usually run in closed loops—once coolant is injected, it continually circulates and, during normal operations, produces no net consumption.
In fact, liquid cooling can, by raising the supply temperature of coolant, often reduce or even eliminate dependence on evaporative cooling, thereby reducing direct water usage.
Evaporative cooling is the real main cause of direct water use: water absorbs heat as it evaporates and is lost as vapor into the atmosphere, forming a continual demand on water resources.
Data Center Water Footprint Far Exceeds Direct Cooling Water Use
Water consumption in data centers is more complex than commonly understood.
According to the US Department of Energy, direct water consumption for data center cooling in the US rose from about 21 billion liters in 2014 to about 66 billion liters in 2023, an average annual growth rate of 12-13%.
If data center capacity continues to expand, the Department predicts this figure could rise to 145-275 billion liters per year by 2028.
However, direct cooling water use is just the tip of the iceberg. A much larger volume comes from the power generation needed to run data centers—thermal power plants (gas, coal, nuclear) all require large amounts of cooling water.
Take the 176 TWh electricity demand of US data centers in 2023; the corresponding indirect water consumption is close to 800 billion liters. The International Energy Agency (IEA) estimates that the world’s data centers, in combination with cooling, power generation, and the semiconductor supply chain, withdraw about 5.2 trillion liters of water per year.
(IEA projections of data center water use under baseline scenario)
Meta’s disclosures are illustrative: in 2024, Meta’s purchased power embedded more than 7.2 billion liters of indirect water use, while direct onsite water withdrawal was only 560 million liters—a nearly 13-fold difference.
This proportion shows that if data center operators only focus on onsite Water Usage Efficiency (WUE) metrics, they’ll drastically underestimate their actual water footprint.
Water intensity varies significantly by power generation technology, but as the grid decarbonizes, indirect water use intensity on the generation side is expected to decrease progressively.
Traditional coal plants use about 70 liters per kWh, some old nuclear plants’ once-through cooling use over 100 liters per kWh; modern natural gas combined cycle plants use only about 10-12 liters; wind and solar PV need almost no cooling water at all.
Hyperscale Cloud Providers’ Cooling Strategies Are Diverging
The four largest hyperscalers are showing significant splits in their cooling strategies, with direct impacts on their choice of equipment suppliers.
(Comparison of water efficiency metrics among supercloud companies)
Microsoft has adopted a zero-evaporation cooling data center design centered on air-cooled chillers, aiming for global water “net positive” operations by 2030, and a 40% improvement in WUE from the 2022 baseline.
The company’s water withdrawal in FY2024 dropped around 20% year-on-year to about 1.04 billion liters; WUE was 0.27 L/kWh, an 18% improvement from 2022.
Meta follows a similar path. The latest generation of Meta AI data centers use direct liquid cooling plus dry coolers, with almost no water consumed in normal cooling operations.
In 2024, even as power demand rose about 21%, water withdrawal only rose about 7% to 560 million liters; WUE was 0.19 L/kWh, and its Beaver Dam facility in Wisconsin uses less water per year than two restaurants combined.
Google’s strategy is noticeably different. Based on Google’s public electricity and water data, Barclays estimates Google’s effective WUE exceeds ~1.3 L/kWh, much higher than its peers.
In 2024, its data center water withdrawal was about 4.2 billion liters, up around 27% year-on-year, completely matching growth in power demand.
Google maintains evaporative cooling in regions with low water-resource risk, reasoning that this cuts power use as well as related indirect water/cabon emissions.
However, in high water stress areas, Google has abandoned evaporative cooling; data centers in Mesa, Arizona; Canelones, Uruguay; and Waltham Cross, UK use air-cooled designs.
Amazon saw global WUE improve 17% year-on-year in 2024 to 0.15 L/kWh, and AWS says evaporative cooling can cut energy use by 25-35% during summer peaks.
Amazon likewise uses a differentiated regional approach, avoiding water cooling in high water risk areas.
Five Major Investment Misconceptions Debunked
The report specifically corrects prevalent market misconceptions, which is crucial for assessing equipment vendors’ investment value:
Liquid cooling ≠ high water use: As mentioned, evaporative cooling is the real culprit; liquid cooling can actually save water.Immersion cooling isn’t the final word: Direct-to-chip (D2C) liquid cooling has become standard for hyperscale deployments; immersion cooling remains niche due to operational complexity, lack of OEM ecosystems, and regulatory pressures on two-phase systems (PFAS, etc.).Air cooling isn’t going away: Except for full-immersion setups, networking, memory, and storage still rely on air. Liquid cooled data centers are hybrid architectures; air cooling will remain in the picture long-term.
(Schematic of air-cooled computer room air conditioning system)Chiller demand isn’t disrupted: The market overinterpreted NVIDIA CEO Jensen Huang’s statement about “45°C supply temperature needing no chillers,” temporarily hitting HVAC vendor stocks.But in hot climates, seasonal peaks, and for redundancy, chillers are still indispensable. As operators shift from wet towers to air systems, demand for air-cooled chillers should rise further.
(Schematic of computer room air handler + C-type water chiller + wet cooling tower system)LG Electronics forecasts data center air and liquid chiller market size will grow from $1.6 billion in 2026 to $12.7 billion in 2030.Waste heat utilization is hard to scale: Due to temperature mismatches, complicated municipal coordination, and high capital investment, reusing waste heat is mainly viable in Nordic district heating grids, and will not become a mainstream cooling design consideration.
Labor Shortage: The Underestimated Risk for Timeline and Cost
Cooling system installation is highly reliant on onsite skilled workers, making it the most labor-intensive subsystem of data center construction.
Installing liquid cooling loops or chiller units involves complex pumps, valves, heat exchangers, and custom piping networks, requiring plumbers, pipefitters, HVAC techs, welders, electrical engineers, and commissioning engineers to coordinate.
This is very different from the standardized, plug-and-play approach for prefabricated power modules or server racks.
Crusoe, developer of the Abilene Stargate project in Texas, USA, revealed (via CEO Chase Lochmiller at Stanford) that data center development labor costs are about $4.7 million per MW, or about 25% of total costs including captive power plants.
(Disclosed cost breakdown for Crusoe data center construction)
He noted that the Stargate project had an average of about 9,000 workers on site per day, and the Claude project elsewhere in Texas had 3,500—over twice the population of its town. Because these projects are remote, Crusoe had to recruit from other areas and provide large retention incentives.
Labor shortages bring multiple risks and opportunities:
First is schedule risk: insufficient labor may delay cooling installation, commissioning, and other key milestones, thus postponing data center operation;Second is cost inflation: intense competition for trades is raising both wage levels and subcontractor quotes;Third is technology tradeoffs: some developers are factoring construction difficulty into their choice of cooling, and more likely to select simpler air designs if labor is tight;Fourth is an opportunity for modular solutions: Prefab cooling units can reduce skilled labor needs onsite, helping compress timelines and reduce execution risks.
This article is from the WeChat public account "Hard AI". For more cutting-edge AI information, visit here
Risk Disclosure and DisclaimerThe market carries risks, and investments should be made cautiously. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions herein are appropriate to their particular circumstances. All investment decisions are at your own risk. ```