AI cooling has moved from an engineering footnote to a front-page infrastructure story. According to Rystad Energy estimates, data centers used 222 billion liters (59 billion gallons) of water worldwide for cooling in 2025, and that figure could reach 644 billion liters by 2030 without adaptive measures. Water Tech Online, which is tracking the shift, frames the question for operators as how to cut resource consumption without slowing the AI build-out.
The industry answer so far is a mix of closed-loop liquid cooling, warmer coolant temperatures, dry coolers, immersion fluids and software that tunes the whole system. The stakes are not abstract. A recent cooling failure at a George Washington University data center that hit nine named systems showed how quickly thermal problems turn into outages. This analysis walks through what vendors and analysts have actually reported, what the numbers imply, and where the approaches differ.
Everything below draws on named reports: Rystad Energy, NVIDIA, the Information Technology and Innovation Foundation (ITIF), Schneider Electric, Huawei and Dow. Where a detail is not confirmed by those sources, this article says so instead of guessing. For wider context on AI infrastructure, see our AI and machine learning coverage.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Numbers Behind AI Cooling’s Water Problem
Start with the baseline. Rystad Energy puts global data-center cooling water at 222 billion liters in 2025. That equals roughly 59 billion US gallons, enough to make the problem a line item in utility planning rather than a rounding error.
The forecast is the part that worries planners. Rystad’s no-action scenario reaches 644 billion liters by 2030. By simple arithmetic, that is about 2.9 times the 2025 level, or compound growth near 24% a year across five years. Those two derived figures are our math, not Rystad’s wording, but they follow directly from its published numbers.
The qualifier matters: “without adaptive measures.” Rystad’s projection is a warning about the default path, not a prediction that the path will be followed. Operators who switch designs can bend the curve, and several already claim to have done so. The rest of this article tests how credible those claims are.
| Metric | Figure | Source |
|---|---|---|
| Global data-center cooling water, 2025 | 222 billion liters (59 billion gallons) | Rystad Energy |
| Projected cooling water, 2030, no adaptive measures | 644 billion liters | Rystad Energy |
| Implied growth multiple, 2025 to 2030 | About 2.9x | Derived from Rystad figures |
| Peak-day water for a 1 GW evaporative-heavy site | 5 million to 10 million gallons | Schneider Electric analysis |
| Water-use efficiency gain, 2022 to 2025 | 25% at Microsoft, 37% at AWS | Company sustainability reports |
Why 1-Gigawatt Campuses Change the Math
Schneider Electric’s analysis gives the clearest picture of scale. A 1-gigawatt AI data center that leans heavily on evaporative or adiabatic cooling during peak conditions could need approximately 5 million to 10 million gallons of water per day, depending on design. That is a daily figure for peak conditions, so annual totals depend on climate and load.
If a site ran at the extremes of that range for a full year, the total would land between roughly 1.8 billion and 3.7 billion gallons. We stress the “if.” Real sites do not run at peak every day, so treat those bounds as a ceiling for thought experiments and not a forecast.
Gigawatt-class campuses are now a planning reality. The scale of spending behind them shows up in corporate budgets, as we covered when Meta raised its AI capital spending to $145B. Water and power are the two constraints that money alone cannot remove, which is why cooling design now gets board-level attention.
How NVIDIA DSX Closes the Loop
NVIDIA’s DSX reference designs target the water problem directly. The company describes closed-loop liquid cooling paired with outdoor dry coolers that reject heat without evaporating water. NVIDIA says facility cooling-water consumption can be reduced to near zero under this approach.
The operating detail is what makes it work. NVIDIA’s system circulates liquid directly through servers, close to the chips. Chip temperatures can exceed 80°C (176°F), and the liquid can enter servers at 45°C. Because the coolant arrives warm, outdoor air can still pull heat out of it for much of the year, so the facility does not need evaporation to finish the job.
What “near zero” does and does not mean
NVIDIA’s claim concerns facility cooling water. It does not cover water consumed upstream in electricity generation or in chip manufacturing. Readers comparing vendor claims should check which boundary each figure uses. A site-level number and a lifecycle number are different things, and the fact sheet we worked from only supports the site-level framing.
Rubin and the 113°F Threshold
ITIF describes NVIDIA’s Rubin-generation AI infrastructure in similar terms. The foundation says the platform can use liquid cooling with water consumption reaching zero, and that the cited design runs liquid loops at up to 113°F with external dry coolers.
The 113°F figure is the same as 45°C, which is the coolant entry temperature NVIDIA cites for DSX. That consistency is useful. It suggests the industry is converging on a warm-water target as the design point for next-generation accelerators, rather than the chilled loops common in older facilities.
Chip vendors are also pushing power and packaging hard. Memory stacks add to the heat load, a point worth keeping in mind alongside Micron’s HBM4 and 400-layer NAND push. Denser silicon makes liquid cooling less optional with each generation.
Schneider’s Temperature Math: Warmer Loops, Less Water
Schneider Electric quantified why warmer coolant matters. Raising the technology-cooling-system supply temperature from 32°C (89.6°F) to 45°C (113°F) cut the water-usage-effectiveness ratio by approximately 66% in Dallas and 82% in Paris, for the analyzed 100-megawatt facilities.
The gap between the two cities is the interesting part. Paris, with a cooler climate, saw the bigger drop. A warmer supply loop lets dry coolers handle more hours of the year, and cooler ambient air extends that advantage. Dallas still gained two-thirds of the benefit despite heat, which shows the lever works even in demanding climates.
| Site analyzed | Facility size | Supply temperature change | Water-usage-effectiveness reduction |
|---|---|---|---|
| Dallas | 100 MW | 32°C (89.6°F) to 45°C (113°F) | About 66% |
| Paris | 100 MW | 32°C (89.6°F) to 45°C (113°F) | About 82% |
These are modeled results for analyzed facilities, and Schneider’s own assumptions drive them. They show direction and rough magnitude. They do not guarantee identical outcomes at every site.
Microsoft’s Zero-Water Cooling Design
ITIF reports that Microsoft has introduced AI-optimized data centers designed to use zero water for cooling operations. The design relies on continuously recirculating, closed-loop direct-to-chip cooling. In practical terms, the same coolant circulates through cold plates on the processors, and the heat leaves the building through a sealed circuit.
This matches the NVIDIA architecture in spirit. Both rely on liquid touching the hot components, warm loop temperatures and dry heat rejection. The convergence between a leading chip vendor and a leading cloud operator tells us this is becoming the default pattern for new AI halls, not a niche experiment.
Note the scope of the claim. ITIF describes it as a design for cooling operations. The fact sheet does not tell us how many sites use it, how fast it will roll out, or what share of Microsoft’s fleet it covers, so we will not speculate on those points.
The Efficiency Paradox: Better Ratios, Bigger Totals
The latest sustainability reports from Microsoft and AWS show a pattern that frustrates simple narratives. Between 2022 and 2025, water-use efficiency improved by 25% at Microsoft and 37% at AWS. Yet both companies’ total water use increased over the same period.
The explanation is growth. Each unit of compute uses less water than before, but the number of units is climbing faster. Efficiency gains of 25% to 37% cannot offset a build-out that expands capacity by much more than that. This is the same dynamic that has played out with energy in earlier computing eras.
| Company | Water-use efficiency change, 2022 to 2025 | Total water use trend |
|---|---|---|
| Microsoft | Improved 25% | Increased |
| AWS | Improved 37% | Increased |
For policymakers, the lesson is that ratio metrics alone can mislead. A falling water-usage-effectiveness number looks like progress, and it is, but absolute consumption is what local watersheds feel. The zero-water designs from NVIDIA and Microsoft matter because they attack the absolute figure rather than only the ratio.
AI Managing Cooling: Huawei and the 21% Result
Hardware is only half the story. Software that tunes cooling in real time is the other half. Huawei says its AI-based cooling and power-optimization solution, called iCooling@AI, improved data-center power usage effectiveness by 8% to 15% at China Unicom Henan.
A separate report on an AI-optimized industrial cooling-control system found it reduced electrical energy consumption by 21% during its first three months of operation, compared with conventional control systems. That result comes from industrial cooling more broadly, not specifically from a hyperscale AI hall, so apply it with care.
The pattern is clear enough. Control software can trim energy use by double digits in some deployments, using the same chillers, pumps and fans already on site. These are vendor and report claims, and independent verification across many facilities would strengthen the case. Still, the ranges are consistent with the idea that cooling plants are rarely tuned to their best operating point.
Software-led optimization also fits the way operators already think about AI workloads. Teams that run inference at scale, such as the ones behind DeepSeek’s open-source DeepInfer framework, track every watt per token. Applying the same discipline to cooling is a natural next step.
Immersion Cooling and Dow’s Silicone Fluids
Immersion cooling takes the liquid idea further by submerging servers in a dielectric fluid. Dow says its silicone-based DOWSIL immersion-cooling fluids are intended to reduce power consumption, extend equipment lifespan and support energy-efficient data-center operations.
Dow’s statement describes intent. The fact sheet gives no measured savings figure, no named customer and no deployment count, so we cannot say how much power the fluids save in practice. What the announcement does confirm is that a major chemical supplier now sees immersion as a market worth serving, which signals growing demand for alternatives to air.
Immersion also changes the supply chain. Fluids, tanks and service procedures differ from direct-to-chip plumbing, and operators must weigh maintenance, material compatibility and vendor lock-in. Those questions are why many large operators have favored cold plates for mainstream racks while testing immersion in selected halls.
Trade-offs: Water Saved, Power Spent
Zero-water cooling is not free. Dry coolers reject heat to air, and that usually means fans, larger heat exchangers and more floor or roof space. Our read is that operators trade water for some combination of electricity, capital cost and footprint. Warm-loop designs soften the penalty because the temperature gap between coolant and outdoor air stays wide.
Climate shapes the balance. Schneider’s Dallas and Paris results show the benefit of a warm loop in two very different climates, yet the hottest days still test dry coolers. Some sites may keep a limited evaporative option for extreme peaks, which keeps total water far below the 5 million to 10 million gallons per day that heavy evaporative designs can need.
Reliability also enters the picture. Liquid cooling puts coolant lines near expensive electronics, which raises the cost of leaks and pump failures. Operators answer with redundant loops, leak detection and careful commissioning. That engineering effort is part of the real price of switching, and it rarely appears in headline claims.
Competitive Comparison of Cooling Approaches
The table below summarizes the approaches covered in this article, using only what the named sources report. Cells marked “not reported” reflect gaps in the sources, not zero values.
| Approach | Reported water outcome | Reported energy or efficiency outcome | Source |
|---|---|---|---|
| NVIDIA DSX closed loop with dry coolers | Near-zero facility cooling water | Coolant enters servers at 45°C | NVIDIA |
| NVIDIA Rubin liquid cooling | Water consumption reaching zero | Loops up to 113°F | ITIF |
| Microsoft direct-to-chip closed loop | Zero water for cooling operations | Not reported | ITIF |
| Warmer supply loop (32°C to 45°C) | WUE down about 66% (Dallas), 82% (Paris) | Not reported | Schneider Electric |
| Huawei iCooling@AI | Not reported | PUE improved 8% to 15% | Huawei |
| AI-optimized cooling controls | Not reported | Electricity use down 21% in first three months | Industrial control report |
| Dow DOWSIL immersion fluids | Not reported | Intended to cut power use | Dow |
The comparison shows two families of answer. Hardware changes such as closed loops and warmer coolant attack water. Software and fluid changes mostly attack energy. The strongest sites will likely combine both, and the vendor claims above are not mutually exclusive.
Historical Context: From Air Halls to Liquid Loops
For most of the history of enterprise computing, data centers relied on air. Chilled air moved through raised floors and aisles, and many operators added evaporative stages because evaporation is an efficient way to shed heat with little electricity. That choice saved power and spent water, a trade that seemed sensible when racks drew modest power.
AI accelerators broke that bargain. Chips that run above 80°C and sit in dense racks overwhelm air as a medium, which pushed vendors toward liquid at the chip. Once liquid is in the rack, the loop can run warm, and a warm loop frees the facility from evaporation. The water story and the density story turned out to be linked.
Public attention followed the physics. As cooling water became a headline number, as in Rystad’s 2025 estimate, vendors began marketing zero-water designs. The shift from an afterthought to a selling point took only a short time, and it shows how quickly infrastructure priorities move when local communities and regulators start asking questions. Smaller projects feel the same pressure, as seen in the Wiikwemkoong data centre project, where community context shapes design.
Market Impact: Who Gains From the Cooling Shift
The cooling transition redirects spending. Suppliers of cold plates, coolant distribution units, dry coolers, control software and specialty fluids stand to gain. Dow’s push into immersion fluids and Huawei’s software pitch are early examples of vendors positioning around the shift. Traditional evaporative-cooling equipment faces weaker demand in new AI builds that adopt the closed-loop pattern.
Chip and platform vendors gain leverage as well. When NVIDIA publishes reference designs, it shapes what facilities get built around its processors. Competing silicon efforts, including Google’s TorchTPU push against Nvidia, will need equally credible cooling stories for customers who now ask about water in procurement.
Cost is the open question. The fact sheet gives no price comparison between evaporative and closed-loop builds, so we cannot say which wins on total cost. What we can say is that water risk is now a siting factor. Regions with scarce supply or tight permits favor designs that need almost none, and that can outweigh a modest increase in capital cost.
Predictions: Where AI Cooling Goes Next
These are our analytical forecasts, based on the sources above. They are judgments and not reported facts.
- Warm-water loops become the baseline. With NVIDIA and ITIF both pointing to 45°C or 113°F coolant, new AI halls will increasingly specify warm loops by default, and chilled-water designs will shrink to legacy retrofits.
- Reporting shifts from ratios to absolute totals. Microsoft and AWS both improved efficiency while total use rose. Pressure from regulators and communities should push disclosure toward absolute liters and gallons.
- Cooling software gets bundled with hardware. Claims like Huawei’s 8% to 15% PUE gain and the 21% energy cut make controls an easy add-on, so expect vendors to package optimization with their cooling gear.
- Rystad’s 644 billion liter path proves too high. If zero-water designs spread in new builds, 2030 consumption should land below the no-action case, though existing evaporative sites will keep it well above zero.
- Power becomes the binding constraint. As water drops out of the equation, the electricity needed to run dry coolers, pumps and the chips themselves takes center stage in siting and permitting debates.
What Operators and Buyers Should Do Now
Start by asking what boundary each vendor claim uses. Facility cooling water, total site water and lifecycle water are different measures. A near-zero claim for the first says little about the other two.
Next, model your own climate. Schneider’s Dallas and Paris results differ by 16 percentage points, which shows location matters even when the design is the same. Run the numbers for your sites rather than relying on a headline percentage.
Finally, track both water and power. A design that cuts water but raises electricity may still be the right call, yet the decision should be explicit. Teams already weighing hardware costs, such as the memory squeeze behind PS6 RAM pricing fears, know that infrastructure budgets are zero-sum. Cooling choices compete with compute for the same dollars.
Frequently Asked Questions About AI Cooling
How much water do data centers use for cooling?
Rystad Energy estimates that data centers used 222 billion liters (59 billion gallons) of water worldwide for cooling in 2025. It projects 644 billion liters by 2030 if no adaptive measures are taken.
Can AI data centers really use zero water for cooling?
ITIF reports that Microsoft’s AI-optimized data centers are designed to use zero water for cooling operations, and that NVIDIA’s Rubin-generation infrastructure can reach zero water consumption with liquid cooling. NVIDIA describes its DSX designs as reducing facility cooling water to near zero. These claims cover cooling operations at the facility, not every upstream water use.
What is a closed-loop liquid cooling system?
It circulates coolant in a sealed circuit that passes near or over the chips, then sends the heat to outdoor dry coolers. Because the coolant is not evaporated, the facility does not consume water for cooling. NVIDIA says its system can send liquid into servers at 45°C while chips run above 80°C.
Why does a warmer coolant temperature save water?
A warmer loop lets outdoor air remove heat for more hours of the year, so evaporative help is needed less often. Schneider Electric found that raising supply temperature from 32°C to 45°C cut the water-usage-effectiveness ratio by about 66% in Dallas and 82% in Paris for analyzed 100-megawatt facilities.
Do efficiency gains mean total water use is falling?
Not necessarily. Microsoft’s water-use efficiency improved 25% and AWS’s improved 37% between 2022 and 2025, but both companies’ total water use increased, according to their latest sustainability reports.
Can AI software make cooling more efficient?
Reported results suggest so. Huawei says iCooling@AI improved power usage effectiveness by 8% to 15% at China Unicom Henan, and a separate AI-optimized industrial cooling-control system cut electricity use by 21% in its first three months versus conventional controls.
What is immersion cooling and who is selling it?
Immersion cooling submerges servers in a dielectric liquid. Dow offers silicone-based DOWSIL immersion-cooling fluids, which it says are intended to reduce power consumption, extend equipment lifespan and support energy-efficient operations. Dow has not provided measured savings in the material we reviewed.
How much water could a gigawatt AI campus need?
Schneider Electric’s analysis says a 1-gigawatt AI data center relying heavily on evaporative or adiabatic cooling during peak conditions could need approximately 5 million to 10 million gallons per day, depending on design. Closed-loop designs aim to cut that sharply.
