Thirsty AI: How Data Centers Create New Opportunities in Water Innovation

Aishwarya Mohan: Welcome to the webinar, Thirsty AI: How Data Centers Create New Opportunities in Water Innovation. My name is Aishwarya Mohan, Research Associate here at Lux Research, and I will be moderating today’s session. Presenting today are my colleagues, Abhirabh Basu, Senior Analyst, and Akshay Chaudhari, Analyst here at Lux Research. Throughout the webinar, you can type any questions you have in the questions box on your screen. Time permitting, we will answer all the questions that we can. If your question does not get answered, please do not hesitate to email it to [email protected], and we will respond. If at any point you experience technical difficulties, such as a frozen screen, simply refresh your browser and check that your internet connection is strong. Before we start the webinar, a word about who we are. Lux Research helps organizations make more confident innovation decisions about what matters next. As an independent research and advisory firm, we work with many of the world’s leading companies to identify emerging opportunities, evaluate technologies, and make high-impact innovation decisions. Our scientists, engineers, analysts, and industry experts combine original research and methodologies with practical decision frameworks to help organizations separate signals from noise. Every day, we help clients answer three critical questions: Where should we focus innovation efforts? Which technologies deserve investment? And which partners can accelerate innovation? Today’s webinar is an opportunity to share some of our thinking with you. Now, let’s jump into the discussion. Over to you, Abhirabh and Akshay.

Abhirabh Basu: Thank you, Aishwarya. Your AI service provider has a water problem. We’ve all seen the big headlines regarding the buildout of AI infrastructure recently. Data centers are increasingly being linked to drought, freshwater depletion, and pressure on local utilities, and this is becoming a growing community concern. The concern isn’t limited to one market like the U.S., where we’re already seeing states pause approvals and demand stronger proof of grid and water impacts. We’re seeing it globally, even where data centers have water recycling built into their operations or are sourcing wastewater for their operations. But these headlines tend to collapse several very different issues into a simple conclusion: that AI is using too much water. We think that’s an incomplete story. The real question isn’t simply whether AI is thirsty. It’s how you can prevent water from becoming a constraint on data center growth. And that’s what we’re here to unpack today. First, we look at whether AI really is thirsty and what the water problem with AI is. We look inside the data center to understand where water is used. Finally, we look at how the industry is responding—technologies and business models that are being used to extend resources, reuse more, and even find ways to eliminate water dependence altogether. Let’s start with the part that is unquestionably true here. Absolute water demand from data centers is rising. That’s 100% happening. On the left, you can see the combined on-site cooling and electricity-related water footprint increasing between 2025 and 2030 as AI infrastructure expands. But the chart on the right gives you a bit more perspective. Even at roughly 23 billion gallons of on-site water consumption in 2025, data centers are still a relatively small category of national water use here in the U.S. Compared with food production and agriculture, we lose more water to leaky pipes than is consumed on-site at data centers. For a global perspective, semiconductor manufacturing—the production of chips that go into data centers—draws about 200 billion gallons. That’s nine times more water than is consumed on-site. So even with rapid AI growth, on-site data center water consumption is not expected to rival something like agriculture by 2030. However, the challenge is the pace and geographic concentration of data center growth and, increasingly, as you’ll see, the much larger water footprint embedded in the electricity needed to power it. So that leads me to a really big question here: When does a relatively small national water footprint become a serious business constraint? And the answer is simple: when that demand becomes concentrated in the wrong place. Water is fundamentally local. It’s always been a location issue. A recent facility-level analysis from Stanford mapped roughly 9,500 data centers globally and found that a quarter of them are already experiencing seasonal water scarcity. Nearly half share water resources with municipal systems, often taking water from the municipal utility itself. However, unlike electricity, water is difficult to move long distances. You can procure power from somewhere else on the grid. You cannot simply import another watershed. So site selection, water rights, the capacity of the utility that you’re working with, and even community relationships become part of your infrastructure planning. And that changes the business problem completely. Now, a national average that looked very manageable becomes a local problem because of constrained water availability or a water-scarce watershed. To add another layer here, the water footprint of a data center is not just what’s consumed on-site. It includes the water associated with the electricity that’s used to power it. And that can be dramatically different given the regional power mix. So the same compute workload for your ChatGPT 5.6 or 6 can have very different water exposure depending on both the watershed it sits in and the electricity system that serves it. And that’s why national averages are useful for context, but they don’t tell you where the business risk actually rises. And I’m going to layer on another issue here, and I think it’s a growing issue: consumer concern and how communities are perceiving the rise in demand. Our consumer analysis shows that concern around data center water use is moving into the mainstream. It’s already in the mainstream, I would say. And those concerns are understandable. People are worried about their local groundwater, depletion of municipal supplies, being overbilled—especially here in Michigan, where I am—and whether communities should be carrying the burden of AI infrastructure growth. So I don’t want to dismiss those concerns. They are legitimate and increasingly consequential. But as we saw in the data on data center water demand versus other applications, it’s only part of the story. Public concern is not just noise around the edges; it can directly affect which projects are accepted, permitted, and given the license to build and grow. So data centers need a social license to build, but how they manage their water is increasingly going to determine their permission to expand. What can operators actually do about this? Well, from the get-go, you should be looking at how efficiently you’re using water. Or, if you’re servicing data centers, ask the question: Are we using water efficiently? Are we using the right source of water? Do we have access to groundwater? Do we need a potable supply? Are we using alternative sources here? Or should we look at alternatives to water dependence altogether? Can you redesign cooling systems so that external water becomes less important altogether? And to give you a little more flavor of that, I’m going to pass it on to my colleague, Akshay Chaudhari, who’s going to talk about the real water-use problem within the data center.

Akshay Chaudhari: Thank you, Abhirabh. To answer those questions, let’s go inside the data center and follow the heat. That will show us where water is actually being consumed and where the innovation opportunities are. So let’s start with the chips and the servers on the left. As they perform computations, they generate a lot of heat, and that heat has to leave the rack. There are three broad ways to collect that heat. Air cooling moves cool air through the servers. Direct-to-chip cooling circulates liquids through cold plates attached to the processors. And in immersion cooling, the servers are submerged in circulating dielectric fluids. The computer room air-handling unit and the coolant distribution unit then transfer that heat to the heat-rejection unit on the right. There, you have several options. Cooling towers reject that heat through evaporation. Adiabatic systems use water to cool incoming air. Dry coolers reject heat directly to the air without using water. And chillers provide mechanical refrigeration, but their water use, again, depends on how they reject that heat to the atmosphere. The choice primarily depends on climatic conditions. Looking at the left side gives us one important distinction: There is a technology cooling system that collects heat from IT equipment, and there is a facility cooling system that takes the heat and rejects it outside. The FCS ultimately determines how much water the data center is going to consume. Now let’s look at some examples. Let’s start with air cooling. There is no water flowing through the servers themselves. But the facility’s effectiveness here is measured per kilowatt-hour of electricity. The actual number will vary by climate and, of course, operating conditions. An air-cooled server does not necessarily mean it is a water-free data center. Now switch that server cooling to direct-to-chip. Here, we use a water-based coolant that collects heat from the processor effectively. Capturing heat at the source reduces air-handling demand and allows the facility to operate at a higher temperature. This is particularly valuable in hotter environments, where dry cooling can be used for more hours and evaporative cooling can be used only during peak conditions, such as extremely hot days. Under the same climatic conditions, this can reduce water consumption. Compared with air cooling, direct-to-chip cooling can definitely achieve water savings. Here’s a different example where you still have a direct-to-chip cooling system on the left, but in colder environments, you can actually use dry chillers. That brings water consumption down even further. The key point is that liquid cooling improves heat transfer in the TCS, of course, but the climate and the final FCS architecture will determine how much water is used in those data centers. The industry is actively trying to optimize these systems. Some key developments are shaping the industry and have direct implications for water use in data centers. The first development, or market evolution, that we are seeing is the rapid adoption of direct-to-chip liquid cooling. Here, you can see NVIDIA’s Vera Rubin computer, which has a completely integrated direct-to-chip cooling system. For operators, the cooling system increasingly needs to be planned alongside the IT equipment. Here, the GPUs, CPUs, and even the memory units are liquid-cooled. So the TCS is becoming more efficient at capturing heat. This allows for a warmer FCS on the heat-rejection side. And we are already seeing some signs and developments on that front. NVIDIA’s latest design showcases coolant entering at around 45 degrees Celsius. A warmer loop gives the facility more opportunities to reject that heat directly to the outdoor air. That can reduce chiller operation, displace evaporative cooling, and directly reduce water use in this case. However, we still need further optimization of the entire architecture to ensure reliability, and we also need to plan for extremely hot days. As you push these systems to their limits, reliability becomes an important factor. That has led to a third change in the market, which is more commercial in nature: Fluid management is becoming an integrated service. Workteam’s acquisition of BurgeRite is one such example. With this acquisition, Workteam added flushing, purging, and filtration capabilities to its liquid cooling services. Increasingly, as coolant moves closer to the electronics, cooling performance and uptime become very critical to data center operations. Data center operators need confidence that the loop is clean and stays within specified limits during operation. That creates opportunities to combine fluid monitoring, maintenance, and services. Now, keeping all these developments in mind, let’s look at the three questions Abhirabh raised. Are we using water efficiently? For that, we need to focus on the facility cooling loop, where evaporative losses occur. Are we using the right source? If we are using alternative supplies, we need suitable treatment and assurance of fluid quality. And can we reduce dependence on external water? For that, we need to look at the complete cooling architecture, starting with chip design and ending with heat rejection. To explore these opportunities further, I will hand it back to Abhirabh.

Abhirabh Basu: Perfect. Thanks, Akshay. I think you bring up an important point. As we go into the next section, which looks at innovation and case studies, do the innovation opportunities depend on whether the constraint is actually water availability outside the technology cooling loop and outside the rack? Or is it more about fluid reliability inside that loop? Or is it a combination of those factors? So, looking at those same three questions, operators must decide whether using water more efficiently can extend their operations. If they’re using alternative sources, what are those sources? How can they substitute for potable water? And finally, when they’re looking at alternative cooling architectures, how can they eliminate water usage while ensuring that the cooling system has the uptime required for computing? So let’s look at some case studies. The first one I want to focus on involves extending operations: increasing cycles of concentration, reducing blowdown, improving filtration, and making each gallon of water work harder. Singapore is a good example because water efficiency is being treated as a prerequisite for continued data center growth. Cooling towers can account for the vast majority of on-site water usage. So generally, the least disruptive lever to pull here is to reduce blowdown. The Public Utilities Board is actually pushing facilities toward higher water-use efficiency per megawatt-hour, with guidance on improving cycles of concentration with potable water, getting them up to seven cycles, and with NEWater, which is already an expensive source of recycled water, getting them up to 10-plus cycles. But with every additional COC—cycle of concentration—what are you doing in that water loop? You’re concentrating salts. You’re increasing solids. You’re also increasing the chances of biological contamination. So once those COCs rise, the bottleneck moves away from basic water supply and toward fine-tuning chemistry control and filtration and implementing automated monitoring systems. Higher COCs are a relatively low-disruption way to reduce facility water demand. We think the winners will be integrated chemistry and water-monitoring providers with the service depth to support hyperscalers. On the flip side, where you’ve got smaller facilities, we are also likely to see specialized startups emerge, provided they can exceed the limits of conventional cycles of concentration. Hydroleap in Singapore is a good example. The company is using a nonchemical, electrochemical approach to achieve seven, 10, or 12 cycles of concentration in its pilots with data centers. That was outside the technology cooling loop, where we’re trying to extend water use. Let’s look within the technology cooling system itself. The problem changes completely. When you’re looking at direct-to-chip cooling, water moves from being a consumable utility to a closed-loop working fluid. And as cooling moves closer to the chip, fluid quality is directly tied to computer reliability. Typical direct-to-chip filtration today is around 5 microns. But smaller cold-plate channels and higher heat fluxes are pushing particulate control into much tighter regimes. At the same time, operators have to manage corrosion in the pipes, conductivity drift, biological growth, and leaks as well. So the operating model for the facility cooling side versus the technology cooling side is going to be different. Instead of periodic maintenance, you’re looking at continuous fluid-condition management. I like the example here of Ecolab and CoolIT. Ecolab acquired CoolIT, and you can see where the market is heading for a service provider or chemical company that serves the water industry. It’s building an integrated solution: cooling hardware on one side and a legacy of water chemistry monitoring and service depth on the other. The winners are unlikely to be typical single-product vendors developing a new filter, sensor, or chemistry. They will be providers that can validate and assure fluid reliability across the full loop and tie that performance directly back to uptime. The second pathway operators should be looking at is substitution. We talked a little bit about reclaimed wastewater or impaired sources of water. Substitution technically sounds pretty straightforward. We’ve been treating wastewater to higher quality levels, including drinking-water quality, so we know the technology exists. We’ve been using it for cooling for decades. But this case study of OpenAI and XDC, their project in Western Sydney, shows the real constraint when you’re dealing with an alternative source of water. The planned project proposed using treated sewage for cooling to reduce potable-water dependence. The problem here isn’t the technology required for treatment. It is the infrastructure needed to get that reclaimed water to the site, specifically the pipeline and associated permitting. So substitution shifts the problem from water availability to infrastructure readiness. You need treatment capacity. You need to have distribution in place. You need to coordinate with the utility itself. And permitting, as we saw earlier, is becoming more and more of a challenge. So reclaimed water is most compelling when infrastructure is already secured or when you’re prioritizing and partnering with providers that can guarantee that supply. Because the challenge here, again, is not the treatment technology. It’s whether you can actually get the water to the site. I want to talk about a broader implication here for service providers, EPCs, and companies operating in this space. As AI demand grows, reclaimed wastewater itself will become a more valuable source. In many markets, treating municipal wastewater is still materially cheaper than, for example, investing in a desalination plant in Sydney and piping the water out to the data center. So reclaimed water is not an unlimited fallback option. It is increasingly a premium local resource. As a data center or a service provider to a data center, you’re going to be competing with other industries, municipalities, and agriculture. For the other example of substitution, I’m going to hand it back to Akshay, who’s going to talk about dielectric fluids.

Akshay Chaudhari: Yeah, so you saw that Abhirabh’s example changes the water source at the facility. This example actually changes the fluid close to the chips in the TCS. Darkconnex selected Accelsius’ NeuCool system for its planned 300-megawatt AI data center campus in Canada. The system uses two-phase direct-to-chip cooling. A nonconductive refrigerant absorbs heat by boiling at the cold plate. Because it uses latent heat for cooling, the fluid absorbs substantial heat during that phase change. Its dielectric properties also reduce the consequences of any leakage near expensive IT hardware. With efficient heat transfer, it also allows warmer cooling loops in the FCS. The agreement between Accelsius and DarkNX is a meaningful commercial signal that two-phase liquid cooling is gaining traction, but the technology, of course, needs further optimization. This does not directly remove or eliminate water use in the FCS, but it does offer industrial companies an opportunity to develop innovative solutions. For fluid developers, there is an opportunity to develop refrigerants with low global warming potential that are PFAS-free. For equipment providers, there are opportunities to innovate in filtration and maintenance to ensure reliable operations. And, of course, a key opportunity for system integrators is determining how to integrate this new generation of technologies with the existing FCS architecture. Now, we talked about extending the cycles and substituting the water source. Next, let’s investigate the final strategy: avoiding and eliminating the use of water in these systems. Dry heat rejection is central to this idea. So let’s look at a few examples. First, here we have an example from Microsoft and Corintis. They are bringing the coolant closer to where the heat is generated. Instead of making heat travel through layers of materials between the silicon and the cold plate, they etch tiny channels into the back of the silicon. Liquid flows through these channels and targets hotspots on the chips. Because the coolant comes into direct contact with the chip, Microsoft reports up to three times better heat-removal performance compared with conventional cold plates. Of course, it depends on the workload and the system design, but the key benefit for water is that removing those thermal barriers allows effective cooling with warmer liquid. You can run the chips at higher temperatures, and that could make dry heat rejection practical under a wider range of conditions. Again, the facility design then has to change to deliver that benefit in an integrated manner. The next challenge is, of course, making this reliable and manufacturable. Particularly with etched silicon, there are challenges involving the etching process and integration. There are opportunities in packaging and sealing, and the fluid channels on the chips need to be optimized even further. So there is an opportunity here for chipmakers and packaging suppliers. This is a development opportunity for them to monitor and validate over the next couple of years. There is also an opportunity for fluid suppliers to develop low-viscosity dielectric fluids that can flow easily through these microfluidic channels and achieve maximum heat transfer. Again, we talked about eliminating water, and we mostly discussed how that can be done in the TCS and FCS. But I want to draw your attention to the source of heat generation itself. Could we reduce the heat generated by the computation itself? That is the direction Vaire Computing is exploring with its adiabatic reversible computing. In simple terms, it aims to recover some of the energy used in transistor switching and reuse it. The company has reported energy recovery in experimental silicon circuits. That is an early technical milestone. It has a long way to go to reach the scale and computational performance of the current GPUs and CPUs we have. But for R&D teams, this is a long-term architectural opportunity. It is important to watch processor-level performance and total system energy and then perhaps adapt a simpler cooling system to reject the heat. It takes the story one step further: from managing water, to designing cooling that avoids evaporation, to reducing the heat that needs cooling. And I will hand it back to Abhirabh now to summarize the talk.

Abhirabh Basu: Perfect. Thanks, Akshay. So I’m going to wrap up today’s webinar with a couple of key takeaways. We’ve seen that water is extremely local. It always has been. And that means the same technology can have a very different business risk depending on the site or location where you’re operating. Take the example of OpenAI and XDC in Sydney, where reclaimed water was technically an attractive option, but you need infrastructure. You need permitting in place. You need local approval. And that becomes a constraint. So, in practice, water access and community acceptance can determine whether a project gets built and whether it grows, essentially requiring that social license to build. What does that mean for technology providers, R&D teams, and VCs? In the near term, I think the market—especially on the water-treatment side—is less about discovering a completely new water-treatment technology, particularly for the facility cooling-loop side. It’s more about deploying proven solutions where there’s less operational risk. Achieving higher cycles of concentration is a really good example of that. There’s an opportunity to work with cooling-system providers to get the chemistry right, lower chemical usage, improve filtration capabilities, and put an automated monitoring system in place so the operator can trust the operations at scale. For investors, especially the CVC and VC teams I’m speaking to right now, that means you want to be cautious about standalone water-technology bets. You’re going to be approached by a lot of companies. Everybody is now targeting data center water treatment as an opportunity. It is a massive opportunity, but you want to be able to validate technologies against new failure modes. Hyperscalers and data centers are unlikely to choose these very early-stage, low-TRL solutions. And finally, when we’re talking about integrated architectures and where you’re looking, Akshay has done a ton of work on thermal-management materials and liquid-cooling systems. When we’re looking at those opportunities, the value shifts from individual components to providers that can assure whole-system uptime. Companies need to integrate those new cooling approaches with monitoring that will ensure reliability and provide performance guarantees. And for R&D teams, especially on the chemicals and materials sides, that means looking further ahead at how heat is being generated at the chip or at how different bio-inspired channels, et cetera, can lower water-quality requirements or require entirely different approaches to water use. All right. Thank you. And I’m going to hand it back to Aishwarya for any questions from the audience.

Aishwarya Mohan: Thank you very much, Akshay and Abhirabh. We will now take questions you may have about the presentation, which you can type into the questions box. If we do not get to your question on this call, someone from Lux will be in touch after the webinar. I can see that we’ve already received several questions, so let’s get started. Our first question for Abhirabh is: Who is likely to capture the most value as this market evolves?

Abhirabh Basu: Yeah, that’s a good question. And we get asked about that quite often, especially on the water-treatment side. It’s an established technology space, right? Commercial maturity is pretty high there. I think when we’re looking at data centers, water usage, and optimizing water use, the best-positioned companies are water-treatment incumbents working with cooling-architecture companies, the OEMs, and likely the companies building these huge facilities—the EPCs and service companies. I think it’s these incumbents that can combine the hardware and produce the water at the quality required by the facility. That means combining the hardware with chemistry monitoring and ensuring commissioning and accountability for uptime. I would say those integrators are better positioned—the Ecolabs of the world that are acquiring companies in this space to provide services. You also see large companies such as Veolia, Suez, Watertech, Evoqua, and Xylem positioned there. But you’re also seeing midsized companies, such as Gradiant and Aquatech, that have built suites of technologies used in power generation and other industries and are now positioning them for this market. And they’re winning projects as well. For standalone technology developers, the most effective route is to go through these integrators rather than approach a hyperscaler directly to try out a new technology. We get asked this by a lot of materials developers working on new membranes, filtration capabilities, and sensors. They should focus on qualifying those technologies for facility cooling loops and technology cooling platforms, where the integrator can absorb the validation risk and guarantee performance.

Aishwarya Mohan: Thank you, Abhirabh. We have time for one final question, and this is for Akshay: Where do chemicals and materials companies have a differentiated role?

Akshay Chaudhari: Very interesting question. We have been getting it a lot from our clients. There are a couple of ways you can look at it, especially as direct-to-chip cooling gains traction and cooling loops become warmer. There is a thermal budget on the TCS side where the cold plate comes into contact with the processor. You have a limited thermal budget because the chip temperature is essentially T-max and the coolant is now at a higher temperature. So as the thermal budget goes down, you have to transfer heat from the chip to the cold plate as efficiently as possible. For that, you need better thermal interface materials. They need good thermal conductivity. You also need mechanical compliance to ensure that they are always in contact and that there is no thermal resistance at the interfaces. We have seen a lot of companies, including startups, working on these applications and targeting this specific issue. However, it’s a very crowded market. A lot of companies are targeting this market, and I think it goes beyond just material performance. You also need to ensure reliability. You need to show that the material can sustain thousands of thermal cycles. Qualification is actually the most critical part, and it may take several months to two years for a specific system. Increasingly, as NVIDIA introduces its own system architectures and maintains an approved-vendor list, it is very hard for a new material supplier to enter the ecosystem unless it is on that list. There may be an opportunity on the coolant and fluid side, and we touched on it during the presentation. Especially when you talk about fluids flowing through these cooling channels—and now, as you use finer features and microfluidic channels on the cold plate or directly on the chips—you need to ensure that the fluids are clean. The required particulate filtration will also need to improve. So there may be opportunities for new membranes. There is also an opportunity involving fluids because most dielectric fluids have high viscosity. Developers can create low-viscosity fluids that match the performance of water, which is one of the best heat-transfer fluids we have. Again, even for fluids, fluid assurance is the most critical part. So just having a fluid is not enough. You need to qualify that fluid at the system level and ensure that it is compatible with all the materials present in the liquid-cooling system. Qualification again takes time, and I think materials and chemical companies targeting this space will have to plan for this as well.

Aishwarya Mohan: Thank you very much, Akshay. That was very informative. That concludes our webinar for today. The slide presentation and recording from this webinar will be sent to all attendees via email today. After leaving the webinar, you will be prompted to complete a survey about today’s presentation. We would appreciate any feedback you may have to help inform and improve future webinars. Take a moment to check out our upcoming webinars on our website. Thank you for joining us, and have a great day.

Abhirabh Basu: Thank you.