The water number in a sustainability report may tell only part of the story.

Artificial intelligence is being built at unprecedented scale. Microsoft, Google, Amazon, Meta and other technology companies are investing hundreds of billions of dollars in new computing infrastructure, with industry analysts and financial reporting pointing toward roughly $1 trillion in cumulative AI infrastructure spending across 2025 and 2026 alone. But there is another infrastructure requirement receiving considerably less attention: Water. As AI data centers become larger, denser and more power-intensive, the question is no longer simply how much water a facility uses to cool its servers. The bigger question may be:

How much water does the entire system require to operate that data center — including the water consumed to produce the electricity powering it?

That distinction could fundamentally change how we understand the resource footprint of AI.

The water you can see — and the water you can’t

When companies publish sustainability reports, data center water consumption is generally discussed in terms of direct water use. Think cooling towers, chillers and other systems located at the data center itself. This is an important measurement. But it doesn’t necessarily capture the water consumed elsewhere in the energy system that supplies electricity to the facility. Power plants can require significant quantities of water for cooling and other processes. The amount varies dramatically depending on the type of electricity generation, the technology used and the location. In other words, a data center can reduce its own on-site water consumption while still relying on an electricity supply whose production has a substantial water footprint. That isn’t necessarily deceptive. It is an accounting boundary. And accounting boundaries matter. A July 2026 Wall Street Journal investigation by technology columnist Christopher Mims highlighted this distinction, reporting that Microsoft, Google and Amazon generally disclose water used directly at their data centers, while Meta stands out as the major operator that also accounts for water used at the power stations supplying its data centers’ electricity. The result is that two companies can appear to have very different water footprints — or similar ones — depending entirely on what their reporting boundary includes.

The 12x problem

This is where the issue becomes particularly interesting. A 2024 report from Lawrence Berkeley National Laboratory estimated that, in 2023, the indirect water consumption associated with electricity generation for U.S. data centers was roughly 12 times their direct water consumption for cooling. That statistic requires an important clarification. It does not mean that every individual data center literally consumes twelve times more water than the figure appearing in its sustainability report — the ratio varies by cooling technology and by the water intensity of the local power grid. It means that, at the broader U.S. data-center level, estimated water consumption associated with electricity generation can be dramatically larger than the water consumed directly for on-site cooling. The distinction is critical. If a facility reports only its direct cooling water, the number can be fully accurate within that reporting boundary while still failing to describe the full resource footprint of the computing system it belongs to. That is the difference between measuring a facility and measuring a system.

AI changes the equation

AI makes this problem more important because AI computing is unusually resource intensive. Training and running increasingly sophisticated models requires enormous amounts of computing power. That computing power translates into electricity demand. Electricity demand translates into additional generation capacity. And depending on the generation mix, electricity production can carry its own substantial water requirements. At the same time, the computing equipment itself generates significant heat. That creates a complicated optimization problem: Use more water for cooling, or use more electricity to reduce direct water consumption? Neither number exists in isolation. A cooling technology that looks extremely water-efficient at the facility level may require more electricity. If that additional electricity comes from a water-intensive generation source, some of the water footprint has effectively moved upstream, not disappeared. Conversely, a facility using evaporative cooling may consume more water directly while requiring less electricity for cooling overall. The right question therefore isn’t simply: “How much water does this data center use?” It is: “What is the total water intensity of delivering the computing services this data center provides?” That is a much harder question. It is also a much more useful one.

Location changes everything

There is no universal “AI data center water footprint.” The answer depends on where the facility is built, how it is cooled, what electricity supplies it and what water sources are available locally. A data center operating in a water-abundant region with a low-water-intensity electricity mix presents a very different resource picture than an identical facility operating in a water-stressed region on a grid still leaning on thermoelectric generation. This becomes especially significant in places such as Arizona, where rapid data-center development intersects with long-standing concerns about water availability. The physical location of compute is therefore not merely a real-estate decision. It is a resource decision. And increasingly, it is an infrastructure-resilience decision.

Reporting water isn’t enough if the boundary keeps changing

There is another problem hiding inside the numbers: comparability. Research into technology-company disclosures has found substantial differences in how companies report water use. Some provide facility-level information. Others report broader corporate figures. Some disclose data-center-specific water consumption, while others publish water metrics that are difficult to isolate to data centers at all. That makes comparisons difficult. Imagine two companies: Company A reports 1 million gallons of direct data-center water consumption. Company B reports 2 million gallons — but that figure includes additional categories of indirect water consumption Company A doesn’t disclose at all. Which company actually uses more water? The answer may be impossible to determine from the headline numbers alone. This is why sustainability reporting needs more than a number. It needs context, methodology and provenance.

The real issue isn’t transparency versus secrecy

It is tempting to characterize this as a story about Big Tech hiding water consumption. The reality may be more nuanced. Companies are often reporting legitimate measurements according to established accounting methodologies. The problem is that those methodologies may not capture the entire system. That distinction matters because the solution isn’t simply to demand that companies “report more.” The solution is to establish consistent, auditable resource accounting across the industry. For AI infrastructure, that could mean reporting at least four layers:

1. Direct water consumption — water used inside the data center.
2. Indirect electricity-related water consumption — water associated with generating the electricity consumed.
3. Local water stress — whether the facility’s water source sits in a water-constrained region.
4. Resource intensity per unit of compute — how much water and energy are required to produce a defined amount of computing output.

That would provide a far more meaningful picture of AI’s physical infrastructure than any single headline number.

From sustainability reporting to infrastructure intelligence

This is where the conversation becomes bigger than water. AI infrastructure increasingly sits at the intersection of energy, water, land, compute, supply chains and local infrastructure. A decision to build a 1-gigawatt AI campus isn’t simply a decision about servers. It potentially affects electricity generation, transmission capacity, cooling systems, municipal water supplies, wastewater infrastructure, construction materials and surrounding communities. Those systems are interconnected. Yet the data used to make decisions about them is often fragmented across companies, utilities, municipalities, regulators and infrastructure operators. That creates a trust problem. If decision-makers can’t see the complete resource chain, they can’t accurately assess the consequences of infrastructure decisions. And if two organizations measure the same resource using different boundaries, their numbers may be technically correct while still producing completely different conclusions about the same facility.

AI needs a better resource ledger

The irony is that artificial intelligence may ultimately help solve this problem. AI systems can correlate infrastructure data across otherwise disconnected sources: energy consumption, cooling requirements, water availability, power-generation mix, weather, grid conditions, facility utilization and geographic constraints. But AI cannot make fragmented data trustworthy simply by analyzing it. The underlying data needs provenance. It needs to be possible to determine: Where did this number come from? What exactly does it measure? What does it exclude? When was it measured? Can another source corroborate it? And perhaps most importantly: Does the number describe the facility — or the system? That distinction is becoming increasingly important as AI infrastructure expands.

The next generation of data-center reporting

The future of AI infrastructure shouldn’t require choosing between economic growth and environmental responsibility. It should require better information. A modern data center should be evaluated not only by its computational capacity, uptime and power availability, but also by the total resource system required to sustain it. Water should be part of that equation. So should energy. So should geographic scarcity. And so should the uncertainty surrounding the data itself. The Wall Street Journal‘s reporting is therefore less a story about whether a particular technology company is “telling the truth,” and more a warning about how easily technically accurate numbers can add up to an incomplete picture. The AI infrastructure race is moving too quickly for resource accounting to remain fragmented. If we’re going to measure the future, we need to measure the whole system — not just the part that’s easiest to report.

The PulseDNA Perspective

At PulseDNA, we believe trustworthy infrastructure decisions require more than a number. They require context, provenance, corroboration and continuity across systems. AI infrastructure is becoming a physical system as much as a digital one. The organizations that can connect those layers — compute, energy, water, location, risk and evidence — will be better positioned to build AI infrastructure that is not only powerful, but resilient and accountable. Because the future of AI isn’t just about how much compute we can build. It’s about whether we can build it intelligently. Image is made by AI and for illustrative purposes only.

---

Works Cited

Mims, Christopher. “AI Data Centers Use Far More Water Than Most Tech Giants Report.” The Wall Street Journal, July 2026. Shehabi, Arman, et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, LBNL-2001637, December 2024. https://escholarship.org/uc/item/32d6m0d1 The Conversation. “Data centers consume massive amounts of water – companies rarely tell the public exactly how much.” August 2025. Information Technology and Innovation Foundation (ITIF). “The Data Center Water Problem Is Soluble.” July 2026. de Vries-Gao, Alex. Research on indirect vs. direct water consumption at technology companies, VU Amsterdam, 2026 (as cited in Hacker News discussion of WSJ reporting). U.S. Environmental Protection Agency. “How We Use Water.” 2025. https://www.epa.gov/watersense/how-we-use-water MOST Policy Initiative. “Data Center Water Use.” Science Note, 2026. Note on the $1 trillion AI infrastructure spending figure: multiple 2026 industry and financial-press estimates (Financial Times, Fortune, Futurum, Goldman Sachs) place cumulative hyperscaler AI capital expenditure in the same general range — roughly $650–750 billion in 2026 alone, layered on top of prior 2025 spending — though exact totals vary by source and by what is counted as “AI infrastructure.” Treat this as a order-of-magnitude figure rather than a single precise number.

The water number in a sustainability report may tell only part of the story.

Artificial intelligence is being built at unprecedented scale. Microsoft, Google, Amazon, Meta and other technology companies are investing hundreds of billions of dollars in new computing infrastructure, with industry analysts and financial reporting pointing toward roughly $1 trillion in cumulative AI infrastructure spending across 2025 and 2026 alone. But there is another infrastructure requirement receiving considerably less attention: Water. As AI data centers become larger, denser and more power-intensive, the question is no longer simply how much water a facility uses to cool its servers. The bigger question may be:

How much water does the entire system require to operate that data center — including the water consumed to produce the electricity powering it?

That distinction could fundamentally change how we understand the resource footprint of AI.

The water you can see — and the water you can’t

When companies publish sustainability reports, data center water consumption is generally discussed in terms of direct water use. Think cooling towers, chillers and other systems located at the data center itself. This is an important measurement. But it doesn’t necessarily capture the water consumed elsewhere in the energy system that supplies electricity to the facility. Power plants can require significant quantities of water for cooling and other processes. The amount varies dramatically depending on the type of electricity generation, the technology used and the location. In other words, a data center can reduce its own on-site water consumption while still relying on an electricity supply whose production has a substantial water footprint. That isn’t necessarily deceptive. It is an accounting boundary. And accounting boundaries matter. A July 2026 Wall Street Journal investigation by technology columnist Christopher Mims highlighted this distinction, reporting that Microsoft, Google and Amazon generally disclose water used directly at their data centers, while Meta stands out as the major operator that also accounts for water used at the power stations supplying its data centers’ electricity. The result is that two companies can appear to have very different water footprints — or similar ones — depending entirely on what their reporting boundary includes.

The 12x problem

This is where the issue becomes particularly interesting. A 2024 report from Lawrence Berkeley National Laboratory estimated that, in 2023, the indirect water consumption associated with electricity generation for U.S. data centers was roughly 12 times their direct water consumption for cooling. That statistic requires an important clarification. It does not mean that every individual data center literally consumes twelve times more water than the figure appearing in its sustainability report — the ratio varies by cooling technology and by the water intensity of the local power grid. It means that, at the broader U.S. data-center level, estimated water consumption associated with electricity generation can be dramatically larger than the water consumed directly for on-site cooling. The distinction is critical. If a facility reports only its direct cooling water, the number can be fully accurate within that reporting boundary while still failing to describe the full resource footprint of the computing system it belongs to. That is the difference between measuring a facility and measuring a system.

AI changes the equation

AI makes this problem more important because AI computing is unusually resource intensive. Training and running increasingly sophisticated models requires enormous amounts of computing power. That computing power translates into electricity demand. Electricity demand translates into additional generation capacity. And depending on the generation mix, electricity production can carry its own substantial water requirements. At the same time, the computing equipment itself generates significant heat. That creates a complicated optimization problem: Use more water for cooling, or use more electricity to reduce direct water consumption? Neither number exists in isolation. A cooling technology that looks extremely water-efficient at the facility level may require more electricity. If that additional electricity comes from a water-intensive generation source, some of the water footprint has effectively moved upstream, not disappeared. Conversely, a facility using evaporative cooling may consume more water directly while requiring less electricity for cooling overall. The right question therefore isn’t simply: “How much water does this data center use?” It is: “What is the total water intensity of delivering the computing services this data center provides?” That is a much harder question. It is also a much more useful one.

Location changes everything

There is no universal “AI data center water footprint.” The answer depends on where the facility is built, how it is cooled, what electricity supplies it and what water sources are available locally. A data center operating in a water-abundant region with a low-water-intensity electricity mix presents a very different resource picture than an identical facility operating in a water-stressed region on a grid still leaning on thermoelectric generation. This becomes especially significant in places such as Arizona, where rapid data-center development intersects with long-standing concerns about water availability. The physical location of compute is therefore not merely a real-estate decision. It is a resource decision. And increasingly, it is an infrastructure-resilience decision.

Reporting water isn’t enough if the boundary keeps changing

There is another problem hiding inside the numbers: comparability. Research into technology-company disclosures has found substantial differences in how companies report water use. Some provide facility-level information. Others report broader corporate figures. Some disclose data-center-specific water consumption, while others publish water metrics that are difficult to isolate to data centers at all. That makes comparisons difficult. Imagine two companies: Company A reports 1 million gallons of direct data-center water consumption. Company B reports 2 million gallons — but that figure includes additional categories of indirect water consumption Company A doesn’t disclose at all. Which company actually uses more water? The answer may be impossible to determine from the headline numbers alone. This is why sustainability reporting needs more than a number. It needs context, methodology and provenance.

The real issue isn’t transparency versus secrecy

It is tempting to characterize this as a story about Big Tech hiding water consumption. The reality may be more nuanced. Companies are often reporting legitimate measurements according to established accounting methodologies. The problem is that those methodologies may not capture the entire system. That distinction matters because the solution isn’t simply to demand that companies “report more.” The solution is to establish consistent, auditable resource accounting across the industry. For AI infrastructure, that could mean reporting at least four layers:

1. Direct water consumption — water used inside the data center.
2. Indirect electricity-related water consumption — water associated with generating the electricity consumed.
3. Local water stress — whether the facility’s water source sits in a water-constrained region.
4. Resource intensity per unit of compute — how much water and energy are required to produce a defined amount of computing output.

That would provide a far more meaningful picture of AI’s physical infrastructure than any single headline number.

From sustainability reporting to infrastructure intelligence

This is where the conversation becomes bigger than water. AI infrastructure increasingly sits at the intersection of energy, water, land, compute, supply chains and local infrastructure. A decision to build a 1-gigawatt AI campus isn’t simply a decision about servers. It potentially affects electricity generation, transmission capacity, cooling systems, municipal water supplies, wastewater infrastructure, construction materials and surrounding communities. Those systems are interconnected. Yet the data used to make decisions about them is often fragmented across companies, utilities, municipalities, regulators and infrastructure operators. That creates a trust problem. If decision-makers can’t see the complete resource chain, they can’t accurately assess the consequences of infrastructure decisions. And if two organizations measure the same resource using different boundaries, their numbers may be technically correct while still producing completely different conclusions about the same facility.

AI needs a better resource ledger

The irony is that artificial intelligence may ultimately help solve this problem. AI systems can correlate infrastructure data across otherwise disconnected sources: energy consumption, cooling requirements, water availability, power-generation mix, weather, grid conditions, facility utilization and geographic constraints. But AI cannot make fragmented data trustworthy simply by analyzing it. The underlying data needs provenance. It needs to be possible to determine: Where did this number come from? What exactly does it measure? What does it exclude? When was it measured? Can another source corroborate it? And perhaps most importantly: Does the number describe the facility — or the system? That distinction is becoming increasingly important as AI infrastructure expands.

The next generation of data-center reporting

The future of AI infrastructure shouldn’t require choosing between economic growth and environmental responsibility. It should require better information. A modern data center should be evaluated not only by its computational capacity, uptime and power availability, but also by the total resource system required to sustain it. Water should be part of that equation. So should energy. So should geographic scarcity. And so should the uncertainty surrounding the data itself. The Wall Street Journal‘s reporting is therefore less a story about whether a particular technology company is “telling the truth,” and more a warning about how easily technically accurate numbers can add up to an incomplete picture. The AI infrastructure race is moving too quickly for resource accounting to remain fragmented. If we’re going to measure the future, we need to measure the whole system — not just the part that’s easiest to report.

The PulseDNA Perspective

At PulseDNA, we believe trustworthy infrastructure decisions require more than a number. They require context, provenance, corroboration and continuity across systems. AI infrastructure is becoming a physical system as much as a digital one. The organizations that can connect those layers — compute, energy, water, location, risk and evidence — will be better positioned to build AI infrastructure that is not only powerful, but resilient and accountable. Because the future of AI isn’t just about how much compute we can build. It’s about whether we can build it intelligently. Image is made by AI and for illustrative purposes only.

---

Works Cited

Mims, Christopher. “AI Data Centers Use Far More Water Than Most Tech Giants Report.” The Wall Street Journal, July 2026. Shehabi, Arman, et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, LBNL-2001637, December 2024. https://escholarship.org/uc/item/32d6m0d1 The Conversation. “Data centers consume massive amounts of water – companies rarely tell the public exactly how much.” August 2025. Information Technology and Innovation Foundation (ITIF). “The Data Center Water Problem Is Soluble.” July 2026. de Vries-Gao, Alex. Research on indirect vs. direct water consumption at technology companies, VU Amsterdam, 2026 (as cited in Hacker News discussion of WSJ reporting). U.S. Environmental Protection Agency. “How We Use Water.” 2025. https://www.epa.gov/watersense/how-we-use-water MOST Policy Initiative. “Data Center Water Use.” Science Note, 2026. Note on the $1 trillion AI infrastructure spending figure: multiple 2026 industry and financial-press estimates (Financial Times, Fortune, Futurum, Goldman Sachs) place cumulative hyperscaler AI capital expenditure in the same general range — roughly $650–750 billion in 2026 alone, layered on top of prior 2025 spending — though exact totals vary by source and by what is counted as “AI infrastructure.” Treat this as a order-of-magnitude figure rather than a single precise number.