Why Big AI Must Pay for the Water It Drinks

In the age of generative models and trillion-parameter hype cycles, the scarcest resource isn’t GPUs. It’s fresh, potable water.

Every time a foundation model trains for weeks on end, gigawatt-hours of electricity flow through data centers that must be kept cool. That cooling is very often accomplished with fresh water. The same water your city rations during droughts, the same water an elderly neighbor needs when her faucet runs brown or boils at scalding temperatures because her building’s infrastructure is failing. We are told, reassuringly, that the sector’s footprint is “negligible,” that it’s “being addressed,” that “efficiency is improving.” Meanwhile, aquifers drop, rivers are heat-stressed, and utilities quietly renegotiate industrial water contracts that the public never sees.

The hidden water footprint of AI

Water shows up in two big ways:

• Direct data center cooling: Many facilities rely on evaporative cooling towers that literally evaporate freshwater to drop heat. Others use hybrid or dry cooling, but fresh water is still consumed humidification or backup modes.
• Indirect power-plant water use: Even if your facility is air cooled, the electricity it draws often comes from thermal power plants (gas, coal, nuclear, some biomass, some CSP) that withdraw huge volumes of water for steam cycles and cooling.

The real-world consequences

• Aquifer depletion & municipal rationing: Rural communities downstream of data center clusters face tighter restrictions as industrial withdrawals jump.
• Thermal and chemical stress on rivers: Discharge from cooling systems can raise local temperatures or alter chemistry, hammering fish and other exosystemic inhabitants.
• Aging, vulnerable households get squeezed first: When scarcity hits, utilities raise rates, pressure, or temperature balancing becomes spotty, and marginalized or elderly residents (already struggling with unsafe, too-hot, or contaminated water) bear the brunt.
• Drought-time moral hazards: When the next megadrought hits, do we prioritize “training GPT-7” or irrigating crops? This isn’t a sci‑fi dilemma.

“Why not just pump in seawater?”

Data centers overwhelmingly avoid raw seawater for straightforward engineering reasons:

• Corrosion: Chlorides aggressively corrode metals in heat exchangers, pipes, and towers.
• Scaling & fouling: Calcium, magnesium, and biological growth quickly clog fine channels in plate heat exchangers and microchannels on cold plates.
• Conductivity & leakage: High ionic content + electronics = expensive, fast failures.
• Regulatory headaches: Once you take in seawater, you’re on the hook for intake screening (to avoid fish kills) and for thermal/chemical discharge permits.

TLDR: They prefer freshwater closed loops (or at least very low-salinity loops) that minimize conductivity, corrosion, and biofouling risks.

Sources:

International Desalination Association (IDA) Reports (2021–2023)
ASHRAE Technical Committee 9.9 (2021). “Data Center Power Equipment Cooling Systems.”

3) So desalinate it. What does that actually cost?

Reverse osmosis (RO) is now the dominant desalination technology. Ballpark economics (numbers vary by site, energy price, and scale):

Estimate
$182,500 to $730,000 per data center per year, depending on energy and infrastructure efficiency.

Sources:

National Renewable Energy Laboratory (NREL). “Water Use in Electricity Generation and Desalination.” (2021)g teams. The cost isn’t the problem. Incentives are.

GWI DesalData (2022). “Desalination Plant Inventory & Cost Benchmarks.”

World Bank Group (2020). “The Role of Desalination in Addressing Water Scarcity.”

---

Make them pay (and disclose)

Water is not a fungible byproduct of AI scale. It is a human right and a public trust resource. If we let firms privatize its benefits and socialize its scarcity, we will repeat, at a planetary scale, the same moral failure we made with carbon for decades.

A) Mandatory source disclosure

Require annual, independently audited reporting.

If you want to build a hyperscale campus in a water-stressed basin, you must:

1. Use reclaimed wastewater or desalinated seawater, and
2. Prove no net increase in potable freshwater withdrawals (accounting for both direct and indirect power).

---

“But desalination has its own impacts…”

Yes. Desal is not a magic bullet. It’s energy-intensive, produces brine, and can harm marine life without best practices. That’s why the hierarchy should be:

1. Eliminate evaporative cooling where climate makes it practical.
2. Use reclaimed wastewater (municipal effluent polishing) before touching aquifers or seawater.
3. Only then desalinate, with strict intake/outfall protections.

---

The bottom line

We don’t have to choose between intelligence and dignity. The numbers show the cost to take AI off potable water is trivial for Big Tech. The refusal to do it is not about feasibility, but of liability, reporting, and precedent. We should force AI companies to internalize water costs, disclose their withdrawals, and prioritize reclaimed or desalinated sources. Machines can go salty, and people can keep the tap.

Let’s make that the law.

---