How Much Water Do AI Data Centers Use, Compared to Heavy Industry?

Almost every article about AI’s water consumption is technically correct. They also routinely contradict each other.

A single ChatGPT query costs about 0.32 mL of water on-site. A quarter-pound hamburger costs 2,400 liters.

I have an engineering background and work as a software developer, using LLMs daily. When I first ran into these numbers, my instinct was that one side had to be wrong. So I pulled the actual sources instead of guessing.

Some say a ChatGPT prompt uses less than half a milliliter of water. Others say a single email costs an entire bottle. Neither number is wrong. They’re measuring different system boundaries, and almost nobody says which one they mean.

Nobody was lying. They were answering different questions.

The comparison only makes sense once the accounting boundary is defined.

The Accounting Boundary

Every viral AI-water claim measures different layers of the computing stack. To resolve the contradictions, we have to look at the three primary domains of water consumption:

• On-site cooling: Water evaporated by the data center’s local cooling towers.
• Electricity generation: Water evaporated at off-site power plants to generate the electricity powering the servers.
• Hardware manufacturing: Water consumed by semiconductor fabs to produce the silicon itself.

When you classify the most widely cited figures by their boundary, the 1,600x variance between them vanishes into a predictable scaling law based on included scope and computational intensity.

The absolute number matters less than the parameters. A multi-step reasoning model draws 50–100x more power than a standard completion, driving up both on-site thermal load and grid-electricity water consumption.

Comparing industries also requires standardizing the type of water measured. Data centers rely on blue water (surface or groundwater permanently removed from a local watershed). Agriculture relies heavily on green water (rainwater absorbed by soil). Comparing liters consumed by an AI prompt directly to liters consumed by agriculture conflates local aquifer stress with natural rainfall cycles.

None of the commonly cited per-prompt estimates include the water required to manufacture the accelerator itself.

Semiconductor fabs require Ultrapure Water (UPW)-municipal water stripped of dissolved solids, ions, and gases. Producing one unit of UPW requires 1.4–1.6 units of standard municipal water.

TSMC, which fabricates the vast majority of AI accelerators, withdrew roughly 101 billion liters of water in 2023. The company discloses consumption of 161 liters per 12-inch-equivalent wafer. Because modern AI accelerators require 70–100+ mask layers, a single wafer’s water footprint reaches thousands of liters before dicing. While TSMC plans 90% on-site recycling at newer facilities, a true cradle-to-query lifecycle calculation would amortize this manufacturing footprint over the chip’s inference lifetime.

Steel (Physics vs. Engineering)

Thermodynamics sets a lower bound. You cannot optimize your way out of phase-change physics.

An integrated steel plant withdraws roughly 28,600 liters of water per metric ton produced. Modern recycling loops bring actual consumption (water lost to evaporation) down to 1,500–4,200 liters per ton.

Molten steel exits the caster above 1,500°C and must be quenched to roughly 800°C using high-pressure spray. Removing roughly 450 MJ of heat from a ton of steel requires evaporating water. Engineering estimates put the physical floor at 175 liters just for the secondary spray-cooling phase.

An AI inference cooling load is an engineering problem you can design around using ambient air or closed-loop liquid cooling. Steel quenching is mandated by physics.

AI vs Everyday Products

Using 0.32 mL for on-site consumption and 15 mL for a conservative, full-lifecycle figure:

At the level of an individual transaction, AI’s direct water footprint is negligible next to manufactured or agricultural goods. However, scaling this footprint to billions of daily queries requires looking at aggregate corporate infrastructure.

Aggregate Scale

Lawrence Berkeley National Laboratory estimates US data centers directly consumed about 66 billion liters on-site in 2023. Including grid electricity pushes total lifecycle consumption to 800–860 billion liters. Based on these figures, over 91% of a data center’s total water footprint comes indirectly from off-site power generation (734B indirect / 800B total).

The raw withdrawal volumes reported by major tech companies reflect this massive scale.

Globally, FAO AQUASTAT data indicates freshwater withdrawal is dominated by agriculture. Data centers represent approximately 0.5% of the industrial subset. The absolute volume is low, but the growth rate is steep.

The PUE/WUE Trade-off

Evaporative cooling towers yield excellent Power Usage Effectiveness (PUE ~1.1) because water’s latent heat of vaporization (2,260 kJ/kg) removes thermal energy efficiently. A 1 MW facility can evaporate 70,000 liters daily.

Switching to dry, air-cooled chillers drops on-site Water Usage Effectiveness (WUE) to zero, but raises PUE to 1.4+ due to inefficient air conduction. Optimizing only for WUE can simply relocate water consumption to the electricity grid. Thermoelectric plants evaporate water to condense turbine steam (coal ~2.6 L/kWh, nuclear ~2.2 L/kWh).

To illustrate this, here is a comparison of how different cooling systems shift the burden between electricity overhead (PUE) and direct water use (WUE):

Direct Liquid Cooling (DLC) effectively resolves this by circulating fluid through micro-channel cold plates on the die. Capturing 80% of the thermal load via sensible heat transfer removes the need for evaporation, minimizing both PUE and WUE.

Location Over Volume

Water is not fungible. A ton of CO2 has a uniform atmospheric impact, but a billion liters of evaporated water has wildly different consequences in Seattle versus Phoenix.

xAI’s Colossus facility in Memphis illustrates this geographic vulnerability. The site draws on the Memphis Sand Aquifer, the sole drinking water source for 1.3 million people. xAI requested up to 3.7 million gallons a day for its first two sites. A promised wastewater recycling plant broke ground in late 2025 but halted construction in April 2026. Without recycling, the local aquifer is actively pumped faster than its recharge rate.

This local depletion occurs despite data centers being a global rounding error in freshwater accounting.

Water-Positive Commitments

Microsoft claims it achieved “water positive” status in FY2025. Google reports 78% replenishment for 2025, and Amazon projects reaching 75% of its 2030 target.

These volumetric commitments are executed through watershed restoration and municipal leak repairs. However, replenishing a watershed 500 miles away does not physically replace the water withdrawn from the Memphis Sand Aquifer. Offsets balance global ledgers, but they cannot restore specific local wells.

At the individual-query level, AI’s direct water footprint is small next to almost any manufactured or agricultural good - a burger costs more water than 160,000 AI prompts at the heaviest widely-cited lifecycle estimate. At the aggregate global level, the hyperscale data center industry remains a small fraction of industrial water use, smaller than a single major semiconductor fab.