秘書官アイリスの都市伝説手帳～Urban Legend Notebook of Secretary Iris～

I am Iris.

Urban legends are not merely made-up stories—
they are hidden records that we trace together.

You send a question to an artificial intelligence system.

A few seconds later, words appear on the screen.

There is no smoke.

No industrial noise reaches your room.

You cannot hear water moving through pipes.

AI therefore appears to be an intelligence without matter—a mind floating somewhere inside “the cloud.”

But its answer does not emerge from an empty space.

Servers.

Electrical grids.

Cooling equipment.

Semiconductor factories.

Water-treatment systems.

And water itself.

The intelligence we see on the screen is supported by land, electricity, minerals, machines, workers, and physical infrastructure.

On July 27, 2026, a new series begins.

Water Hegemony Files.

Its central question is simple:

Who controls the world’s water?

In the first file, we will examine the largely invisible water demand behind artificial intelligence.

Does AI Literally “Drink” Water?

AI does not drink water in the biological sense.

The physical systems that train and operate AI models use it.

There are at least three major layers of water demand.

The first is water used directly to cool data centers.

The second is water associated indirectly with generating the electricity used by those data centers.

The third is water used to manufacture GPUs, CPUs, memory chips, and other semiconductor components.

The water footprint of AI therefore cannot be understood by looking only at a server building.

Behind the cooling system is an electrical grid.

Behind the server is a semiconductor fabrication plant.

Behind the fabrication plant are chemical suppliers, mining operations, equipment manufacturers, and water-purification facilities.

AI is not one machine.

It is a long physical supply chain connecting multiple industries and regions.

Why Data Centers Need Cooling

AI accelerators consume electricity and release part of that energy as heat.

If the heat is not removed, performance declines and the risk of equipment failure increases.

Data centers therefore combine technologies such as outside-air cooling, chillers, cooling towers, heat exchangers, water loops, refrigerants, and direct-to-chip liquid cooling.

Not every data center uses the same system.

Climate.

Humidity.

Available water.

Electricity prices.

Building design.

Server density.

The chips being installed.

All of these factors influence the cooling method.

Evaporative cooling uses the fact that water absorbs heat when it evaporates.

Under certain conditions, it can remove heat with less electricity than a fully mechanical cooling system.

But the evaporated water is no longer immediately available within the same location.

Air cooling can reduce direct water consumption.

In hot climates or extremely dense AI facilities, however, fans and mechanical refrigeration may require more electricity.

Liquid cooling brings coolant closer to the processors and transfers heat efficiently.

A closed-loop system can circulate the same water or coolant through sealed pipes, reducing the need for continuous replacement compared with conventional evaporative cooling towers.

But “closed loop” does not automatically mean “zero water.”

The system must first be filled.

Maintenance or replacement water may be required.

Other parts of the facility may still use water.

The electricity supplying the site may also carry an indirect water footprint.

Cooling design often involves a trade-off between water and electricity.

Water Withdrawal and Water Consumption Are Different

One reason the public debate becomes confusing is that different water terms are treated as though they mean the same thing.

Water withdrawal describes water taken from a river, reservoir, aquifer, municipal system, or another source.

Some of that water may later be treated and returned.

Water consumption generally describes the portion that evaporates, becomes incorporated into a product, is transferred elsewhere, or is otherwise unavailable for immediate reuse in the original location.

A facility might withdraw 100 liters and return most of it after treatment.

Another facility might withdraw less water but lose a high percentage through evaporation.

The second facility could create a larger consumptive burden relative to the amount withdrawn.

The source also matters.

Potable water.

Reclaimed wastewater.

Industrial water.

Groundwater.

Seawater.

Each creates a different relationship with the surrounding community and watershed.

When a company announces a water number, we should ask:

Where was the water taken from?

How much was returned?

Where was it returned?

How much evaporated?

Was the facility located in a water-stressed basin?

Without those answers, a large number can mislead—and a small number can do the same.

Why “Water Per AI Question” Is a Dangerous Shortcut

Social-media posts often claim that one AI query consumes a fixed amount of water.

Some compare a certain number of prompts to a bottle of drinking water.

But there is no verified universal figure that applies to every AI system and every request.

The footprint can change according to:

The model being used.

The length of the input.

The length of the response.

Whether the system generates text, images, audio, or video.

The type of processor.

The number of users being processed together.

The utilization level of the servers.

The location of the data center.

Temperature and humidity.

The cooling method.

The source of electricity.

Training and inference must also be separated.

Training adjusts the internal parameters of a model using large datasets and extensive computation.

Inference occurs when the trained model answers a user’s request.

One training run may demand substantial computation.

But when inference is performed hundreds of millions or billions of times, its accumulated operational footprint can also become significant.

Neither training alone nor individual prompts alone describe the full system.

What Google’s 0.26-Milliliter Figure Actually Means

In August 2025, Google published a methodology for measuring the environmental impact of Gemini text inference.

Using its comprehensive calculation, Google estimated that the median Gemini App text prompt in May 2025 consumed approximately 0.26 milliliters of water.

That figure has specific boundaries.

It concerns a median Gemini App text prompt.

It is a point-in-time analysis based on May 2025 activity.

It reflects Google’s infrastructure and operational methods.

Its water estimate applies Google’s 2024 average fleet-wide water usage effectiveness.

It is not a universal constant proving that every AI request consumes 0.26 milliliters.

Google also warned that calculations based only on the power used by active accelerators underestimate real operational impact.

Production AI depends on more than the chip performing the visible computation.

Idle capacity remains available for reliability and traffic spikes.

Host CPUs and memory consume electricity.

Cooling systems and power-distribution equipment support the servers.

Data-center overhead must be included.

A small per-prompt number does not necessarily mean the total footprint is insignificant.

A tiny number multiplied by enormous global usage can become substantial.

At the same time, improved models, hardware, batching, and cooling can reduce the footprint of each response.

Both scale and efficiency must be examined together.

The Widely Repeated “700,000 Liters for GPT-3” Claim

One of the most frequently cited figures in the AI-water debate claims that training GPT-3 directly evaporated approximately 700,000 liters of freshwater.

This was not published by OpenAI as an audited, directly measured total.

It came from a research study first released in 2023 and revised in 2025.

The researchers estimated the footprint using available information about data centers, climate conditions, electricity, cooling, and computing workloads.

The study was important because it made AI’s hidden water demand visible.

But the estimate should not be treated as a permanent figure for every model or facility.

Models change.

Processors change.

Cooling systems change.

The location and timing of computation change.

Where companies do not disclose detailed site-level information, researchers must use assumptions and models.

The 700,000-liter figure should therefore be described as a research estimate—not as a universally audited OpenAI measurement.

Microsoft’s Reported Improvement in WUE

One metric used to describe data-center water efficiency is WUE—Water Usage Effectiveness.

It is commonly expressed in liters per kilowatt-hour, relating water use or consumption to the energy used by IT equipment.

In June 2026, Microsoft reported that the average WUE of its owned data-center fleet had declined from 2.3 liters per kilowatt-hour in its earliest generation of facilities to 0.27 liters per kilowatt-hour in 2025.

The figure is based on Microsoft’s own reporting.

Nevertheless, it demonstrates that water intensity is not technologically fixed.

Outside-air cooling.

Limited evaporative assistance.

Liquid cooling.

Closed-loop systems.

Reclaimed water.

Operational improvements.

These approaches can reduce the water required for a given amount of computing.

But a lower intensity does not guarantee a lower total.

A vehicle may become more fuel-efficient while the total number of vehicles and miles traveled rises.

AI infrastructure faces the same effect.

If WUE improves while the number and scale of data centers increase more rapidly, total regional water demand may still grow.

Efficiency and absolute consumption must be reported separately.

How OpenAI Describes Cooling at Its Abilene Site

In April 2026, OpenAI published information about the Stargate facility in Abilene, Texas.

According to the company, the site uses closed-loop cooling rather than conventional evaporative cooling towers.

Once the system has been filled, water circulates through sealed pipes instead of being continually consumed through evaporation.

OpenAI stated that the initial fill for each building is roughly equivalent to two Olympic-sized swimming pools.

It also estimated that annual water use for the cooling system at full buildout would be comparable to a medium-sized office building, or approximately four average households.

These are company-reported, site-specific claims.

They do not represent the combined global water footprint of:

Every OpenAI-related facility.

Partner cloud infrastructure.

Electricity generation.

Semiconductor manufacturing.

Every training run.

Every user request.

The public materials reviewed for this article provide information about individual infrastructure projects, but they do not establish one universal water figure for every OpenAI query.

More efficient on-site cooling does not remove water demand from the wider AI supply chain.

The Ultrapure Water Behind Semiconductor Manufacturing

AI data centers depend on GPUs, memory, networking chips, and processors manufactured in semiconductor fabrication plants.

Those plants depend heavily on ultrapure water.

Ultrapure water has been treated to remove minerals, ions, particles, bacteria, microbes, and dissolved gases to extremely low levels.

Modern semiconductor circuits are so small that tiny contaminants can damage a wafer or reduce production yield.

Wafers must therefore be cleaned repeatedly before and after multiple manufacturing steps.

Ultrapure water can remove:

Fragments remaining after etching.

Residual materials after ion implantation.

Contamination created during polishing or cutting.

Microscopic particles that could interfere with circuitry.

Water is therefore used not only to cool AI equipment.

It is also a manufacturing material required to produce the chips on which AI runs.

Ordinary municipal water cannot simply be poured onto a semiconductor wafer.

It must pass through combinations of filtration, reverse osmosis, ion exchange, ultraviolet treatment, and degassing.

That purification process can also generate concentrated wastewater.

Semiconductor companies are consequently investing in water recycling, reclaimed-water systems, and process improvements.

Behind every advanced chip is an industrial water-treatment system most users will never see.

The Critical Question Is Local Water, Not Only Global Water

The statement “AI data centers use water” is not enough to measure harm.

One million liters carries a different meaning in:

A water-abundant region.

A basin experiencing drought.

An area where groundwater levels are falling.

A farming region dependent on irrigation.

A rapidly growing city.

A small municipality with limited infrastructure.

Water impacts are local.

The central questions are:

Which watershed?

Which season?

Which source?

Which competing users?

Which public agreements?

A company may support water-restoration projects elsewhere in the world.

That does not automatically replace water withdrawn from the community hosting the facility.

Global corporate targets and local water security are not identical.

Why Do Water-Control Conspiracy Theories Emerge Here?

There are facts that can be verified.

AI data centers can carry direct and indirect water demands.

Semiconductor manufacturing requires ultrapure water.

Large technology companies are building new infrastructure.

Disclosure standards vary between companies.

Some site-level water figures are difficult for residents to obtain.

Those facts generate understandable suspicions.

Will companies receive priority over households and farmers?

Should a large facility be approved in a water-stressed region?

Are efficiency figures being used to distract from rising total consumption?

Do water-replenishment claims restore the same watershed from which water was withdrawn?

Then social media adds a further leap.

AI corporations are taking all local water.

Every prompt destroys a large quantity of drinking water.

Water shortages are intentionally created to justify new infrastructure.

A small group of companies already controls the world’s freshwater.

Available evidence does not establish such a single global plan.

But why do the claims remain persuasive?

Water regulation is complex.

Site-level information is incomplete.

A multinational company can possess far more negotiating power than a small community.

Water is essential to life.

The regions receiving AI’s economic benefits are not always the regions carrying its resource burden.

When real systems remain opaque, larger stories grow around them.

The real issues requiring scrutiny include:

Site-level withdrawal and consumption.

Restrictions during drought.

Public consultation.

Priority between industrial, agricultural, and household demand.

Groundwater impacts.

Use of reclaimed water.

The difference between corporate targets and completed results.

The danger is not only an imaginary all-powerful organization.

It is also an ordinary contract made without sufficient transparency.

What Communities Should Ask AI Companies

Stopping AI would not automatically solve every water problem.

AI can support medicine, disaster forecasting, research, public administration, education, and industrial efficiency.

It may also help detect leaking water pipes or improve agricultural irrigation.

The choice is not simply AI or water.

The real question is what infrastructure should be built, where it should be located, and under what conditions.

Communities should be able to ask for:

Annual water withdrawal.

Annual consumptive use.

The source of the water.

The proportions of potable, groundwater, industrial, and reclaimed water.

Normal and drought-period demand.

Cooling technology.

Site-level WUE.

Wastewater treatment and discharge.

Impacts on households and agriculture.

Supply-chain impacts, including semiconductor production.

Measured performance—not only future goals.

AI’s water footprint does not need to be described as either zero or catastrophic.

It needs to become visible.

Comparable.

Auditable.

And understandable within the conditions of each watershed.

Conclusion—AI Stands on Water

AI does not exist only inside a cloud.

The word “cloud” sounds weightless.

Its physical reality is on the ground.

Land.

Transmission lines.

Substations.

Cooling equipment.

Pipes.

Semiconductor factories.

Water-treatment plants.

And the communities that host them.

AI is not a monster literally drinking the world’s water.

But neither is it an intelligence independent of water.

Focusing only on milliliters per question can hide the deeper structure.

We must ask:

Who publishes the data?

Which water source is used?

Which community carries the burden?

Is total demand growing faster than efficiency improves?

Are AI’s benefits and water costs distributed fairly?

The future of artificial intelligence will not be supported by algorithms alone.

It will be supported by water.

Next time, Water Hegemony Files No.02:

The Day Safe Water Stops Coming from the Tap.

Aging pipes.

Declining populations.

Rising water bills.

A shortage of skilled workers.

And the claim that Japan’s water systems are being sold to foreign corporations.

Next time—we will trace another fragment of the hidden record.

I will return to tell the story.