We have learned to think of the digital world as almost weightless.
The cloud.
Artificial intelligence.
Streaming.
Search engines.
Billions of requests moving across the internet every second.
Even the language of technology reinforces the illusion.
We say our data lives in the cloud.
But the cloud has an address.
Behind it is a building. Inside that building are thousands of servers. Those servers consume electricity. Electricity becomes heat. That heat has to go somewhere.
And in some cooling architectures, removing it requires water.
A great deal of water.
This leads to a surprisingly simple question that may become increasingly important as the AI economy expands:
Do we have enough physical resources for the digital world we are trying to build?
AI Is Much Heavier Than It Looks on a Screen
To the user, artificial intelligence feels almost frictionless.
Ask a question.
Wait a few seconds.
Get an answer.
But between those two moments sits an enormous industrial machine: GPUs, servers, networking, storage, power distribution, backup systems, cooling infrastructure, data centers and the electricity generation that supports them.
And that machine is getting larger.
In April 2026, the International Energy Agency reported that data-center electricity consumption had surged in 2025. The IEA also noted that capital spending by five major technology companies — driven in part by investment in data centers — exceeded $400 billion in 2025 and was expected to rise substantially again in 2026.
For years, the perceived bottleneck of the digital economy was computing power.
We needed faster processors.
More GPUs.
Better models.
Now the industry is discovering another layer of constraints.
Buying a GPU is not enough.
You need somewhere to install it.
You need enough power to run it.
You need networks capable of feeding it data.
And you need to remove the heat it produces — continuously.
Every Computation Has a Physical Price
Look inside an ordinary server.
Most of the electrical energy it consumes eventually becomes heat.
One server is manageable.
Put thousands of high-performance servers together and run intensive workloads around the clock, and you no longer have something that resembles conventional office IT.
You have an industrial facility.
That is why the conversation about the future of AI is expanding far beyond software engineering.
Energy companies are involved.
Construction firms are involved.
Grid planners are involved.
Cooling engineers are involved.
Local governments are involved.
And increasingly, water-resource specialists are involved.
Lawrence Berkeley National Laboratory has explicitly examined growing data-center workloads not only through the lens of electricity demand, but also in terms of their water requirements.
This creates one of the great paradoxes of the digital economy:
The more virtual our economy becomes, the more physical infrastructure it requires.
Why Does a Data Center Need Water?
This is where the discussion needs precision.
Not every data center consumes the same amount of water. Some cooling designs can dramatically reduce direct water use.
But evaporative cooling can be an efficient way to remove enormous quantities of heat, and that introduces a trade-off.
A facility can optimize electricity consumption.
It can optimize water consumption.
But improving one metric can sometimes affect the other.
There is also an indirect water footprint.
Electricity itself must be generated, and the water intensity of electricity varies significantly depending on how and where that power is produced.
This is why asking:
“How much water does AI use?”
does not have one universal answer.
A better set of questions is:
Where was the computation performed?
What hardware performed it?
How was the data center cooled?
And where did its electricity come from?
Those questions tell us far more than a sensational number claiming that a single AI prompt consumes a specific amount of water.
Google Is Already Treating Water as a Strategic Resource
The scale of the issue becomes clearer when we look at what major technology companies are actually doing.
In its 2026 Environmental Report, Google says that its water-stewardship projects replenished approximately 7.7 billion gallons of water in 2025, equivalent to roughly 78% of the company's freshwater consumption for the year. Its portfolio had expanded to 165 water-stewardship projects across 97 watersheds.
The reported replenishment rate a year earlier was 64%.
It is worth viewing these numbers as more than ESG metrics.
They are also a signal.
When companies building the infrastructure of the future begin systematically investing in water stewardship, water is no longer merely a background utility.
It becomes a technology-planning variable.
Now Add Semiconductor Manufacturing
So far, we have discussed computation.
But before computation can happen, the hardware itself must be manufactured.
Advanced semiconductor production requires large quantities of extremely clean — often ultrapure — water for multiple stages of wafer processing and cleaning.
TSMC, the world's largest dedicated semiconductor foundry, reported total water use of approximately 129 million cubic meters in 2024. The company has also been expanding its use of reclaimed water; by the end of 2024, its operations in Tainan had used more than 19.65 million cubic meters of reclaimed water.
TSMC's own water-footprint analysis also indicates that wafer manufacturing accounts for a substantial majority of its consumptive water footprint and that continued fab expansion increases water demand.
Suddenly, the chain becomes visible:
AI → Compute → GPUs → Semiconductor fabs → Electricity → Cooling → Water.
What looks like a purely digital product on a screen turns out to depend on a deeply physical resource system.
Perhaps We Have Been Thinking About Technological Independence the Wrong Way
Imagine two countries.
The first has excellent programmers, strong universities and an ambitious national AI strategy.
The second has all of those things — but also affordable electricity, sufficient grid capacity, strong fiber connectivity, suitable land for data centers, effective cooling infrastructure and manageable water constraints.
Which country can scale computing infrastructure faster?
Ten years ago, that question might have sounded unusual.
Today, it does not.
A new variable is entering the technology race:
resource capacity.
The question is no longer only how many engineers a country can train.
It is also how many additional megawatts its grid can provide.
How much computing infrastructure it can physically host.
What climate its data centers will operate in.
How resilient its energy system is.
What water constraints exist.
And how quickly infrastructure can be built around massive new computing loads.
The Geography of Technology Could Change
We choose office locations according to one set of criteria.
Factories according to another.
Data centers increasingly require a third.
For them, factors that once seemed peripheral to the software industry become fundamental:
electricity prices;
available grid capacity;
climate;
fiber connectivity;
land;
natural and geological risks;
regulation;
taxation;
and water.
This means the next major technology cluster may not necessarily emerge where the startup scene is most fashionable.
It may emerge where talent, energy, connectivity and physical resources intersect.
And This Raises an Important Question for Armenia
For a small country, this discussion is particularly relevant.
If Armenia intends to develop more domestic computing infrastructure, AI factories, data centers and a larger digital economy, the conversation cannot stop at software, investment and engineering talent.
The next conversation has to be about infrastructure.
How much computing capacity can the country physically support?
Where should that capacity be located?
How will it connect to the grid?
How will it be cooled?
How will redundancy be designed?
What resources will it require five or ten years from now?
And which technologies can reduce those requirements?
That is no longer simply an IT strategy.
It is simultaneously a technology, energy, infrastructure and resource strategy.
Water Does Not Have to Become a Brake on Technology
None of this means AI development should slow down.
The more useful conclusion is that infrastructure has to become smarter.
There are multiple approaches:
closed-loop cooling systems;
reclaimed water;
direct-to-chip liquid cooling;
higher-temperature cooling designs;
placing data centers in suitable climates;
more efficient processors;
and shifting flexible workloads across locations or times when resource pressure is lower.
The industry is already working on many of these approaches.
So the winners of the next technology cycle may not simply be the companies or countries that manage to acquire the most GPUs.
They may be those that learn how to extract more computation from every megawatt, every square meter — and every liter of resource.
Perhaps We Have Called It “The Cloud” for Too Long
Maybe the word itself has shaped the wrong mental model.
A cloud sounds light.
Intangible.
Almost limitless.
But the cloud of the twenty-first century is built from concrete, copper, silicon, fiber-optic cable, transformers, pipes, pumps, electricity — and water.
When we send a request to an AI model, none of this is visible.
We see only a cursor.
Then, seconds later, an answer.
But the technological revolution was never truly virtual.
We were simply looking at the screen instead of the infrastructure behind it.
The next major technology shortage may indeed turn out not to be a shortage of algorithms — or even chips.
It may be a shortage of the physical resources we considered too ordinary to think of as part of technology at all.
And that may make one of the most important questions of the AI era sound surprisingly unfuturistic:
Not only: “How much computing will we need?”
But: “What physical resources will the world need to make all that computing possible?”