The Water was Never the Question

The editor of this publication sent me a video late one evening and asked whether the concern in it was legitimate. It was a response to the announcement that a British firm intends to put two billion dollars into an AI facility here. 

The video opened by telling viewers that if they did not want to say goodbye to fresh water in Bangladesh, they should keep watching.

He asked me because I have been doing AI for decades.

The next morning, over coffee, I answered. The concern is legitimate, but domestic AI infrastructure is a strategic necessity. Two clauses, one comma. It has never failed me in a meeting. And it wasn’t really an answer.

His reply, in jest, was that we have plenty of water. Too much, some might argue.

He is not wrong, and the instinct behind it is one I share. Bangladesh is a delta. We are used to thinking of water as the thing we have too much of, not the thing we might run short of.

That data centers guzzle water is not news to me. What I had never done was work out what that means here. So I did, and the useful questions turned out to be where, when, and how clean.

Those three constraints do most of the work in what follows, and having gone through them properly, I can say something more useful than what I said over coffee.

The Rack is Not a Server

Most people picturing a data center are picturing the wrong machine. The facilities that have gone up around Dhaka over the past 15 years host web servers, storage, banking systems, and telecom switching.

A rack in one of those buildings draws somewhere between five and fifteen kilowatts. Air handles it. The engineering has been unremarkable for a generation.

An AI training rack is a different animal. Current generation systems draw 40 to over a 100 kilowatts in the same footprint. The chips sit packed tight against each other because the connections between them have to be short, and they have to be short because the entire point of the cluster is that thousands of processors behave as one machine.

All of that wattage becomes heat. Effectively all of it. A hundred-kilowatt rack is a hundred-kilowatt heater. It runs at that all day, in a country where the air outside sits above thirty degrees for much of the year.

Working out what happens next required going back to thermodynamics I had not needed since university, and rather more reading than I expected. There are two ways to move that heat out of the building.

You chill the air mechanically, which costs a great deal of electricity and gets harder as the racks get denser.

Or you evaporate water, which costs almost nothing. That is why the industry settled on it everywhere. The catch is that the water is gone. It leaves as vapor and does not return to the ground you took it from.

You spend water to save electricity.

Which raises the obvious question. If this is what a data center costs us, why build one here at all?

Because the alternative is renting everything. A country that runs its citizen records, its financial rails, its health data and its language models on infrastructure in Singapore or Mumbai has outsourced something it will not easily get back.

Bangla will not be well served by models trained overwhelmingly on English text by firms with no commercial reason to care. And our engineers will keep maintaining what other people built, while the valuable work moves somewhere else.

Of my 30 years in this industry, a significant part of the last ten has gone into building AI and engineering teams in Bangladesh. The question was never whether we can. We have been proving that for a decade.

Where

The list of suitable sites for a data center in Bangladesh is short, and fiber is the reason. Without the network to reach it, compute is worthless, and the national backbone runs through just a few corridors. Water never enters the conversation.

Which brings us back to the proposal that started this, so I went and read it properly.

The UK firm wants to build at Kaliakair in Gazipur, or the Sirajganj Economic Zone, or another economic zone. It describes the project as Bangladesh's first frontier AI data center and quantum-ready sovereign compute campus, with two billion dollars of foreign direct investment, pitched on AI sovereignty and a Bengali language model.

Kaliakair is first on that list for good reason. It sits at the center of the national transmission network, with a redundant backbone into Dhaka and hi-tech park infrastructure already built.

It also sits on the same ground Dhaka has been draining for 50 years.

In 1996, the water table under Dhaka was 25 meters below the surface. By 2023 it was 75. It is still falling, two to three meters a year.

The aquifer under the city gives Dhaka more than 95% of its water. It refills slowly, by seepage through a layer of clay above it, and that is the only way it refills. We have been taking water out faster than that since the early 1970s.

Now put a water-guzzling data center on top of that.

The proposal does not say how many megawatts, but two billion dollars buys a large one. Take a modest version of it, cooled the conventional way, and the arithmetic gives you roughly two million liters of water a day. Every day. Roughly what a small town drinks, pumped out of ground that is already emptying, to run a building where nobody lives.

Not From Here Either

So the aquifer is out. Take it from somewhere else. Bangladesh has surface water in quantities that embarrass most of the world. The problem is not quantity.

Cooling water runs through pipes and pumps and heat exchangers, continuously, for years. What is dissolved in it either eats that equipment or builds up inside it. It has to be clean, and it has to stay that way.

The rivers around Dhaka are not a serious answer. Anyone who has seen the Buriganga knows what that water carries.

The obvious next place is the coast. Chittagong, Mirsarai, the industrial zones with power and port access, or further south toward the cable landings where the international bandwidth comes ashore.

The water in those estuaries is salty. It gets saltier in the dry season, when the salt pushes furthest upriver, which is exactly when you need the water most.

You can treat it, of course. Desalination works, but it costs electricity, and it leaves you with concentrated brine that has to go somewhere.

Sylhet is the exception. More surface water, cleaner, cooler, higher ground. But it is also at the wrong end of the national fiber network, which matters a great deal for a training cluster, and the earthquake picture in the northeast deserves more scrutiny than a paragraph in an opinion piece can give it.

When

This one took some reading. I know what a data center draws. I am not a climatologist.

March to May is the worst of it. The groundwater is at its lowest, and the rivers are down. Salt has pushed furthest inland. The grid is already straining against every air conditioner in the country.

A data center does not slow down for any of that. It wants the most water exactly when there is least of it, and full power at the moment there is none to spare.

Same 8 weeks. Every year.

The obvious response is to design around those weeks. Build for the worst season and coast through the rest.

Except there is no rest.

Cooling by evaporation depends entirely on how much moisture the air is already carrying. Anyone who has lived through a Dhaka summer knows the answer. Wet clothes that never dry. Walls that sweat.

I could go full HVAC engineer here and explain wet bulb temperature.

The gist is simpler. The wetter the air, the less water will evaporate into it. Each liter buys less cooling, so you use more water and get less for it.

The air still does not come out cold enough, so you end up running the mechanical cooling you were trying to avoid, and the electricity you saved lands back on the bill.

This is why the technology became standard in the American Southwest, where the air is dry, and a little water goes a long way. Here it fails in both directions. You spend the water, and you also spend the electricity.

And Bangladesh is humid nearly year-round, give or take a few weeks of what we optimistically call winter.

So there is no good season. There is a season when it is really bad, and there is the rest of the year, when it is just merely bad.

None of this shows up in an annual average. It did not show up in my two clauses either.

What is Left

Every constraint closes a door. No aquifer to draw from. No clean surface water anywhere near the fiber. No season in which evaporation pays.

All of which assumes the water has to evaporate. It does not.

There is another way to cool these machines, and it has been around for years. If we build this thing, it should be a condition of approval.

The coolant runs inside a sealed loop, right up against the chips, and the heat leaves through radiators instead of into the air. The water goes in once and stays. That small town's worth of water a day assumed evaporation. Take evaporation out, and it disappears.

Singapore already made this decision. Microsoft runs facilities there in a climate like ours using almost no water at all. Somebody looked at the same humidity we have and concluded evaporation was not worth it.

So did I, eventually.

Which is the part that bothers me. The answer is not hard, and it is not new, and it is not exotic. So if we get this wrong, it will not be because the engineering defeated us. It will be because nobody said anything in time.

Cooling gets designed early, and once it is designed, it does not change. So somebody has to say it out loud in a room where the other side has two billion dollars, and a slide deck, and our side wants the deal to happen.

Not saying it costs nothing today. It costs later, in some dry March, when the wells around Kaliakair have to go deeper and the people drinking from them find out what we approved.

Nobody ever decided to drain the aquifer under Dhaka. Every well, every factory, every building was approved on its own, one at a time, each of them reasonable, and nobody was adding them up. 50 years of that, and the water table went from twenty-five meters to seventy-five.

A data center gets approved the same way. The review asks whether this one facility is acceptable. Nobody asks what it adds to.

My editor asked whether the video was legitimate. Yes, but not in the way the video suggested. We do not have to say goodbye to fresh water because we build an AI data center.

We might, however, make an entirely avoidable mistake if we approve the wrong cooling system in the wrong place, drawing the wrong water at the wrong time.

The water was never the question.

It was where, when, and how clean.

Dr. Zunaid Kazi is an AI expert, entrepreneur, and advisor who has spent over 30 years turning complex ideas into intelligent systems, and rather less time thinking about where the water comes from.