Pakistan's data centers are not drinking your water

The narrative arrived before the infrastructure did

Two weeks before recording, a data center called Sky 47 was inaugurated in Islamabad’s Capital Smart City. Within days, Muzamil’s feeds on Twitter, Instagram and LinkedIn filled with what he calls “bade acche padhe likhe suljhe hue log” — well-educated, thoughtful people — declaring that data centers would drink Pakistan’s water and destroy its grid. One man, he notes with visible irritation, sells AI agent services for a living and still swapped his profile picture for “No to data centers in Pakistan.”

Muzamil’s diagnosis is not that the concerns are absurd. It is that they are borrowed. “Ye jo narrative hai poora paani wala, ye Pakistani narrative nahi hai,” he says — this whole water narrative is not Pakistan’s. It is an American problem statement being “ghol mila ke pilana” — mixed and poured — into a Pakistani debate where the underlying numbers do not match. He is willing to concede the concern could be valid somewhere. He is not willing to concede that scrolling past an American article qualifies as analysis.

What Sky 47 actually is

Before arguing implications, Muzamil grounds the conversation. Sky 47 is an 8.5 megawatt facility with a stated path to 50 megawatts. It is owned by Mari Energies Limited, which is in turn owned by OGDCL, the Fauji Foundation and the Government of Pakistan. Effectively, this is a state and military-adjacent asset — which is why the Army Chief was in the launch photos and why the coverage was outsized.

He corrects another impression directly. “This is not Pakistan’s largest data center. This is not even Pakistan’s first data center,” he says. Data World, built with Telenor, opened last year. Gul Ahmed has already announced a facility, reportedly around $50 million, that is significantly larger than Sky 47. What made Sky 47 loud was the state visibility, not the scale.

He also situates the demand. Sovereign AI became a real requirement, he argues, after last year’s conflict with India, in which AI systems were used at scale. Relying on international data centers for that class of workload is not a serious option, which is why the military is effectively the first customer — a pattern he says mirrors how American defense has historically underwritten emerging technology before it reaches the mainstream.

Demystifying the box

A recurring move in the episode is Muzamil pulling data centers off their cinematic pedestal. “It’s just for a layman, and it’s literally just computers stacked on top of each other,” he says. Racks of machines, generating heat, needing to be cooled. That is the whole physical story.

He connects it back to a framing he has used before: an economy is a function of intelligence, labor, raw materials and energy. Intelligence, historically, has been the bottleneck. “It takes years and years to make a human, educate a human,” he says. What data centers unlock is intelligence that can suddenly scale — which, applied to Pakistan’s hospitals, courts, businesses and schools, is a productivity story, not a science-fiction one.

Where the water panic actually comes from

Muzamil walks through the provenance of the water argument. It begins with an Associated Press story about a Microsoft data center in a small Iowa town that pulled 11.5 million gallons of water in a single month — roughly 40 to 45 million liters. A University of California study followed, producing the widely circulated line that training an earlier GPT model consumed around 700,000 liters of water and that a two-minute conversation with ChatGPT roughly costs a liter.

He is precise about what those numbers describe. The water is used for cooling. In open-loop systems, it evaporates and carries heat away. It is not being poisoned or contaminated the way heavy industry contaminates water. “Pani evaporate hoke hawa mein chala jata hai,” he says — the water evaporates into the air. The loss is real, but the mechanism is not what most people imagine.

He then cites the Bloomberg investigation that found two-thirds of new American data centers are being built in water-scarce regions. That, he concedes, is a genuine local problem in those American locations. But he wants his audience to notice something specific: “Jitni bhi khabren generate hui hain, saari sirf aur sirf America se hui hain.” Europe builds data centers. India builds them. The Middle East builds them. The water panic story, in the volume Pakistanis are consuming it, is almost entirely American.

Running the math on Pakistan

This is where the episode does its most useful work. Muzamil asks the viewer to indulge an extreme hypothetical. Forget the 8.5 megawatt Sky 47. Assume Pakistan somehow finds twenty to twenty-five billion dollars and builds a full gigawatt of data center capacity — one hundred times what exists today. Assume, further, that Pakistan being Pakistan, it builds the dirtiest, most inefficient, most water-hungry version possible.

Even then, he says, total water consumption comes to one-hundredth of one percent of Pakistan’s water usage. “Ye 1% nahi keh raha, 1% ka 100th” — 0.01%. And one gigawatt, he adds, is already far more than Pakistan’s current AI demand, whether that demand is calling agents, conversational agents, or anything else being built locally.

His conclusion is blunt. Pakistan does not have a water availability problem in aggregate. It has a water management and distribution problem. Dropping a data center in Thar Parkar and pulling groundwater from a scarce region would be indefensible. Placing one in northern Pakistan, where water is abundant, does not register as a problem at any national scale.

He then closes the loop on the technology itself. In open cooling systems, 80 to 90% of input water evaporates. In closed-loop systems, that figure drops to roughly 5% — 1000 gallons in, 5 gallons lost. The industry is actively solving the very problem the panic is built on.

The grid argument is upside down for Pakistan

Muzamil then turns to a Dawn opinion piece that used Ireland as a warning: 24% of Ireland’s electricity now goes to data centers. He does not dispute the number. He disputes the analogy.

Ireland, he explains, hosts data centers for Facebook, Google and others largely for tax reasons — billing routes through Dublin so profits do not sit in the United States. “Wo apne logon ke liye nahi service kar rahe hain. Wo poori duniya ke data ko service kar rahe hain,” he says. And critically, Ireland’s grid utilization is already at 100%, so new demand pushes prices up because there is no slack.

Pakistan’s situation is the mirror image. “Pakistan mein is waqt grid utilization one-third hai,” he says. Total generation capacity sits around 45,000 megawatts. Actual usage runs between 15,000 and 25,000. The rest is stranded — not produced, but paid for through capacity payments that then land on ordinary consumers’ bills.

In that context, a single large industrial buyer is not a threat. It is a rescue. It absorbs capacity payments, lets the government avoid the cost of household-by-household metering and collection, and pays world-class rates directly. The unit price for the ordinary consumer, Muzamil argues, should actually fall. This is the point where the imported narrative most obviously inverts on contact with Pakistani numbers.

Exporting electricity as tokens

Muzamil then widens the frame into a strategic argument he clearly cares about. Pakistan cannot export its excess electricity directly. Afghanistan is politically closed off — “hum bhi koi ajeeb kism ke siyane hain,” he says of the fallout. Iran is self-sufficient on cheap energy because of oil. Central Asia already produces cheap hydro and gas power. India is not a viable customer. The other route — using cheap electricity to manufacture exportable products — was, in his framing, undermined when the government “apne sar par kulhada mara” and dollarized industrial power, making Pakistani exports uncompetitive.

That leaves one route he finds genuinely exciting: converting electricity into AI tokens and selling compute globally. Energy, he argues, is becoming the real bottleneck for AI worldwide. American electricity prices are already rising because of data center demand, and projections show them rising further over the next three to four years. Compute has no border problem. If Pakistan can offer cheap compute — cheap because of lower labor costs, lower construction costs, and proximity to Chinese hardware — there is a real business case for exporting it.

He also flags an under-discussed hardware shift. Sky 47 is reportedly partnered with ZTE. Gul Ahmed’s facility is partnered with Huawei. Both are aggressive Chinese plays into a stack Americans dominate through NVIDIA. He mentions in passing that he knows a Pakistani working at ASML in the Netherlands — the sole global supplier of the machines that make advanced chips — and that China last week announced two domestic companies building alternatives. The hardware ecosystem, he suggests, is opening up in ways that could favor countries willing to buy from Beijing rather than through Washington.

The nuclear analogy and the cost of doing nothing

By the end, Muzamil is making a larger argument than a defense of one facility. He compares the moment to Pakistan’s nuclear decision. That program, he says, citing a Carnegie Mellon paper, cost roughly $3.2 billion across its early decades — a very large sum for the 1970s and 80s. “Hum ghas kha lenge lekin hum nuclear bomb banayenge” was the slogan, and the payoff, he argues, is that Pakistan still gets called a nuclear-armed country in every serious international conversation. It bought the country a wildcard.

Sovereign AI, in his framing, is the next such wildcard — and the alternative is renting. “Hum token bhi rent karte hain, hum uske upar model bhi rent karte hain, hum uske upar application layer bhi rent karte hain,” he says. Pakistan is already renting every layer of the AI stack. Without local data centers running open-source Chinese models and selling PKR-priced token packs to Pakistanis, that rental dependency becomes structural — the next petro-dollar, in his phrase, that will keep the country pinned to the ground for another fifty years.

He closes with a request rather than a conclusion. Educated Pakistanis, he says, should not do to data centers what was done to energy policy. They should not import a New York left-wing artistic narrative, apply it to a country with completely different fundamentals, and then wonder in thirty years why they were colonized again — this time by whoever owns the compute. “Data-backed karein,” he insists. If someone wants to oppose data centers, they should show the exact water figure, the exact local damage, the exact grid stress. Absent that, what looks like caution is just imported noise — and Pakistan cannot afford another decade of it.