Pakistan needs 5.5 gigawatts of AI compute in 10 years

The raw numbers of the global compute race

Muzamil opens with the current stock of AI compute capacity around the world. America sits at 25 gigawatts. China at 32 gigawatts. India, the neighbour, at 1.5 gigawatts and building fast. Pakistan, by contrast, has roughly 8.5 megawatts — small enough that it does not register on the same chart.

The 2030 projections widen the gap. America is expected to reach 80 gigawatts, China around 60 in the American analysts’ estimate (Muzamil suspects the real Chinese number will be closer to 100–120 because “American analysts always try to show China a bit lower”), and India 5 gigawatts. Data centers already consume more electricity than all industrialised nations combined, and that consumption is set to double in the next five years.

This is where Muzamil sets the frame. “It is not business as usual,” he says. “It is not like the last 40 years where the world just kept running the same way.” He points to the first industrial revolution — the one that took the subcontinent from 30% of global GDP to 0.3% by 1947 — and warns that whoever builds this infrastructure now becomes the coloniser of tomorrow.

Per capita is the honest measure

Raw gigawatts hide the real story. Muzamil pushes the conversation to per-capita compute, which he calls the honest measure of how much intelligence a country is producing per citizen.

America has around 158 megawatts of data center capacity per million people. China, despite the larger raw total, has only 22 megawatts per million because of its population. Europe sits at 27. Pakistan is at 0.04 megawatts per million — three hundredths of a single megawatt.

“America has roughly 800 times more computing power per citizen than we do,” Muzamil says. “800 times. Remember this. This is not 8 times, not 80 times. 800 times.” He connects the number to GDP per capita — roughly $2,000 in Pakistan against $80,000 in America — and argues the compute gap is what will let that ratio compound further. “That is what should scare you.”

The 5.5 gigawatt benchmark

Muzamil then sets a specific target. He rules out matching America — “they may be a century ahead of us” — and picks China’s current per-capita figure as the north star. Applied to Pakistan’s population, that means 5.5 gigawatts of data center capacity, built over ten years.

The math is deliberate. At today’s frontier prices, using the latest closed-loop cooling and Nvidia hardware, one gigawatt of AI data center costs around $40 billion. Almost 50% of that is hardware; the rest is building, land, and labour. McKinsey projects the world will spend $7 trillion over the next four years on AI compute build-outs. Saudi Arabia alone is putting $77 billion into data centers by 2030.

Applying American prices, 5.5 gigawatts comes to roughly $220 billion — more than half of Pakistan’s $400 billion annual GDP. Muzamil is upfront that this “doesn’t make sense” as a lump sum. But he argues three things bring the real number down sharply: staged compounding deployment across a decade, year-on-year declines in compute prices, and the fact that Pakistan’s construction and labour costs are a fraction of America’s. His revised estimate lands closer to $20 billion over ten years.

What the ten-year deployment curve looks like

Muzamil walks through the shape of the build-out year by year. Year one: 50 megawatts, which he notes is the full campus limit of the already-announced Sky 47 project. Gul Ahmed and two or three others have also announced builds. At international prices, 50 megawatts needs roughly $1 billion of investment — a figure he argues is well within reach of Pakistan’s private capital.

Year two adds another 100 megawatts. Year three, 200. Each year roughly doubles the previous year’s addition. By year five, cumulative capacity crosses one gigawatt. By year seven, two gigawatts. By year ten, the full 5.5 gigawatts.

“This kind of plan is what the government should ideally be giving,” Muzamil says, “but they don’t give it. So at least have a graph in your head — year one target, year two target, year three target — so you know if you’re on the right track.”

The compounding matters for more than capex. Pakistan’s IT services already bring in about $5 billion annually on the books, which Muzamil estimates is closer to $10–12 billion in reality. Goldman Sachs projects India’s services industry will grow from $400 billion today to $900 billion by 2030 — a curve Pakistan cannot join without its own compute layer. Muzamil concedes AI may kill bottom-of-the-pyramid freelance work, but argues top operators can extract far more value if they own the infrastructure.

The money is already there — it is just being burned

This is the sharpest section of the episode. Muzamil argues Pakistan does not have a capital problem. It has a capital allocation problem.

He runs the counter-ledger. Pakistan imports $6 billion of oil every year — $60 billion over a decade, $40 billion more than his entire data center plan. Another $4 billion a year goes to palm oil imports ($40 billion per decade), which he describes bluntly: “technically speaking it is not that essential for staying alive, but our people enjoy drowning food in oil.” Cars: around $2.5 billion a year, $25 billion per decade. Phones: $1 billion. Non-essential luxury imports the government itself flags: $6 billion a year. Capital flight to Dubai and elsewhere: another $6–8 billion a year.

“It is not a matter of capital not being there,” Muzamil says. “It is a matter of us putting that capital, with some intelligence, toward something that gives us long-term benefit — instead of just consuming.”

He draws the distinction plainly: consumption is what you eat, burn, and restart the next morning. Investment compounds. A sensible country facing Pakistan’s poverty levels, he argues, would spend $2.5 billion a year on public transport instead of importing cars to please a few hundred families.

A direct message to Pakistan’s IT operators

Muzamil then turns to a specific audience: Pakistan’s IT company owners parking dollars abroad and holding rupees domestically. He warns their trajectory mirrors the textile sector’s.

Today, he says, IT founders earn in dollars, pay their Pakistani employees a small share, and keep the rest offshore. But those employees will increasingly be replaced by AI compute the founders themselves are already buying — Cursor subscriptions, cloud accounts, agent infrastructure. Globally, cutting-edge companies are already spending $2,500 to $5,000 per employee per year on AI tools.

“You will earn one dollar,” Muzamil says, “and 50 to 70 cents will go straight back out on buttons and cotton and zippers — just like the textile guys. That is the business you will be running.”

His alternative: use the offshore capital to buy hardware, bring it into Pakistan, build the data center, use the tokens yourself — paying yourself in dollars while spending rupees on token consumption — and sell the excess capacity to other local and international service providers. “For that you need vision. You need a bit of risk appetite. You need to focus on a long-term horizon.”

AI sovereignty as the new nuclear bomb

Muzamil closes by reframing the entire four-part series. He acknowledges he has been “screaming about this for four days straight” and that this is the final video in the arc. He wants the closing frame to land.

He references a recent conversation where someone told him Pakistan only exists as a functioning state because of its nuclear bomb. He does not fully endorse the claim, but uses it as a pivot. “If you believe the nuclear bomb was existential for our survival,” he says, “then I am telling you today that the sovereignty of your AI stack and the sovereignty of your intelligence is as important as a new nuclear bomb.”

He points to the historical pattern. Every industrial revolution produces two kinds of countries — those that build the machinery and those that rent it. “Renters are always servants. They always exist as slave-style subjugated vassal states. The builders control the world.” Fifty years ago, he adds, Pakistanis ate grass to build a bomb. Today the country is materially better off and the ask is smaller — redirect some consumption spending into compounding infrastructure.

Muzamil ends with an editorial request: engage in the comment section with research, not reflex. “If I am wrong, go prove me wrong. Do the research, come back with facts, counter me properly so I also learn and grow.” The point of the series, he suggests, was never to deliver conclusions — it was to force a public that spends its attention on Uber controversies and Instagram outages to spend some of it on the infrastructure question that will decide whether Pakistan builds or rents its next century.