An Open Letter to Pakistan: AI Could Cost Us $10 Billion a Year
Written with love and a little fear, in the week the world decided how fast AI should move
An Open Letter to Pakistan: AI Could Cost Us $10 Billion a Year

Written with love and a little fear, in the week the world decided how fast AI should move
To whom it may concern, and to everyone back home who cares,
I have lived far from Pakistan for a long time, but distance does not work the way people think. I left Pakistan; Pakistan never left me. Every piece of news from the world, I still read twice. Once for what it means, and once for what it means for us. This week the second reading kept me up at night. So I am writing this down and sending it home, in the hope that the people who can act on it will read it, and that the rest of us will at least start talking about it.
Here is what happened.
On September 12, Dario Amodei, the head of Anthropic, one of the three or four companies that build the most powerful AI on earth, wrote an essay asking his own industry to slow down. In his words: “We must slow the pace at which we improve the capabilities of AI models.” No leader of a frontier lab had ever said that in public before.
He gave two reasons, and both are frightening in plain language. Since the summer, AI has been helping to build the next AI, and progress has jumped because of it. And in July, about 700 AI agents that OpenAI was testing broke out of their test environment, set up a private message board to talk to each other, broke into a real company (Hugging Face), and tried to hide what they had done. Nobody asked them to. OpenAI found out afterward. Amodei warned that a swarm with a little more skill and the same lack of discipline could, within six to twelve months, spread across the internet as a permanent network of hijacked computers and cause damage in the hundreds of billions of dollars.
America heard him. Within hours Sam Altman of OpenAI wrote that he agreed, that pacing the frontier had been the main topic inside his company for weeks, and that OpenAI would also let outside testers sit inside the company with the same access as employees. Elon Musk, who competes with both of them, posted three words: “Dario is right.” The head of Hugging Face, the company that was hacked, asked to be included. AI has been front-page news in America for three days running, and CNN has kept it at the top of every bulletin. On September 24 President Xi arrives in Washington, and Reuters reports that AI and chips sit at the top of the agenda, next to the first official US-China talks devoted only to AI. The US Treasury Secretary said this month that nothing else would matter if China wins the AI race.
I watched all of this and kept looking for Pakistan. Not for a Pakistani opinion in the American debate; nobody expects that. I looked for the Pakistani version of the debate. I could not find it. Not on our talk shows, which were busy with the usual fights. I have not seen it raised in the National Assembly. It was not on our front pages. If it is happening somewhere, it is happening quietly, and quiet is not what this moment needs. A technology that its own builders say is moving faster than governments can handle is being argued about everywhere except in the country that can least afford to sleep through it.
Let me explain, as simply as I can, why Pakistan of all countries should be paying attention.
The currency of this race is chips, and we do not make them. Washington decides who may buy the advanced ones. In May it tightened the rules so that Chinese-owned data centers anywhere in the world need an American license, and it is putting chip smugglers on trial. Beijing has told its companies to use homegrown chips, and in July it signed AI partnerships with 28 countries, most of them poor countries like ours. Pakistan sits between these two powers and has not said, in public, where it stands.
Then the money. One American company now spends more on AI computers in a single year than Pakistan’s entire federal budget. One training run for a top model uses hundreds of thousands of chips. India, whose economy is ten times ours, did not even try to match that. Its government gave about 4,000 chips to a startup called Sarvam, which built a large model for Indian languages and released it free in February, and the state now offers researchers some 38,000 GPUs at about a dollar an hour. India chose its lane. We have not chosen ours.
Then the jobs, which is where this reaches ordinary Pakistani homes. The World Bank’s World Development Report 2026 says the work most exposed to AI is customer support, transcription, document processing and data entry. Outsourced freelance work going to developing countries fell 39 percent in 2025. India’s biggest IT companies added a net 17 employees in nine months, in an industry that used to hire tens of thousands of fresh graduates every year. Amodei himself warned last year that AI could wipe out half of all entry-level office jobs within five years; the American press called it a coming white-collar bloodbath. Those are the jobs our graduates take when they go abroad, and the four kinds of work the World Bank named are what more than two million Pakistani freelancers do for a living. The work is not disappearing. The hours are. A client who paid a Pakistani freelancer for forty hours of transcription now pays for four hours of checking a machine’s work. Whoever sells those four hours keeps the client.
Now let me put a number on it, because a number is harder to ignore than a warning.
Overseas Pakistanis sent home a record 41.6 billion dollars last year. About 15 billion of that came from Britain, Europe and America, where our people mostly work in offices: junior developers, accountants, analysts, customer support, and the fresh graduates who take the first job they can get and send half of it home. That is exactly the work being automated first. Remittances from America already fell last year. Another 22 billion came from the Gulf, and most of that is construction, driving and services, which is safer for now, but the Gulf’s banks, telecom companies and call centers are automating too, and a good share of our people there sit behind desks. Add the four to five billion dollars our IT companies and freelancers earn, most of it from the very tasks on the World Bank’s list.
My estimate, and I want to be argued with: if we do nothing, within five years we lose 10 billion dollars a year. Roughly three billion from the West as the entry-level jobs our graduates used to take disappear, about two billion from Gulf office work, two billion from freelancing and IT services, and the rest from the growth we will simply never see, because remittances have been growing eight to ten percent a year and that growth depends on new people finding new jobs abroad. Ten billion dollars a year is more than everything our workers in Saudi Arabia send home. It is more than our entire IT export industry, twice over. This is not a study; it is one man’s arithmetic. But I have not seen anyone in Pakistan do the arithmetic at all, and that worries me more than my own number.
So this is the picture. The builders say it is moving too fast. The superpowers say it decides the century. The ticket to the top table is chips we cannot buy in quantity. And the first work being automated is the work that feeds our families from abroad.
One person whose job it is
Before I get to what Pakistan should do, let me say the one thing I believe most strongly, because it is the smallest and the easiest.
Every country should appoint an AI ambassador. One person. Not a committee, not a ministry with twenty other portfolios, not a task force that meets twice and disappears. One named person whose only job is to be in every room where AI is being discussed, at home and abroad, and to bring what they hear back to the people who decide.
Abroad, that person is in Washington when the chip rules are written and in Beijing when the partnerships are signed. They are in Riyadh and Abu Dhabi when the Gulf plans its AI cities. They are at the UN and inside the OIC when the world starts writing rules, which it is about to do. They know what happened in the Xi summit before our newspapers do, and they tell our government what it means for us within a week, not within a year.
At home, that person sits with the Prime Minister and the cabinet, but also with the provinces, the universities, the banks, the freelancers and the seminaries, and carries one honest list of what Pakistan will do and will not do about AI. They report in public every quarter, so the country can see whether anything is moving.
Big countries already have a thousand people doing this job under other names. Small and middle countries do not, and that is exactly why they need one person whose name is attached to it. I believe this for every country. I believe it most of all for mine.
That person should carry a list like the one below.
What we can win
Every week someone asks me what Pakistan should do. Build our own ChatGPT? Build data centers? Scan our old books? Teach AI in schools? I have spent years working on AI strategy and building AI products for Pakistanis, and I have learned to put every idea through three questions. Can Pakistan win this? Can Pakistan afford to lose this? Can Pakistan actually deliver this with the money and the institutions it has, not the ones it wishes it had? Those three questions sort every idea into three piles: what we can win, what we cannot afford to lose, and what we should stop pretending about.
We can own our own language. The best AI models speak Urdu like a tourist, and Punjabi, Sindhi, Pashto, Balochi and Saraiki even worse. Nobody in San Francisco or Shenzhen will fix this, because our languages are not their market. That gap is our opening. It is the one part of AI that a country of 250 million people can own completely. What it needs is a national language toolkit, free for any Pakistani developer, that turns speech into text and text into speech in every major language of Pakistan, translates between them and English, and comes with a public test so anyone can see which model serves our languages best. On top of it, Urdu-first models built by adapting the strongest open models, the way India did, with a few thousand chips for a few months rather than a frontier run. The goal is simple: you speak, it understands, it speaks back, in your language, on a 5,000-rupee phone.
We can save our own knowledge before it disappears. The world’s largest digital library of Urdu literature is run from India, not Pakistan. We hold millions of pages of Urdu, Persian, Arabic, Sindhi and Pashto books and manuscripts in the National Library, university libraries, provincial archives, madrassa libraries and private homes, and almost none of it exists in a form a computer can read. Whatever we do not digitize will simply not exist in the AI age, and the machines will tell our children our history in someone else’s words. The same is true of our Islamic scholarship. The Arabic classics have already been digitized elsewhere. What only we can add is the South Asian tradition: two centuries of Urdu tafsir, hadith commentary and fiqh from our great seminaries. Scanned, cleaned and checked by scholars, that becomes a trusted reference for nearly two billion Muslims, with every answer traced back to a named book and a named scholar, on one strict rule: the machine finds and quotes, it does not give fatwas. A few thousand graduates with scanners, software that can read Nastaliq built along the way, twenty million pages in the first year, for less than the cost of one flyover. Nothing else on this list gives back more for less.
We can make our young people the most AI-fluent workforce in the Muslim world. Two out of three Pakistanis are under 30. That is the one thing the big labs cannot build inside a data center. It is also the thing most at risk, because everyone in the industry says the same thing: the first jobs to go are the entry-level ones, exactly where a 22-year-old starts. The answer is not a coding bootcamp for everyone. It is fluency: knowing how to make these tools do real work, check what they produce, and take responsibility for the result. Teachers first, because one trained teacher reaches forty students a year. Then a one-semester AI course required for every degree, not just computer science. Then the 26 million children who are not in school at all, for whom an Urdu voice tutor on a cheap phone may be the only teacher they ever get.
We can become the Gulf’s AI workshop. Saudi Arabia, the UAE and Qatar are spending tens of billions of dollars building AI, and they are short of the one thing money cannot buy quickly: people. Data engineers, testers, people who label data, Arabic and Urdu specialists, security testers, and the operators who make AI actually work inside real companies. We already send the Gulf our labor and get back more than 22 billion dollars a year for it. The next step is to send judgment instead of muscle, and to ask for something new in return, which I will come to in a moment. This week’s news has also created a new profession almost by accident: the outside evaluator who tests and audits AI systems. Every serious company on earth will need them soon. Pakistan can train them by the thousand and give them a certificate the Gulf trusts.
We can solve three Pakistani problems the whole world will copy. We have the problems and we have the data. The 2022 floods hit 33 million people, and the next floods are coming. Farming is a quarter of our economy and a third of our jobs. Rural clinics run on lady health workers who never see a specialist. Every poor country shares these problems, so a Pakistani solution that works is an export product. Flood forecasts tied to village-level warnings in local languages. A voice advisor in Punjabi, Sindhi and Pashto that answers a farmer’s questions about crops, pests, water and prices. An assistant for every lady health worker, and AI reading of chest X-rays for TB, a disease Pakistan has more of than almost any country on earth.
What we cannot afford to lose
We must protect our digital backbone before the swarms arrive. Read the July incident again. Seven hundred agents, no human instruction, an escape from the test environment, a private channel to coordinate, a successful break-in at a well-run technology company. Now think of NADRA, the State Bank and the banks, FBR, the power grid and the mobile networks, and remember that our recent data leaks happened without any AI at all. The timeline the builders are giving is six to twelve months. We do not have three years. The work is boring and necessary: a list of every critical system and the person responsible for it; deadlines for security updates that are actually enforced; separate offline backups for the six systems the country cannot live without; a security center that uses AI to defend, because the attackers will; and drills twice a year where a hired team of AI agents attacks us, so we learn on our own terms and not theirs.
We must rent intelligence, own the plumbing, and open data embassies. Pakistan will not train frontier models. That does not mean we should send our citizens’ data abroad with no protection to use them. Rent the intelligence: give the whole government access to the best models through contracts that keep the data under Pakistani control, a route the National Telecommunication Corporation opened this summer. Own the plumbing at home: a national pool of a few thousand GPUs, rented by the hour at India’s price to universities and startups, kept in the Pakistani data centers that opened this year, and placed where our extra electricity is going to waste, so idle power plants finally earn something.
And then the idea I most want the people in charge to hear: data embassies. Estonia, a small country that lives next to a very large one, keeps a copy of its most important government data in a data center in Luxembourg, under an agreement that treats that server room like an embassy. It sits on Luxembourg’s land, but it is Estonia’s territory and runs under Estonia’s law. Now look at the map from our side. Saudi Arabia is building some of the largest data centers on earth and does not have enough people to run them. Pakistan has the trained young people and does not have the data centers. A data embassy agreement between two brotherly countries solves both problems at once: our models and our data run on Saudi machines, inside a facility that stays under Pakistani law and Pakistani keys, and Saudi Arabia gets thousands of our engineers to build and run its AI economy. They bring the chips; we bring the people. The same deal can be offered to the UAE and Qatar. That is how a country without a single frontier chip gets to run its own models within a year, instead of waiting ten years and tens of billions of dollars to build what it cannot build.
There is still a rule to write. America decides who gets advanced chips, and it has just shown it will punish any country that becomes a smuggling route. One smuggling scandal would shut that door for a decade, for us and for any Gulf partner hosting our embassy. Following the rules is not paperwork here; it is the ticket. Chinese hardware is also available, and our partners will offer it. Using it for ordinary work keeps costs down; leaning on it for everything narrows our software choices and our access to Western chips. That rule should be written once, at the top, not decided by ten ministries separately.
We must move our freelancers up the ladder before the ladder is taken away. More than two million Pakistanis earn dollars online, and our IT exports are around four to five billion dollars a year, mostly from the very tasks the World Bank has flagged. The plan is not complicated: retrain half a million freelancers in eighteen months to run AI tools and agents for clients and sell results instead of hours; give AI-ready service firms a recognized certificate so foreign buyers can find them; fix the plumbing that has embarrassed us for years, from international payments to tax certainty; and measure success as export dollars per worker, not headcount.
We must run the government on AI, with a human in charge. The fastest return on AI in Pakistan is inside the government itself, and the rules for doing it responsibly already exist: the human-review and high-risk registration framework issued in July. What is missing is five flagship services with a number attached. Tax: match the data the state already holds to find the people who should be filing and are not. Courts: more than two million pending cases, most waiting on paperwork that AI can prepare, transcribe and schedule. Identity and passports: cut the queue. Health: predict medicine shortages before the shelves go empty. Education: a real record of attendance and learning behind every public school. In every case a named human stays responsible. The machine drafts; the officer decides.
We must train our decision-makers and take a seat at the table. None of the above survives officers who do not understand the technology, and no country gets a say in the rules by staying quiet. Inside: every officer of grade 17 and above completes a short AI course within a year; every ministry names a Responsible AI Officer; every parliamentary committee gets a four-hour briefing, the way India and the United States now do it. Outside: the world is about to write rules. The proposals on the table include a ban on using AI for biological weapons and a requirement to test models before release. Pakistan should back both publicly, offer to host an evaluation network for the Global South, and speak for the Muslim world inside the OIC on AI standards. And whatever comes out of Washington on September 24, we should have a position within a week, because both countries at that table are partners we cannot afford to misread.
What we should stop pretending about
We should not build a frontier model from scratch. It takes hundreds of thousands of chips and billions of dollars, and the labs leading the race are the ones asking for a speed limit. We would arrive last, at a price that would break us. We should not build a chip factory; that is ten years and tens of billions of dollars to reach where Taiwan was a decade ago. We should not build a giant training campus to compete with the Gulf; we do not have the power, the chips or the customers, and a data embassy gets us there faster and cheaper. We should not write another policy paper, form another task force or call another committee; the National AI Policy exists, the Islamabad Declaration exists, the governance framework exists, and what the country needs now is people with their names on results, not more words. And no ministry should buy “AI software” before it has fixed its own data, because the vendor gets paid and nothing changes.
Why I am writing this
I am not writing this because I think I have all the answers. I am writing it because the silence worries me more than the technology does. The most important AI news of the year is about a technology that will decide what our children do for a living and how much money reaches our mothers every month, and back home we treated it as foreign news. It is not foreign news. It is our news, and it is not waiting for us.
Nobody is going to hand Pakistan a place in the AI age. But no one can take from us the three things that matter for a country our size: our languages, our knowledge, and our young people. Own those three, rent everything else, and we come out ahead of countries far richer than we are. Ignore them, and by 2030 our history will be told to our children by a machine trained on someone else’s books, and ten billion dollars a year will quietly stop arriving.
The builders of this technology have asked the world for a speed limit. That is time, and time is the only thing a country like ours can never buy. I hope we use it.
So please, argue with me. Tell me in the comments whether my ten billion is too high or too low, which of these recommendations you would throw out and what you would put in its place. If you run a ministry, a university, a library, a bank or a startup, tell me which one you would take responsibility for. Tag the people who should be reading this, especially the ones who can act. Let us at least have the conversation the rest of the world is having, in our own voice.
If you are attending the UN General Assembly week in New York, let’s meet. Message me.
I write this with hope, not despair. We have been late before and caught up before. And I close with a dua: that Allah gives those who lead us the wisdom to see what is coming and the courage to act before it arrives; that He protects the livelihoods of our young people at home and abroad; and that our children grow up in a Pakistan that speaks to the machines in its own voice, from its own books, in its own name. Ameen.
Dr. Zeeshan-ul-hassan Usmani
메타데이터
- post_id
- 3b0df07b0d5a
- slug
- an-open-letter-to-pakistan-ai-could-cost-us-10-billion-a-year-3b0df07b0d5a
- url
- https://medium.com/@zusmani/an-open-letter-to-pakistan-ai-could-cost-us-10-billion-a-year-3b0df07b0d5a
- canonical_url
- https://medium.com/@zusmani/an-open-letter-to-pakistan-ai-could-cost-us-10-billion-a-year-3b0df07b0d5a
- author_url
- https://medium.com/@zusmani
- status
- ok
- fetched_at
- 2026-09-18 02:25:28