Pakistan can build sovereign AI without a Saudi chequebook

The claim Muzamil is defending

Muzamil opens by returning to a thread from his previous two episodes: Pakistan pays in dollars for what it lacks — electricity, gas, oil, petrol — and because dollars are finite, those things stay permanently rationed to a thin slice of society. He extends the argument to intelligence itself. If AI is about to become a critical input in the economy and Pakistan rents it from abroad, then intelligence will remain a scarce, dollar-priced luxury too.

He acknowledges the pushback head-on. “जिस मुल्क के अंदर बिजली के बिल देने के पैसे नहीं है, उसके अंदर आप क्या बातें कर रहे हैं एआई की?” — a country that can’t pay its power bills has no business talking about AI. Most comments on the previous videos were exactly this cynical. The purpose of this episode, Muzamil says, is to defend the claim with data and to show that “आधी दुनिया इस वक्त वही कर रही है जो मैं आप लोगों को या पाकिस्तान को करने का बोल रहा हूं” — half the world is already doing what he is asking Pakistan to do.

What sovereign AI actually means

Muzamil gives the term a precise definition: “Sovereign AI is a nation’s ability to produce its own intelligence” — to develop, deploy and govern it using its own infrastructure, its own data, its own models and its own people. Local doesn’t just mean built inside Pakistan. It means Pakistanis building it, running it and owning every layer.

He traces the phrase back to Jensen Huang, who popularised it in a 2023 Dubai talk. Muzamil quotes the argument directly: “Every country in the world needs to own the production of their own intelligence. You cannot allow that to be done by other people.” The reasoning, as Muzamil unpacks it, is threefold. First, if you hand your data, literature and civilisational knowledge base to a foreign model, that model “in a way चोरी कर रहा है” — steals it, then sells it back to you wrapped as a product. Second, the foreign model neither understands your culture nor renders it faithfully. Third, you end up buying your own knowledge back in dollars, in token form, forever.

He notes that Satya Nadella and the CEO of Palantir have since made similar arguments. Sovereign AI has moved from a Jensen Huang keynote line into the operating assumption of serious institutions.

The countries already building this

Muzamil walks through who is actually doing it. India has committed over a billion dollars through its IndiaAI Mission, deployed 34,000 GPUs, and backed Sarvam AI to build a foundational model — most likely by forking an open-weight base like DeepSeek and training it on Indian languages and knowledge. The UAE built Falcon, which at launch was the largest and most capable open-weight model in the world at roughly 70 billion parameters, and has committed to five gigawatts of AI data centre capacity, with one gigawatt already commissioned.

Muzamil sharpens the comparison: “पाकिस्तान 250 मिलियन की कंट्री है। हम 8.5 मेगावाट ऑफ कंप्यूट या डेटा सेंटर के ऊपर हम शोर डाल रहे हैं” — Pakistan, a country of 250 million, is panicking over 8.5 megawatts, while the UAE with roughly a million locals is building five gigawatts.

Saudi Arabia’s Humain, owned by the Public Investment Fund and headed by the Crown Prince, is targeting two gigawatts by 2030 at an estimated cost of 77 billion dollars, with a first order of 18,000 cutting-edge Nvidia chips already placed. The EU has committed 20 billion euros for five “AI gigafactories,” and Muzamil highlights their justification: “to enable sovereign access to computing and reduce dependence on non-European suppliers.” His line lands: “Even the richest club of nations on Earth is scared of renting.” Japan is putting roughly 65 billion dollars into AI and chips. France is backing Mistral. The pattern, Muzamil argues, is universal among serious countries.

Why Indonesia is the right template

Muzamil pivots to the comparison he considers most useful. The Gulf and European examples are apples to oranges — those countries have chequebooks Pakistan doesn’t. Indonesia is different. Similar population, similar cultural weight, similar history of military takeovers and corruption, roughly fifteen years ahead in political-economic evolution but still a developing nation with familiar problems.

Indonesia did not announce a fifty-billion-dollar programme. It built Sahabat AI, a model trained in Bahasa Indonesia and regional languages including Javanese, Sundanese and Balinese. On the infrastructure side, telecom operator Indosat and tech company GoTo have partnered with Nvidia to build the local ecosystem — “slow and steady,” structurally, without the government sitting at the centre.

Muzamil argues this is directly emulable. Pakistan doesn’t have a Saudi chequebook, but it has Jazz — “एक बहुत बड़ा कंग्लमरेट बन चुका है” — and it has large software players who understand the industry and have made real money. He also mentions meeting a Pakistani engineer building local-language AI models for the Malaysian government, and notes the irony that Pakistanis are building these systems for Malaysia but not for Pakistan.

On the local-language point Muzamil becomes most direct. If Urdu, Punjabi, Saraiki, Pashto and Balochi speakers aren’t given the same AI interactions available in English, “हमारे लोग बैठे रह जाएंगे और फिर 20 30 साल जो है ना पीछे बैठे रहेंगे.” He draws the parallel to Google: the world’s knowledge sits there in English, and Pakistanis who can’t read English end up on TikTok instead. He gives a shoutout to a Pakistani startup working on local-language AI, but notes it’s a difficult place to operate without wider industry and narrative support.

The three-layer stack, and where Pakistan stands

Muzamil then teaches the architecture. The AI stack is a pyramid. At the base sits hardware — Nvidia chips, RAM, storage, the physical data centre. In the middle sit models — the LLMs, whether Claude, ChatGPT or the flood of open-weight releases. He notes with genuine surprise that Qwen 3.8 is now delivering the same quality Opus 4.6 was six months ago, and can run on a 64GB MacBook. At the top sits the application layer — the software you actually interact with.

Mapped onto Pakistan: the application layer, the country fully owns. “पाकिस्तान का वर्ल्ड क्लास टैलेंट है। इसके ऊपर कोई डाउट ही नहीं है.” Software houses in every city, a strong freelancing ecosystem. The model layer, Pakistan does not yet own — but because China played “the open-weight game” for geopolitical reasons, the benefit accrued to poorer countries. You can now fork a Chinese model, train it on your own data and language, and produce a sovereign version without reinventing the wheel.

At Atlanta Tech Week, Muzamil met a Georgia Tech-incubated company that takes open-weight models in the 50-to-100-billion-parameter range and distils them down to focused 3-to-4-billion-parameter models for specific organisations. The hardware layer is where the real investment must go — the foundation everyone else builds on. Miss any one layer, Muzamil warns, and you’re still paying dollars in rent somewhere.

The electricity myth

The most common objection Muzamil faces is electricity. Load shedding, no power, how can you run data centres? He is emphatic: “The reason why people say that is because they don’t understand the problem of electricity to begin with.”

Pakistan’s installed capacity is roughly 46 gigawatts. Peak summer demand is around 20 gigawatts — less than half of capacity. Winter demand sometimes falls to 7,000 megawatts, one-seventh of total capacity. The problem is not shortage. The problem is capacity payments — the dollar-indexed take-or-pay contracts signed with IPPs, front-loaded into a ten-year debt repayment schedule. Muzamil delivers the number: 2 to 2.5 trillion rupees a year in capacity payments, roughly 7 billion dollars, for plants Pakistan doesn’t fully use. “क्या खूबसूरत दुनिया है। क्या जबरदस्त मोयदा है.”

The solution, he argues, isn’t to shut plants. It’s to add demand. A data centre is the ideal customer — a single buyer, 24/7 offtake, high margins, no distribution cost, no recovery risk. He addresses the inference-versus-training economics: yes, AI companies are losing money, but that’s on training. “इनफेंस के ऊपर इवन क्लॉट पैसे बना रहा है। इनफेंस के ऊपर चैट जीबीटी पैसे बना रहे हैं.”

He credits the government for allocating 2,000 megawatts to AI data centres, but pushes back hard on rumours that this power will be subsidised. “You will do the same rent-seeking बेवकूफी again.” The right guarantee is uninterrupted supply, not a discount. He also points out that the peak of the capacity payment crunch will taper off after 2027 as CPEC-era ten-year debt schedules complete, and that Pakistan has already deployed 27 to 28 gigawatts of rooftop solar in the last three to four years — capacity currently going to waste, but likely storable within three to four years as sodium-ion and iron-based batteries mature globally.

The chip cost problem, and why timing matters

The second objection is hardware cost. Roughly 50% of the total cost of a data centre is chips, RAM and storage — and right now, thanks to global demand, all three are extraordinarily expensive. Muzamil concedes the point completely. Pakistan cannot afford to build a one-gigawatt data centre today. Doing so at peak prices would guarantee a loss when the market normalises.

But he argues an AI bubble is already forming, and analysts widely expect it to pop, just as the Cisco-era internet hardware bubble did. Prices will fall. Moore’s Law — technology halving in price roughly every two years — still applies. China is entering chip manufacturing the same way it entered solar panels and EVs: not necessarily by matching ASML’s cutting edge, but by producing something 95% as good at half the cost, or by turning one supplier into fifteen. Huawei’s Ascend chips are already in the market. Muzamil notes that the Gul Ahmed 250-million-dollar data centre is being built with Huawei, and the Sky47 project with ZTE. Pakistan is already partnering with Chinese suppliers.

His answer on the dollar-import question is honest. Yes, chips will always involve some dollar cost. But there is a fundamental difference between renting tokens — used once, gone — and buying an asset that produces value for three to five years. He draws the parallel to mobile phones: Pakistan used to import fully-assembled handsets at $200 each; now it imports roughly $50 of components and assembles locally, and total phones produced annually has grown exponentially without a proportional dollar outflow. The same logic applies to AI hardware.

Why the government should stay out

By the end of the conversation, Muzamil is unusually blunt about the role of the state. “Personally I do not support government getting into anything, especially Pakistani government getting into anything, because they just खराब कर देते हैं दिमाग.” The government’s job is to provide an enabling environment, a level playing field, justice and infrastructure. Not to build data centres.

The right builders, he argues, are Pakistan’s IT services companies — Systems Limited, NetSol and their peers. He is direct about the pressure they’re already under. Remote work broke the staff-augmentation arbitrage after Covid, where a middleman charged clients $5,000-$6,000 and paid the developer $500 in rupees. Now clients pay developers $2,000-$2,500 directly. AI is now automating the low-end augmentation work itself. These companies need to move to products and invest in the infrastructure layer. “If AI is replacing you, then you need to go and own the stack that runs the AI.”

Muzamil estimates that Pakistan’s stated IT services revenue of $5 billion is actually closer to $10 billion, with the difference parked in Dubai, Singapore and the US. “This is not a hidden fact. Anyone who’s in the IT sector knows this.” If a smart investment banker structured a $500 million hedge fund with a $500,000 minimum entry, seeded by thousands of services-industry operators who genuinely have that money stashed, a data centre becomes buildable without a single dollar of government money.

He closes by distinguishing this from nuclear. The state built the bomb because security concerns and single-buyer economics demanded it. Sovereign AI is just as strategically important, but its buyers, investors and ecosystem sit all over the world. It is a private-sector job. And, Muzamil insists, it is not charity — it is a real business opportunity for anyone willing to do the diligence today for what the market will look like in three to five years. The numbers, he promises, are coming in the next episode.