Rabia Sundho
Pakistan has set its sights on artificial intelligence as a path to higher productivity, new jobs and stronger exports. The National AI Policy 2025, approved by the federal cabinet in July 2025, lays out ambitious targets: training one million AI-skilled professionals by 2030, establishing AI hubs in seven cities, creating a National AI Fund, expanding compute access for universities, and supporting startups and research. Optimistic projections speak of a multi-billion-dollar domestic AI market and a meaningful contribution to GDP growth by the end of the decade. These goals are understandable. Pakistan already has a young population, a growing IT and freelancing base, and IT exports that reached roughly $4.6 billion in fiscal year 2025-26. Freelancers have become a significant part of that story. The policy correctly identifies skills, infrastructure, innovation funding and international partnerships as necessary pillars. Yet one fundamental constraint receives far less attention than it deserves: the reliability and cost of electricity.
Artificial intelligence is not software that runs on good intentions. Training and running modern models, hosting data centers, and supporting continuous research require stable, affordable and high-quality power. Without it, the entire ambition rests on a weak foundation. Pakistan’s power system presents a paradox. Installed generation capacity has grown. By the end of March 2026 it stood at approximately 49,651 megawatts, helped by rapid additions of net-metered solar. Cleaner sources hydel, nuclear and renewable have increased their share of generation. At the same time, the system continues to struggle with low plant utilization, high capacity payments, transmission constraints and a large circular debt stock. Power-sector circular debt stood near Rs 1.675 trillion at the end of June 2026. Thermal plants have often operated well below their potential, with utilization rates in the low-to-mid 40 percent range in recent assessments. Consumers and the federal budget continue to carry the cost of capacity that is not fully used.
This is not a temporary inconvenience. Data centers and serious AI compute facilities need continuous, high-quality electricity. Interruptions, voltage fluctuations and unpredictable tariffs raise operating costs and deter investment. International experience shows that power is one of the largest recurring expenses for data centers. In a country where industrial and commercial users already face high effective tariffs and residual reliability concerns, the economics of local AI infrastructure become difficult. The recent solar boom has further complicated the picture. Distributed solar has reduced demand on the grid during daylight hours, contributing to lower utilization of thermal plants and adding pressure on the financial model of the power sector. While solar is welcome for households and many businesses, large-scale AI compute cannot rely solely on intermittent rooftop generation. It requires firm, dispatchable or well-managed power with storage and grid support.
Some policymakers have spoken of allocating surplus capacity figures in the range of 2,000 megawatts have been mentioned in discussions around data centers and related activities. The competitive electricity market reforms, including wheeling arrangements that allow bulk consumers to buy power directly from generators, offer a potential pathway. These are positive signals. But surplus capacity on paper is not the same as reliable, affordable power delivered to the precise locations where AI infrastructure would be built. Transmission bottlenecks, contractual rigidities and the unresolved circular debt problem still stand in the way.
The human capital side faces a parallel risk. Training one million people in AI tools is a worthy target. Yet if the broader economy cannot absorb higher-value digital work because firms lack reliable power and competitive costs, the result may be more freelancers competing for the same global tasks rather than a deep domestic AI industry. Entry-level opportunities in software and digital services have already shown signs of pressure in some markets as basic tasks are automated. Pakistan needs to move up the value chain into specialized applications, local-language models, sector-specific solutions for agriculture, health and logistics, and higher-end services. That shift requires more than coding boot camps. It requires research compute, quality data, and firms that can operate without constant power anxiety.
The policy framework itself still has implementation gaps. The National AI Council and related institutional machinery have moved more slowly than the original announcements suggested. Funding mechanisms such as the National AI Fund need clear capitalization, transparent allocation rules and measurable outcomes. Compute access for universities and startups remains limited relative to the scale of ambition. Without these pieces in place, the risk is that Pakistan produces policy documents and training certificates faster than it produces usable AI capability.
What, then, would a more realistic path look like?
First, treat reliable power as a core AI infrastructure priority rather than a separate energy-sector problem. This means accelerating the conversation of bilateral contracting and wheeling so that data centers and research facilities can secure long-term, competitively priced power directly from generators, preferably from cleaner or surplus sources. It also means prioritizing grid stability and quality in the locations identified for AI hubs. Second, be selective about compute strategy. Pakistan does not need to compete with the largest global players in training frontier models from scratch. It can focus on fine-tuning existing models for local languages and local problems, on inference and application layers, and on sector-specific tools for agriculture, public service delivery and industry. These uses still require power, but the scale and cost profile are more manageable. Third, link skills training tightly to demand. Mass awareness programs are useful, but the higher priority should be advanced, specialized training tied to actual research centers, startups and industry projects. Quality matters more than headline numbers of people trained. Fourth, resolve the circular debt and capacity payment distortions that keep electricity expensive and the system financially fragile. As long as the power sector remains a major fiscal and operational drag, every knowledge-economy ambition will face an uphill battle. Fifth, measure progress by outcomes rather than announcements. The relevant indicators are not the number of policy documents or the number of people who completed a short course. They are the volume of high-value AI-related exports, the number of local products and services actually deployed, the availability of usable compute for researchers, and measurable productivity gains in priority sectors.
Pakistan has genuine strengths: a large young population, an expanding digital workforce, rising IT exports, and a policy framework that at least recognizes the importance of artificial intelligence. These are not trivial advantages. But ambition without foundations produces disappointment. An AI economy cannot be built on intermittent power, unresolved circular debt and under-utilized capacity. Until electricity becomes reliable and competitively priced for the facilities that AI requires, the gap between the National AI Policy’s vision and the country’s actual capability will remain wide.
The choice is not between having an AI policy and having none. It is between treating power reliability as a peripheral concern and treating it as the indispensable condition for any serious digital transformation. Pakistan will not become an AI economy by declaration. It will become one only when the lights stay on for the machines that make artificial intelligence possible.
The writer is a PhD Scholar in Economics and she can be reached at rabiasundho19@gmail.com
