Artificial intelligence (AI) is being viewed primarily as a computing and software challenge. That is understandable, given the race for models, graphics processing units (GPUs), talent, and applications. But the next constraint on AI may be much less virtual: electricity, transmission, cooling, water, and connectivity. That makes the case for an infrastructure policy for AI, much like power became a matter of long-term planning when industrialisation made reliable electricity indispensable. The scale of the coming requirement is already visible.
Data-centre capacity has risen from 375 megawatts (Mw) in 2020 to about 1,575 Mw now. The Central Electricity Authority estimates that demand for power from data centres could reach 17 gigawatts (Gw) by 2031-32. The number may be small relative to the country’s total installed generation capacity, but that misses the point: AI loads are large, concentrated, and increasingly continuous. A few large campuses can throw up significant local requirements for generation, transmission, and cooling.
To be sure, some thought has gone into planning for power requirement. Installed generation capacity is around 524 Gw, and the national transmission system is being expanded alongside renewable capacity. AI and large data-centre demand has been factored into the estimates. The question, therefore, is not whether another layer of government planning is needed. It is whether the existing pieces are being brought together quickly enough. An AI infrastructure policy could identify locations where large compute clusters are likely to emerge and align transmission, generation, fibre connectivity, and industrial infrastructure accordingly. That would be preferable to a model in which data centre operators choose locations independently and the supporting infrastructure catches up later.
The physical footprint of AI goes beyond electricity. High-density computing produces substantially more heat than conventional workloads, making cooling a critical consideration. The industry is already moving towards direct-to-chip liquid cooling, immersion cooling, and closed-loop systems to reduce water use. Standards for power usage effectiveness, carbon usage, cooling efficiency, and water usage are already being developed. These efforts are welcome, but standards alone will not resolve the question of where the next generation of data centres should be located. Water availability, grid capacity, renewable power access, and transmission costs should become part of that decision. Nor should the policy be confined to today’s data centres. The IndiaAI Mission has already onboarded more than 38,000 GPUs through 14 providers, offering subsidised compute to startups, researchers, and others. As compute becomes cheaper and more widely available, demand for the physical infrastructure underneath it will only broaden.
The case for an AI infrastructure policy, then, is not an argument for subsidies or about the state picking winners among data centre operators. It is about anticipating infrastructure requirements before they become a bottleneck. Such a framework could integrate AI policy with power, transmission, renewable energy, telecom, and water policy; set predictable standards for energy and water efficiency; encourage round-the-clock clean power and storage; and provide greater clarity on where large-scale compute can be built without creating new infrastructure stress. The recent move to strengthen the nuclear framework and explore the use of small modular and micro reactors for emerging sectors such as AI and data centres shows how closely the energy and AI questions are beginning to intersect. The larger lesson is simple. AI may be a digital technology, but its economic potential will increasingly depend on physical infrastructure. The race for compute should, therefore, not be allowed to run ahead of the infrastructure needed to power, cool, and connect it.