India’s AI infrastructure debate is beginning to resemble a question of location more than a question of ambition. The country can announce larger computing targets, attract hyperscale investment and expand access to advanced accelerators, but none of those moves changes a basic physical reality: AI workloads eventually have to run somewhere. That “somewhere” needs a dependable electricity connection, a way to remove heat and, depending on the cooling architecture, access to water or an alternative resource strategy.

This creates an unusual tension for India’s AI expansion. Demand for computing can spread rapidly across industries and cities, while the resources that sustain computing remain geographically uneven. A company may want AI capacity close to its customers, but the infrastructure supporting that capacity may work better somewhere else. The result could be an infrastructure market in which geography becomes an increasingly important part of AI strategy. The central question is therefore not simply how much computing India can add. It is whether the country can place that computing where the physical infrastructure can sustain it without turning local constraints into national bottlenecks.

A national AI target can produce very local infrastructure pressure

India’s AI ambitions are national, but the consequences of adding compute capacity are not. Government data released in 2026 showed substantial operational data centre capacity concentrated in Mumbai and Navi Mumbai, followed by Chennai, Bengaluru, Hyderabad and Delhi-NCR. CBRE similarly reported that Mumbai accounted for 53% of India’s data centre capacity at the end of September 2025, with Mumbai, Chennai, Delhi-NCR and Bengaluru together representing nearly 90%. That concentration has commercial logic. Major connectivity routes, enterprise demand and established digital infrastructure naturally attract operators.

AI, however, could make concentration more complicated. Adding another facility to an established market does not necessarily create the same infrastructure effect as adding an equivalent facility in a region with greater electricity availability or different climatic conditions. Two sites with identical IT loads can therefore present very different engineering and infrastructure requirements. That distinction is easy to lose when national capacity becomes the headline metric. India could add thousands of megawatts of data centre capacity and still encounter regional constraints if too much of that growth accumulates in a limited number of locations. The issue is not whether the country has resources in aggregate. It is whether the right resources exist close enough to the infrastructure that needs them.

Water scarcity can change the economics of the same compute cluster

Water introduces another variable that cannot be averaged across India. WRI India reported that more than half of the country’s data centres are located in water-stressed regions, with 75% concentrated across Maharashtra, Tamil Nadu, Karnataka, Telangana and Uttar Pradesh. The figure does not mean that every facility consumes water in the same way. Cooling design, climate, water source and operational practices can materially alter requirements. It does show why a national discussion about data centre water use can obscure the more important issue: local availability. A facility operating in a water-stressed region encounters a different resource environment from an equivalent facility in an area with greater water availability.

This is where AI’s rapid scaling could create an unexpected planning problem. The industry can reduce water intensity per unit of compute, but the absolute demand for compute may increase much faster. Efficiency gains can therefore coexist with growing local resource pressure. The same principle applies to electricity. More efficient chips and cooling systems can reduce the resources required for each unit of computing, but a much larger volume of computing can still increase the total requirement. India will have to manage both sides of that equation simultaneously.

The geography of AI could separate compute from consumption

One of the more consequential changes could come from treating different AI workloads differently. Not every workload needs the same proximity to a user. Inference supporting latency-sensitive applications may benefit from being close to population centers and enterprise networks. Large training jobs, research workloads and other compute-intensive processes can have different latency characteristics and may offer greater flexibility in where they run. That creates an opportunity for India to develop a more distributed AI infrastructure model. Instead of concentrating every form of compute around established data centre markets, developers could place different workloads according to their physical requirements.

Some capacity could remain close to major demand centers, while power-intensive clusters could move toward regions that can support their electrical and cooling needs more comfortably. Such a model would not eliminate the importance of Mumbai, Chennai, Bengaluru or Hyderabad. It would change the reason those locations compete. Connectivity and demand could remain decisive for some workloads, while electricity and resource resilience could become decisive for others. The resulting AI infrastructure map could look considerably less familiar than today’s data centre map.

India can scale AI faster if it stops treating location as an afterthought

India does not face a choice between AI growth and infrastructure realism. The more useful choice is whether the country allows physical constraints to dictate AI expansion reactively or incorporates them into where and how compute gets built. A national AI strategy can set the direction. It cannot make a substation appear where capacity is scarce, create water where a basin is stressed or remove heat without an appropriate cooling system. Those constraints are local, even when the technology they support is global. That is why India’s AI infrastructure opportunity may ultimately depend less on building everywhere and more on building intelligently across different geographies.

The country does not need every region to become an AI data centre hub. It needs the right regions to support the right forms of compute, with enough power, cooling capacity and resource resilience to sustain them. The question facing India’s AI build-out is therefore becoming more precise. It is not simply whether India can build enough computing capacity. It is whether India can build that capacity in the places where the physical systems underneath it can keep up