The city pays, the value leaves: the economics of India’s data-centre boom

The states competing to attract AI infrastructure are waiving taxes on its heaviest input at the very
moment that input becomes the fastest-growing load on the grid. The incentives do not add up.

On 29 July, the government told Parliament that AI data centres would add 26.3 GW to India’s power demand by 2031-32. This was nearly double the 13.56 GW it had projected only in March. In four months, the official estimate of how much electricity this single industry will draw from the grid had doubled.

That revision should change how we read the data-centre build-out now being celebrated as India’s entry into the AI economy. The states competing to host these facilities are, at the same time, waiving taxes and discounting tariffs on the very input that this projection measures.

A Capital-Heavy Bet With a Light Footprint

The scale of the build-out is not in question. Wood Mackenzie expects India’s operational data-centre capacity to rise more than fivefold, from 2.2 GW in 2025 to 12 GW by 2030, with AI-dedicated capacity rising from 275 MW to over 6,500 MW.

Investment is following. Google is building a $15-billion hub in Visakhapatnam, Blackstone-backed AirTrunk has committed around ?3 trillion by 2030, Meta is leasing a 168-MW facility from Reliance in Jamnagar, and Microsoft is investing $3 billion. The Centre is adding to this through the ?10,300-crore IndiaAI Mission, which subsidises GPUs and shared computing capacity.

This is, arguably, the largest industrial commitment India has made in a generation. It is also one of the least labour-intensive.

A hyperscale data centre the size of a stadium employs a few hundred people, whereas an older IT campus employed tens of thousands. The Institute for Energy Economics and Financial Analysis notes that an average Indian data centre consumes as much power as an aluminium smelter, or roughly 100,000 homes, while the sector’s cumulative investment of about $6.5 billion produced only around $1.2 billion in revenue last year.

The model is capital-intensive, power-intensive and creates relatively few jobs. In this respect, it differs from the services economy that built Bengaluru, Hyderabad, Gurugram and Chennai.

The Subsidies Point in the Wrong Direction

This is where the economics becomes difficult to justify. States are competing to be the cheapest location for this capital, and they compete using the very resource the grid is struggling to supply.

Telangana built power and fibre for data-centre loads before the hyperscalers arrived. It classifies data centres as an “essential service”, which guarantees them uninterrupted power, an assurance no household receives. It also offers subsidised electricity, a ten-year exemption from electricity duty, and land at up to half price for large campuses.

Tamil Nadu’s 2021 policy allowed qualifying projects to buy power at industrial tariff, waived the electricity tax on that power for five years, and charged only half the usual open-access fees.

In effect, the state gives up revenue on the input and discounts the tariff. At the same time, it must fund the transmission and generation required to serve a load that could reach 26.3 GW by FY32.

In April, the Central Electricity Authority warned Maharashtra, Andhra Pradesh, Telangana and Tamil Nadu to add capacity or risk grid instability.

Public money subsidises the input, public infrastructure carries the load, and the computing capacity produced is sold elsewhere.

Computing Capacity Has No Fixed Address

This is the central point, and it explains why a data centre differs from the factories and offices these incentives were designed for.

A factory employs local workers and often sells to local customers. A data centre does neither. It sells computing capacity, and that capacity has no fixed location. A model trained in Chennai can serve a bank in Mumbai, a start-up in Bengaluru, or a client in California.

This changes which risk deserves attention. The common concern is that these campuses will stand idle as India’s technology workforce moves to smaller cities such as Coimbatore, Indore and Kochi. That concern is misplaced. The servers remain fully used wherever the workforce lives because demand for computing comes from everywhere.

The real difficulty is the reverse: the infrastructure runs at full capacity, yet the host city captures very little of the value it creates.

Chennai is the clearest example. The city ran out of water in 2019, and it now holds close to a fifth of national data-centre capacity.

The costs that make these facilities possible — power, water, grid upgrades and higher land prices — are local and shared across the public. The value they produce is dispersed and privately held.

Concentration by Design

The design of these incentives makes the problem worse. A gigawatt-scale campus requires investment that only five or six firms in India can finance. A subsidy on megawatts and chips therefore cannot reach a company that has neither, whereas a subsidy on skilling or on small-business AI adoption could.

The policy did not accidentally favour a small number of large conglomerates; its structure made that outcome likely.

It is worth noting that even Tamil Nadu, which pursued the sector aggressively, allowed its data-centre policy to lapse in March 2026 without a clear replacement, suggesting that the trade-off is being reconsidered.

A New Case of Virtual Water Export

India has run a version of this bargain before. The Green Revolution was a national success, built around the legitimate priority of food security. Subsidised power, irrigation and assured procurement made wheat and rice attractive, helping India move from scarcity to self-sufficiency. But the same incentives also encouraged the concentration of water-intensive crops in regions where groundwater was already under pressure.

The sharper parallel is not that GPUs are the new wheat. It is that India already exports a scarce resource without calling it an export: water embedded in agricultural commodities. This is known as virtual water.

When India exports rice, it is not only exporting the grain; it is also exporting the water consumed to produce it.

A 2024 study on India’s rice trade reports that each tonne of rice exported carries 3,423 cubic metres of water consumed in India. For the period 2014-15 to 2018-19, the study calculates total virtual-water exports through rice at between 35.97 and 43.49 billion cubic metres a year, reaching 41.15 billion cubic metres in 2018-19. A 2023-24 estimate by Sathguru Management Consultants puts the virtual water associated with India’s rice exports at 40.87 billion cubic metres.

The policy logic is familiar. Subsidised electricity, irrigation and procurement make particular crops economically attractive; farmers respond to those incentives; scarce groundwater is drawn down locally; and the resulting commodity is sold into national and international markets.

The benefits of the production system are distributed through farm incomes, food security and export earnings, while part of the environmental cost remains concentrated in the places from which the resource was extracted.

The data-centre model presents a new version of the same question. A data centre does not export water in a physical commodity. It consumes land, electricity and water in one location and converts them into computing capacity that can be consumed anywhere: in another Indian city or on another continent.

The resource remains local; the economic value of the output is geographically mobile.

The comparison should not be overstated. Rice is a physical commodity and food security is a fundamental public objective; computing is a strategic infrastructure service with very different economic characteristics. But the underlying policy question is similar: when public policy subsidises a resource-intensive activity, who receives the value and who bears the cost of the resource it consumes?

India therefore needs to examine its emerging AI infrastructure not only through the lens of investment attracted or gigawatts installed, but also through the lens of resource exports.

If a water-stressed city supplies the power and water that enable computing capacity to be sold elsewhere, the country risks creating a new form of virtual resource trade — one in which the city bears the ecological and infrastructure burden while the highest-value output leaves its boundaries.

Subsidised inputs, scarce resources consumed locally and high-value output captured elsewhere. India has seen this pattern in agriculture. The AI data-centre boom could become its next, more technologically sophisticated expression.

What Development Should Deliver

None of this makes the build-out worthless. Sovereign computing capacity is a genuine strategic asset: a country that rents all of its AI capacity from foreign firms does not control its own digital future, and affordable public compute lowers the barrier for Indian researchers and start-ups.

The state incentives also generate construction work and a short-term increase in local activity. However, temporary construction and a few hundred permanent jobs are not the large-scale employment and consumption that the term “development” implies.

Strategic infrastructure and local economic development are different objectives, and India is funding the first while presenting it to citizens as the second.

The timing raises the stakes. India’s youth share of the population is projected to fall from 27% in 2021 to 23% by 2036, so the period in which a young workforce can be converted into rising incomes is narrowing rather than widening.

The gig economy, which the Economic Survey expects to approach 23.5 million workers by 2030, is dividing into high-skill and low-skill segments with a shrinking middle. In a labour market that is already under strain, AI introduces a further geographic divide, separating those who can relocate their work to a cheaper city from those who cannot.

The solution is not to spend more through the same channel. Public subsidies and state concessions should be linked to outcomes that reach people, such as jobs, skilling and technology adoption, rather than to the number of gigawatts installed.

The wider costs should be priced accurately, so that the campuses, rather than households, pay for the power and water they consume. And a private, export-oriented computing utility should not be classified as “urban development”.

India is right to build for AI. But the 26.3 GW of demand it will add by 2032 must be paid for by someone, and as the incentives are currently designed, the cities that host this infrastructure will pay the most and retain the least.