Nageswaran & Poojari: Let's welcome data centres -- but on the right terms

Summary

AI hyperscalers have their eyes on India, but we must carefully weigh all commitment costs of hosting data centres against the benefits before extending scarce resources. The terms must be settled before we pay in advance.

Emerging markets don't have a lot of those data centres, and it's not coming, trust me." So said World Bank president Ajay Banga in New York last month. Even where an emerging economy can arrange the land, water, power and fiscal space they demand, policymakers should ask whether hosting them is the best use of scarce resources, and whether the incentives on offer have been weighed against their costs.

The resources that data centres need are largely spoken for in most developing economies. Every unit of power committed to a data centre is a unit not available to a factory, an irrigation pump, or a household still making its way up the energy ladder. Finances are no less contested. Water is a third constraint: cooling needs are rising while groundwater across much of the emerging world is depleting faster than it can be replenished.

With power, water and finance this contested, what do the public subsidies and private credit diverted towards data centres buy in return? In his recent paper, 'Financing the AI Buildout,' Stijn Van Nieuwerburgh of Columbia Business School estimated that about two-thirds of the cost of an artificial intelligence (AI) campus goes on computing hardware, which has an economic life of around six years.

The building and its power supply will outlast the chips, but the tenant's economics rests on the chips. Its commitment to any one site can therefore be far shorter than the concessions and grid investments made to attract it. Nor do these facilities create many jobs once built. Virginia's legislative audit commission found in 2024 that a typical facility runs on around 50 full-time staff, about half of them contractors.

Hosting a facility does not necessarily keep citizens' data at home either. Where a facility sits and whose data it holds are separate questions. The work that must stay at home, such as government records or payments data, needs far less capacity than the hyperscale now being attracted.

On the wrong terms, then, hosting a hyperscale data centre can amount to a new kind of value transfer. The host country supplies land, subsidized power and scarce water, and absorbs the strain on its grid, while the rents accrue to a handful of firms headquartered in advanced economies. These firms bring capital, but much of it is spent abroad on equipment, and it is increasingly structured to keep the risks off their own books.

Hyperscalers increasingly lease their data centres from developers or from vehicles funded by private credit. As Ellen Carr noted in the Financial Times, these obligations stay largely off their balance sheets during construction.

In an emerging economy, the developer, its lenders and the distribution company that serves the facility are all exposed to a single tenant, on lease terms shorter than the life of the assets. If demand falls short or the hardware ages faster than expected, the tenant may walk away or default, leaving the building, the loans and idle grid capacity with domestic banks, utilities and, ultimately, taxpayers. The gains would flow abroad, while the costs of disappointment would be borne at home.

Such arrangements are already under scrutiny in the home markets of these very firms. Amsterdam has barred new data centre construction until at least 2030. In July, New York paused environmental permits for large data centres for up to a year. Niti Aayog has itself cited Virginia, Dublin, Amsterdam, Johor and Bangkok as places where capacity outran the grid, water and land.

Indian states, meanwhile, are moving the other way, competing to attract data centres with electricity duty exemptions of up to two decades, tariff subsidies, waivers on transmission and wheeling charges, and industrial rather than commercial tariffs.

Newer policies tie these concessions to a rising share of renewable power. That makes them greener, not cheaper. The exchequer and the distribution company lose the same revenue whatever the source of the power, and renewable capacity has real resource costs of its own. How many of these policies came with a public estimate of the revenue forgone over their full term, or of the load they add to stretched grids?

While the Union Budget 2026-27 offered a tax holiday till 2047 to foreign companies selling cloud services abroad from data centres in India, it is worth noting that the concession is different. It gives up tax on income that would most likely be booked elsewhere.

Meanwhile, state concessions on power, water and land hand over scarce real resources below cost, and leave utilities and lenders holding the risk. These most need a full public accounting.

None of this means that data centres bring no benefits. Their services can be exported. They can anchor a domestic AI ecosystem and give the country computing capacity of its own. A large customer paying the full tariff can also strengthen a distribution company's finances.

States now have a sound guide for weighing the two in the data centre evaluation framework Niti Aayog released this month. It ties approvals to verified readiness of power, water and land at each stage of capacity addition. It asks states to protect existing water users and keep projects out of stressed basins.

It separates grid upgrades caused by data centres from normal demand growth so that costs fall where they belong. It asks states to set the support each project receives against its net fiscal contribution, and to release incentives in tranches tied to outcomes, with triggers to suspend them.

Above all, it urges states to compete on readiness rather than on investment announcements. The framework is advisory. States would do well to adopt it.

However, two gaps remain. The framework says little about what happens if the tenant leaves. Grid and water assets built for one customer on a short lease need minimum-offtake commitments or guarantees that run as long as the assets do. Nor does it ask that data centres pay cost-reflective tariffs for power and water, alongside the emissions-linked charge the International Monetary Fund has suggested. Only when these costs are priced in can we judge what incentives, if any, are justified.

Ajay Banga may well be right about most emerging markets. India is an exception, because the hyperscalers are coming. That makes it more urgent, not less, to settle the terms before we pay in advance for the privilege of hosting them.

These are the authors' personal views.

The authors are, respectively, chief economic advisor, Government of India, and former consultant, ministry of finance.