Most of us who have lent to roads and power plants would say no. A toll road takes 15 years or more to earn back its cost, and a power plant around the same. Yet that is now being said about artificial intelligence (AI) compute in America. Jensen Huang says he has heard of investments returning their capital inside 12 months. Elon Musk thinks six months next year is “not impossible”. Nebius put the payback on its larger recent deals at a year and 10 months.
Returns like these go to whoever owns the chips, and in India that is mostly not us. But a good part of the boom is being built here, and I find it useful to think of it in two layers. The first is the data centre itself: Land, building, substations, cooling, water, the generator sets in the yard. The second is what goes inside it, the chips. Indian banks can finance the first without breaking a sweat. The second needs a lot more thought.
India’s boom is much smaller, and of a different kind. Operational capacity should touch about 2 gigawatt (Gw) by December, and most forecasts for 2030 sit between 5 and 8 Gw. The announcements are higher. Adani 5 Gw by 2035, AirTrunk 5 GW by 2030, Google 1 Gw at Vizag, a multi-Gw Reliance campus at Jamnagar. Interestingly, commitments since 2021 add up to about $175 billion. US’ four largest hyperscalers will spend nearly four times that amount this year alone.
Most of what we are building is real estate with a very large electricity bill. Apart from Reliance and Yotta, Indian money is going into land, power and buildings, rented to Google, Microsoft and Amazon, who bring their own chips. NTT’s Indian business earns about ₹9.7 crore of revenue per megawatt (Mw), and assuming a 40 per cent margin a new building needs something like 15 years to repay itself. No equity investor gets excited by that. A project lender should.
Two layers, two very different loans
The building, with its power and cooling, costs ₹50-70 crore per Mw, or $5-7 million, depending on whose estimate you prefer. The chips cost several times as much. By Jensen Huang’s own reckoning, Nvidia’s share of a Gw is about $25 billion for Blackwell and $40 billion for the coming Rubin generation: $25–40 million per Mw, counting only what Nvidia sells. The building will last 25 years. The chips run the risk of looking dated in a few years.
Layer one: The building
Taking India from 2 Gw to 5-10 Gw means building another 3-8 Gw. At ₹60-70 crore a Mw, that is ₹2-5 trillion by 2030, before anyone builds the power plants to run them. If 60-70 per cent is debt, banks are looking at ₹1.2-3.5 trillion of fresh lending.
Put that next to what banks already hold. Bank credit to all of Indian infrastructure was ₹15.1 trillion in June 2026. The buildings alone could add 8-23 per cent to that book in four years, somewhere between a fifth and more than half of what banks now add to infrastructure lending each year. The solar farms and transmission lines behind them come on top.
This is lending our banks understand, and the credit is better than a fair bit of what they already carry. A toll road lives on traffic forecasts, and a power plant on the goodwill of a state discom. A data centre’s tenant is Microsoft or Google, often signed up before the building is finished, in a market where JLL puts vacancy at a record low of 2.8 per cent. Some banks are already at it: SBI and Central Bank have lent ₹5,500 crore to CtrlS, and eight global banks up to $1.44 billion to AdaniConneX. Insurers have a part to play too. A recent tax change makes sale-and-leasebacks and data-centre InvITs possible, so a bank can build, sell the finished asset to an insurer or a pension fund, and move on to the next campus.
The second layer is where the real money is, but also the real risk. Most chips that come to India will sit on Microsoft’s or Google’s balance sheet and never come near an Indian banker. What is open to us is the Indian-owned slice. Reliance is building several gigawatts at Jamnagar, and Yotta wants to grow from 180 Mw to 800 Mw by 2029. If even one Gw of India’s build-out is filled with chips owned by Indian companies, that is $25-40 billion, or ₹2.4–3.8 trillion. At two gigawatts, the chips would need more money than all the buildings put together. A chip is seemingly weaker collateral. It is bought in dollars, rented out on contracts of three years or less, and its value in five years depends on whatever Nvidia launches next. Only a few of our largest banks can even attempt this. The Reserve Bank of India caps a bank’s exposure to any one corporate group at 25 per cent of its Tier-I capital, and a single Gw of chips would take even State Bank of India past that limit. In practice it will be a handful of large banks lending together, with alternative investment funds (AIFs) and credit funds alongside.
It will also need clever structuring, and we have plenty now from America. Reading about it took me back to the good old days of sale-and-leaseback and residual-value financing. Meta moved its Hyperion campus into a joint venture 80 per cent owned by Blue Owl funds, which borrowed $27.3 billion against it; Meta leases it back four years at a time and guarantees its residual value should it walk away. CoreWeave borrows against its customer contracts and graphic processing units (GPUs), on five-year loans against three-year contracts, so its lenders are betting on what the chips will fetch later. Its unsecured bonds now are underwater, yielding over 15 per cent.
For the buildings, power is the hard part, and the problem is the grid more than generation: In the first quarter of 2026 India curtailed 300 gigawatt-hours (Gwh) of renewable power because transmission could not carry it. Each Mw needs 25-30 million litres of water a year, often in cities already short of it. Developers have increased from 5 to 18 since 2019, Icra says rents are already softening, and 12 Gw of announcements against 5-8 Gw of forecast demand in itself has the makings of a glut.
The chips do carry sharper risks — obsolescence, contracts that run out before the loans do, dollar debt against rupee income. A company that borrows to rent out assets that lose value this quickly is, for practical purposes, a non-banking financial company, and we have seen how those behave when the good times end. None of this is a reason to stay away. It is a reason to lend on signed leases and secured power, and to be cautious about residual value.
The landlord’s dividend
America is paying for scarcity, and the money is being made mostly by those who own the chips. We have neither the chips nor the cloud businesses built on them, and I don’t see that changing this decade. It is a little disheartening.
The first layer, though, is ours for the asking. It is a ₹2-5 trillion book of long, contracted lending to some of the best credits in the world. I can think of very few better opportunities for our banks and insurers over the next few years.
The second layer will be for the few large banks and AIFs willing to learn what a used chip is worth. We will not be designing these chips anytime soon. With a bit of old-fashioned financial engineering, we could at least be financing them and hoping to earn the returns provided by short payback periods.
The writer is a Professor of Practice at IIT Bombay. He retired as MD & CEO of Kotak Mahindra Bank. Views are personal
Disclosure: Entities controlled by the Kotak family have a significant holding in Business Standard Pvt Ltd
