India is about to spend the next decade building data centres it should have built five years ago. The demand is contracted, not forecast — and the market is pricing the wrong layer of it.
Start with one number, because it does most of the work. India has roughly one megawatt of data centre capacity for every million people online. The United States has fifty one. China has four. The global average is five. India is not a little behind. It is an order of magnitude behind, sitting on 969 million internet users and 229 exabytes of annual wireless data that grew 17.5% last year. All of that traffic has to land somewhere physical. Right now, it largely does not have anywhere to land.
Our argument is about something specific — India AI infrastructure is a contracted demand story, and the value from it accrues unevenly. It flows first to power, grid, cooling and buildout, not to the IT services names that show up every time someone screenshots India AI stocks. If you own the theme through the obvious vehicle, you own its lowest slope expression.
A market that is already sold out
The cleanest evidence that this is real and not narrative? India ran a 4.3% data centre vacancy rate in the first half of 2025. In commercial real estate, anything under 5% means the asset class is effectively full. Net take up was 97.9 MW in six months, up 48% year on year, against a live base of about 1,123 MW. But the number that should stop you is from 2024 — India absorbed 407 MW of capacity and added only 191. Demand beat new supply two to one.
And the pipeline is already claimed. Cloud providers have pre-committed 800 MW specifically for AI workloads — more than half of everything currently live in the country, reserved before the concrete is poured. Behind them, Microsoft, Google and Amazon have committed on the order of 68 billion dollars of capex into India. It essentially separates this theme from the usual hype cycle. The buyers have already signed.
The bottleneck is not code. It is the grid
Consensus and reality do part ways — the instinctive way to play AI in India is to buy software: TCS at around 16x, Infosys at around 15x. They are fine businesses and the AI tailwind is genuine, but it is incremental to a services model that is already mature. It’s a cheap, defensive, slow slope.
The whole buildout is bound by electricity constraints, not software capacity. Data centre power demand in India is set to go from about 13 TWh in 2024 to 57 TWh by 2030, roughly 4.4 times in six years. India curtailed 300 GWh of renewable generation in the first quarter of 2026, with more than 50 GW of renewable capacity effectively stranded because the network cannot move the power to where it is needed — the strain is visible. You can pour a data centre in eighteen months. Building the transmission, switchgear and reliable backup to feed it takes longer, and that lag is exactly where pricing power forms.
Follow the scarcity, not the headline
When the scarce input is power and not code, the economics favour whoever sells power and keeps it reliable. The more interesting names sit one layer below the obvious ones. The buildout and integration layer — Netweb Technologies, E2E Networks, Anant Raj, Black Box — turns committed capacity into working racks. The power and cooling layer — CG Power, Hitachi Energy India, Siemens, and Cummins India — supplies the constraint itself. Cummins is the tell. Data centres are already 30 to 35% of its power generation revenue, growing around 20% a year through FY29. A company most investors file under “boring industrial” is, functionally, an AI infrastructure stock.
None of this is a buy recommendation. Several of these names already carry the optimism in their multiples, and that is the work an investor still has to do. The point is narrower — the market keeps paying up for the software expression of this theme while treating the binding constraint layer as cyclical industrials. That mismatch is the opportunity, and the risk that it closes is the reason to do the homework now rather than after the next pre-leasing headline.
The second bottleneck nobody is modelling: water
There is a bear case worth taking seriously, and it strengthens the same conclusion. Power is not the only physical constraint. Water is the next one. Cooling is thirsty, and India is structurally water stressed — 18% of the world’s population on 4% of its freshwater. Sector water use could roughly double from about 150 billion litres in 2025 to 358 billion by 2030, and S&P Global estimates 60 to 80% of India’s data centres will sit in high water stress zones this decade. That is real permitting friction, real community opposition, and real operating cost, concentrated in exactly the metro clusters where the campuses want to be.
But read it carefully and it cuts the same way as power. Water risk is most dangerous to the expensive pure play that is priced for a frictionless buildout. It is least dangerous — arguably bullish — for the layer that solves scarcity: efficient cooling, closed loop and liquid systems, grid and water infrastructure. The constraints do not kill the theme — they decide who inside it gets paid.
What we are watching
Pre-leasing velocity — does committed AI capacity push past 800 MW before new supply catches up.
Grid throughput — renewable curtailment and transmission buildout.
Cummins India’s data centre revenue mix, quarter by quarter.
Water permitting in Mumbai, Chennai, Hyderabad and Delhi NCR.