The pattern in one line: India's AI data-centre build-out is concentrating in a handful of coastal and near-coastal clusters — Chennai above all — not because that's where power is cheapest, but because that's where the submarine cables that can actually move an AI-scale training dataset already come ashore. Land and power get a company into the game; a cable landing station gets its data out of the country fast enough to matter.

India's operational data-centre capacity is roughly 1.2–1.5 gigawatts today. Forecasts for 2030 disagree sharply — S&P Global's consensus range is 6–6.5 GW with a 9 GW bull case, Mordor Intelligence projects 15.21 GW by 2031, and a Jefferies estimate cited by industry trackers runs to 17 GW — but the one figure with an actual government commitment behind it is more concrete: the Ministry of Power built a projection of 13.56 GW of data-centre electricity demand by FY2031–32 directly into national transmission planning, published via the Press Information Bureau in March 2026. Whichever forecast turns out closest, every one of them describes a four-to-tenfold expansion inside five years, and private capital is already moving at that scale: Adani and Reliance alone have committed more than $210 billion combined to AI infrastructure and the renewable power to run it, through 2035.

The projects aren't spread evenly across India — they cluster around a short list of cities, each chosen for a specific mix of land availability, renewable power access, and, critically, proximity to a submarine cable landing station.

State-level policy is stacking incentives on top of this: Uttar Pradesh's Data Centre Policy 2026 targets 2 GW of capacity and ₹2 lakh crore of investment with capital and interest subsidies; Telangana has a stated vision for 5 GW by 2030; Maharashtra, Tamil Nadu, and Andhra Pradesh run comparable schemes.

The reason Chennai shows up so often in that table isn't coincidence: it's India's densest submarine-cable landing point, with six active cables offering a combined 14.8 Tbps of capacity, and it sits closer to the US West Coast than any other Indian landing city. Mumbai is the other major hub. Training and running large AI models means moving enormous, continuous volumes of data between compute clusters and, often, between countries — a single frontier-model training run can involve moving petabytes of data repeatedly between storage and GPU clusters, and hyperscalers serving global customers need low-latency routes out of India entirely. A data centre with cheap land and abundant power but no nearby cable landing station is still, in practical terms, an island.

Chennai and Tuticorin together are why Tamil Nadu keeps appearing in the project table above — it isn't only a land-and-power story, it's a cable-landing story that predates the current AI build-out by well over a decade (i2i and the Tata Indicom Cable have served the Chennai–Singapore corridor for some 15 years) and is now being actively reinforced with new capacity purpose-built for hyperscale demand.

Cable capacity and campus announcements don't resolve the two constraints every source describing this build-out flags as the real bottleneck: transmission, not generation. India's data centres currently draw well under 1 per cent of national electricity (estimates range 0.5–0.8 per cent), so aggregate generation capacity isn't the near-term limit — the problem is that the specific substations serving Mumbai and the National Capital Region are approaching local capacity limits, and India's transmission and distribution losses, at 16.64 per cent, run more than double the global average. A hyperscale campus can secure land, a renewable power-purchase agreement, and a cable interconnect, and still be waiting on a substation upgrade. Water is the other quietly binding constraint in several of these same locations: a single 30 MW facility can consume close to 780 million litres of water a year for cooling, in states where water stress is already a live policy issue.