India's AI Ambition Needs 40–50 TWh of New Power

Every large model trained and served at scale is, underneath the architecture diagrams, a power procurement problem. India's AI infrastructure ambitions are projected to require 40–50 TWh of new annual generation by 2030 — a load that has to come from somewhere firm, clean, and fast to build.

China currently generates 420 GW of hydropower to India's 50 GW, a gap that understates the opportunity more than it explains it: India's eastern Himalayas alone hold an estimated 150 GW of untapped hydro potential, largely undeveloped.

The infrastructure to move it is already being planned. India's North-East river-link program targets 76 GW of evacuated hydro-capacity by 2047, explicitly designed to convert monsoon-season damage into exportable, dispatchable clean power.

The arithmetic is straightforward: the compute race needs firm power at a pace renewables-plus-storage alone may not match, and the Himalayas are the largest undeveloped firm-power reserve within India's own borders.