AIBRIEF
India's grid planners see 26.3 GW of new data centre demand by 2031, and Gartner sees AI infrastructure spend nearly doubling in 2026. Cheaper AI and pricier power can be the same curve.
AI is getting cheaper to run. The infrastructure needed to run it at scale is getting more expensive.
India's Ministry of Power now projects that data centres will add 26.3 gigawatts of electricity demand by 2031-32, according to a Business Standard synthesis of the government figures. Applications for roughly 17 GW have already been filed with state transmission utilities; another 9.3 GW sits outside the formal grid-connection queue.
The same cycle shows up in the spend data. Gartner forecasts worldwide outlays on AI-optimised infrastructure-as-a-service will reach $42.3B in 2026, nearly double the $21.5B level of 2025 and on track for $66.1B in 2027, while total IaaS grows at about 29.3%. The always-on load is doing the work: global inference spend is projected at $23.3B this year, surpassing the $19B earmarked for AI training for the first time.
That crossover is the mechanism. Cheaper per-query inference lowers the price floor, invites more queries, and migrates the savings up the stack into power bills and grid build-out that someone else eventually pays.
Forecasts are forecasts: the 26.3 GW figure can slip, and supply-side efficiency gains can compress the curve. The 2026 print is the test.