Racks are drawing far more power
A typical enterprise rack drew about 11 kW in 2026, up from 9 kW the year before, according to the Uptime Institute's annual survey. A single Nvidia GB300 NVL72 rack, a system built for AI training, can draw up to 142 kW, according to Nvidia's reference architecture. That is more than ten times as much in one rack.
Sources: Uptime Institute, 16th Annual Global Data Center Survey; Nvidia NVL72 reference architecture, as reported by Data Center Knowledge, August 2026.
Air cooling becomes impractical above roughly 50 kW per rack, so these racks also need liquid cooling and new ways of delivering power to the rack.
Then multiply by thousands of racks
Roland Berger expects global data centre capacity to head toward 165 GW by 2030, with AI driving much of the growth. In India, forecasts for 2030 range from 4.5 GW (Colliers, May 2025) to 12 GW (Wood Mackenzie, July 2026), up from roughly 1.3 to 2.2 GW in 2025. Even the lower figure is a large new load for a grid to take on in a few years.
Getting connected takes time
Generating power is only half of it. A campus also needs transmission, substations and a connection agreed with the utility. Industry voices quoted by Data Center Knowledge say that in many markets the limit is how quickly operators can get enough grid power, not demand or access to GPUs. Roland Berger describes grid availability as an ongoing bottleneck, alongside the problem of rack-level power.
What this means for how campuses are planned
If the grid is the slowest part, the sensible move is to plan power first. At InduS.AI Ventures that means solar, wind and storage at the core of each campus, with the grid as a backup for resilience rather than the only source. The aim is a campus that can be built when it is needed, with clean power designed in from the start.
Sources
- Data Center Knowledge: AI Rack Density's Real Limits: Power, Cooling, Failure Risk (28 August 2026)
- Roland Berger: Power, not GPUs, is the limiting factor for AI data centers (31 March 2026)
- Colliers, May 2025, and Wood Mackenzie, July 2026, India data centre capacity forecasts (as cited on our home page)

