AI data centres need a fundamental rethink of how they are powered — that is the message Greg Funk of power and cooling specialist Vertiv delivers in an interview with The Economic Times.

Funk's argument lands at a moment when hyperscale operators are straining the electrical limits of existing facilities. Training and inference clusters now draw power densities that legacy data centre designs never anticipated, and the interview frames power — not compute — as the binding constraint on AI infrastructure growth.

Vertiv sits close to the problem. The company supplies the power distribution and thermal management equipment — switchgear, UPS systems, busways, liquid and air cooling — that sits between the grid and the GPU. When rack densities rise, the electrical architecture of the building has to change with them, and Funk's core point is that incremental upgrades will not keep pace.

Why does power, not chips, decide AI buildout speed?

The interview's central claim is that the queue for electricity — grid connections, substation capacity, transformer lead times — now gates how fast AI capacity can actually come online. Operators can order accelerators; they cannot order megawatts on demand.

That dynamic shifts engineering priorities. Instead of treating power as a fixed facility input, data centre designers have to treat it as a scarce resource to be engineered around:

• higher-density power distribution inside the white space;
• cooling architectures matched to dense GPU racks rather than general-purpose servers;
• flexible electrical designs that can absorb successive generations of AI hardware without a full retrofit.

What does a 'rethink' actually mean?

Funk's framing implies redesign rather than expansion. If compute density keeps climbing, the classical model — a building engineered for a stable watts-per-rack profile over a 15-year life — breaks down. The replacement model is an electrical and thermal plant built for change: modular, dense, and reconfigurable as AI silicon generations turnover.

This is where Vertiv's commercial position and the industry's problem converge. Every reconsideration of data centre power design is, in effect, demand for the equipment Vertiv sells. The interview doubles as a case for why power infrastructure vendors move from afterthought to critical-path supplier in AI projects.

Who bears the cost?

The Economic Times conversation also touches the economics. Reworking power architecture is capital expenditure that arrives before a single model can be trained on the new site. That pushes data centre operators, chipmakers and power equipment suppliers into closer, earlier collaboration — the power plan has to be set when the GPU roadmap is set, not after.

For markets such as India, where The Economic Times anchors its coverage, the question is sharper still: AI demand growth collides with grid capacity that is already contested by industrial and consumer load.

The forward signal from the interview is clear. As long as accelerator power consumption keeps rising generation over generation, the companies that control electrons inside the data centre — not only those that make the chips — will set the pace of AI infrastructure deployment.