June 1, 2026 · 7 min read · Backplane Team
How Much Power Does a Data Center Actually Need?
Power demand for AI infrastructure is measured in megawatts of critical IT load, not square footage — and that single shift in units explains most of what's changed about how sites get chosen. A building that would have been plenty for a traditional enterprise data hall can be badly undersized for an AI deployment, while a much smaller footprint with the right power position can outperform it.
From a single rack to a hyperscale campus
A single high-density GPU rack can draw anywhere from roughly 30 kW to well over 100 kW depending on the hardware generation and cooling approach — an air-cooled 8-GPU node sits toward the lower end of that range, while a liquid-cooled rack-scale system like NVIDIA's NVL72 architecture is built specifically to push far past it.
Scale that up and the numbers move fast. A modest training cluster might run a handful of megawatts; a large frontier-scale training run or a big multi-tenant inference deployment can require tens to hundreds of megawatts of continuous, reliable power. There's no single 'typical' number — the honest answer is that power requirement is the first question, not an afterthought, in scoping any deployment.
Why site selection has shifted to power fundamentals
Traditional data center site selection weighted things like fiber proximity, tax incentives, and labor market access heavily. Those still matter, but they've been overtaken by a more basic question: can this site actually get the power it needs, on a timeline that matters?
That's a direct consequence of AI compute demand growing faster than new generation and transmission capacity can be built. The bottleneck isn't chip supply as often as it's assumed to be — it's the physical and regulatory process of getting meaningful new power capacity interconnected to the grid, which in many U.S. markets now commonly runs into multi-year timelines.
What this means if you're evaluating a site
For a property owner, the practical question is simpler than it sounds: what's already on site? Existing service entrance capacity, an on-site substation, or direct proximity to transmission infrastructure can turn what looks like a stranded industrial asset into a highly competitive site — often faster than a greenfield build sitting in a multi-year interconnection queue.
- Existing service entrance rating (in MW, not amps at a residential voltage)
- Whether there's an on-site or adjacent substation, even if it's not currently energized
- Transmission-level vs. distribution-level interconnection — transmission ties generally support far larger loads
- Any remaining or transferable interconnection agreements from the site's prior use
The bottom line
Power capacity, not building size, is now the first filter for whether a site is viable for AI infrastructure. If you're sitting on a property with real existing power service — retired, idled, or simply underused — it's worth a serious look before anyone assumes the only path forward is demolition. See our breakdown of what qualifies as a powered site for the fuller picture.