AI has moved power from an afterthought to the first question in any data center decision. Here is why, and what it means for enterprises.
The hardest part of bringing an AI cluster online is rarely the hardware. It is getting enough power to the rack, and being certain it will stay there.
For most of the history of enterprise computing, power was an afterthought in a facility decision. You chose a data center for location, price and available space, and you assumed the electricity would simply be there. AI has ended that assumption.
An AI deployment draws far more power, in a far smaller footprint, than the workloads data centers were originally built to serve. That single change has moved power to the front of the process. It is now the first question an enterprise has to answer before signing anything, and often the question that decides whether a deployment goes live on schedule or waits.
How AI Workloads Changed the Power Equation
The scale of the shift is easy to underestimate. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global electricity consumption. The IEA projects that data center electricity demand will more than double to around 945 TWh by 2030, with AI emerging as the most important driver of that growth. In the United States, Goldman Sachs Research has projected data center power demand rising from about 31 gigawatts in 2025 to 66 gigawatts by 2027. That represents an increase of roughly 35 gigawatts of power demand in just two years, putting significant pressure on electricity generation and grid infrastructure.
The practical result for any enterprise is that deliverable power has become scarce. When something is scarce, you compete for it, and the winners are usually the buyers who understood the constraint before everyone else did.
Why Utilities Have Become the Bottleneck
The reason power now governs an AI deployment is a mismatch of timelines. A data center building can go up in 12 to 18 months. Connecting a new load to the electrical grid can take five to seven years in constrained markets. The equipment that carries that power has its own backlog: large transformers now average more than two years from order to delivery.
This is why the phrase “available power” on a facility brochure deserves a second look. Available can mean energized and ready today, or it can mean planned and waiting in an interconnection queue. Those are very different things. Some markets show how tight the situation has become. Northern Virginia, long the largest data center hub in the world, now carries multi-year connection queues, and Dublin paused new data center grid connections in its area entirely. When a site depends on new utility capacity, that wait is the real deployment timeline, whatever the marketing says.
How Rack Density Drives Power Consumption
A rack is simply the cabinet that holds the servers. For years, a fully loaded enterprise rack drew somewhere between 5 and 15 kilowatts, and facilities were designed around that range. AI has broken it.
AFCOM’s 2026 industry survey put the average rack density at 27 kilowatts, up from 16 the year before and about 7 in 2021, the largest single-year jump in a decade of the survey. AI-specific systems run far above even that rising average. Current high-density AI racks draw 80 to 140 kilowatts, and the roadmap is public: NVIDIA’s Vera Rubin platform, entering volume production in the second half of 2026, is rated at roughly 190 to 230 kilowatts per rack, with a later generation specified near 600.
The takeaway does not require an electrical engineering background. A single AI rack now draws more power than an entire row of older racks. That is why the unit that matters when planning a deployment has shifted from square feet to kilowatts, and why a facility built for yesterday’s densities cannot simply be topped up to reach today’s.
What Enterprises Should Evaluate Before Choosing an AI Data Center
Because power is the binding constraint, it belongs at the start of any facility evaluation rather than buried in the technical annex. A few questions separate a site that can actually host an AI deployment from one that only looks ready.
The first is how much contracted power is available per rack today, not on a future roadmap. The second is whether that power is already energized or still depends on a utility connection in the queue, since the answer sets your true timeline. The third is whether the cooling matches the density, because racks above roughly 40 kilowatts need liquid cooling and cannot be served by air alone. The last is headroom: whether the facility can support the next generation of denser hardware, or whether you will be shopping again in two years.
A site that scores well on floor space and price but poorly on these four points is not a bargain. It is a delay waiting to happen.
What Power Really Decides
AI has changed what limits a data center. Land and capital still matter, but they now follow electricity rather than lead it. For the enterprise choosing where to run its AI workloads, that means power is no longer a detail to confirm at the end. It is the first thing to verify and the one most likely to determine whether the project moves.
The organizations that treat power as question one, and that insist on firm, energized capacity before committing to a site, are the ones that deploy on a workable timeline. Those that assume the electricity will be there tend to find out, months in, that it will not.
