Data Center Power Constraints Are Now a Procurement Problem, Not Just an Infrastructure One

Worldwide data-center power demand is set to rise 27% in 2026, and grid interconnection for new capacity can now take years rather than months. Most of the electricity currently requested for AI data-center projects in the US won't actually materialise. None of this used to be a question procurement teams asked — now it has to be.

The Constraint Moved From Chips to Power

For most of the last two years, GPU supply was the bottleneck everyone planned around. Power availability and grid interconnection timelines have now overtaken it — a facility with all the GPUs it needs but no reliable power commitment is not a functioning data center. If you're planning any large-scale AI infrastructure buildout, power timelines belong in your project plan alongside chip lead times, not as an afterthought.

What to Ask Before You Commit to Capacity

• Is the power commitment behind this capacity firm, or is it part of the roughly 72% of requested US grid capacity that industry data suggests won't materialise?
• Does your provider have behind-the-meter power generation as a backstop, or are they entirely dependent on traditional grid interconnection timelines?
• How is your provider's infrastructure handling the power-demand volatility that's currently damaging batteries, generators, and cooling systems industry-wide — is that risk priced into your contract or silently absorbed by you as downtime risk?

The Practical Takeaway

Treat power reliability as a genuine vendor selection criterion, not just price-per-GPU-hour — a cheaper provider whose capacity depends on a grid connection that's years from materialising is not actually the cheaper option once you factor in the risk of your AI infrastructure being delayed or unreliable exactly when you need it most.