AI's Power Swings Are Damaging the Data Centers Built to Run It

Worldwide data-center power demand is expected to rise 27% in 2026, reaching 132 gigawatts, up from 104 GW in 2025, en route to an estimated 290 GW by 2030. The problem is not just volume — rapid swings in AI workload power draw, spiking as much as 50% above a facility's design capacity, are straining batteries, generators, and cooling systems, causing them to malfunction or wear out far sooner than expected.

Why AI Workloads Draw Power So Differently

Traditional data-center loads are relatively steady. AI training and inference workloads ramp up and down sharply as jobs start, pause, and finish — producing power demand curves that look nothing like what most data-center electrical and cooling infrastructure was engineered to handle. Equipment rated for steady-state loads simply was not built for this kind of volatility, and it's failing accordingly.

The Supply Side Is Even Worse Than It Looks

More than two-thirds of the electricity currently being sought for US AI data-center projects is unlikely to actually materialise, due to speculative or 'phantom' project pipelines — grid operators and utilities are realistically committing to only about 28% of the 1,066 gigawatts requested. Connecting a new facility to the traditional grid can now take years, which is pushing operators toward behind-the-meter power generation sited directly at the facility instead of waiting on grid interconnection.

What This Means for Anyone Planning AI Infrastructure

• Power availability and interconnection timelines, not just chip supply, are now the binding constraint on how fast you can scale AI infrastructure — factor multi-year grid timelines into any large-scale on-prem or colocated AI build
• Cloud providers absorbing this power volatility cost is a real input into future compute pricing — expect it to show up in your bill eventually, even if it hasn't yet
• Workload-level power efficiency (right-sizing models, batching inference, avoiding unnecessary always-on capacity) is becoming a genuine cost and reliability lever, not just a sustainability talking point