Data Centers Could Eat 20% of US Power by 2035

New grid analysis projects that data centers could consume up to 20% of US electricity by 2035, with AI workloads the dominant growth driver. Behind the headline number sits the real story of this hardware cycle: the binding constraint on AI expansion is no longer chips. It is megawatts, transformers, and interconnection queues measured in years.

Why power became the bottleneck

A single modern AI rack draws more electricity than a neighborhood — hundreds of kilowatts where legacy racks needed tens. Hyperscale campuses are now specified in gigawatts, which makes them utility-scale industrial customers with lead times to match. The sequence that limits every announcement: grid capacity studies, transmission upgrades, substation construction, then — only then — servers. NVIDIA’s DSX platform and AMD’s Open Rack Wide standard both exist because the industry has accepted this reality and is trying to industrialize around it.

• Rack density: moving from tens to hundreds of kilowatts, with liquid cooling mandatory above ~50kW
• Interconnection queues: multi-year waits for grid capacity in key markets — the true gate on every AI campus announcement
• Generation scramble: nuclear restarts, gas turbines on multi-year backorder, renewables-plus-storage hybrid designs all being courted

The second-order effects rippling out now

• Municipal pushback on datacenter water and power use is becoming a permitting variable
• Power purchase agreements are being signed years ahead, locking in AI economics
• Grid operators are re-planning load forecasts around AI demand curves for the first time

The implication for everyone outside hyperscale

The power crunch is quietly making the case for right-sized local AI. Every inference run on a desk-sized machine instead of a hyperscale GPU is demand that never touches the grid queue — and for latency, privacy and cost reasons, local inference is increasingly the better technical answer anyway. The strategic pattern for businesses: cloud for frontier capability and elasticity, local for the high-volume workloads, and avoid building anything at the scale where you compete with hyperscalers for the same strained resources.