Cheaper AI Won't Cut Power Bills: McKinsey Forecasts 24% Annual Growth in Data Center Electricity Consumption
nashnova research
McKinsey and BCG published separate reports this week warning that falling AI costs are driving larger-scale deployment, with global data-center power demand projected to grow 24% annually through 2030 — the savings from cheaper models get swallowed by more usage.
Cheaper per task, more power overall — how does that work?
McKinsey identifies a structural paradox: as the cost and energy per AI task drops, the range of viable use cases expands, and total power consumption rises.
This means → efficiency gains do not translate into aggregate savings — they are fully offset by broader adoption.
In plain terms = it is the same dynamic as widening a highway — more lanes attract more cars, and congestion stays. Economists call it the Jevons paradox.
How fast will data-center power demand grow?
McKinsey forecasts global data-center electricity consumption will grow at 24% per year through 2030, then slow to 5% per year through 2040.
Data centers are now labelled the fastest-growing load segment in OECD power markets, and in several markets they are the single largest driver of electricity-demand growth before 2030.
This reflects a shift from AI pilot programs to full-line enterprise deployment — grid pressure is moving from "future risk" to "present reality."
How are leading companies managing compute spend?
BCG surveyed 1,300 executives across more than 20 industries and found that the most AI-mature firms — dubbed "future-ready" — are not letting compute costs run unchecked.
Three-quarters of these firms set explicit targets for compute (token) spending, tying each target to a concrete return.
This means → the companies furthest ahead already treat AI spend as a managed investment line, not an overhead utility bill.
Encourage use or cap it — why are companies splitting?
About half of leading firms actively encourage employees to use paid AI tools to maximize adoption, compared with just 25% among less mature peers.
A separate 22% of leading firms take the opposite approach — imposing usage caps or controls to contain costs.
In plain terms = both groups are "managing," but one manages for "not enough adoption" and the other for "spending too fast." The strategy depends on where a company sits on the AI-maturity curve.
What happens after 2030?
McKinsey flags significant uncertainty in the long-term trajectory of data-center growth beyond 2030.
The key variable: whether AI can keep delivering proven value across a widening set of business applications — if returns disappoint, companies may pull back deployment, and growth could slow faster than forecast.
This means → the current 24%-per-year projection rests on the assumption that AI adoption keeps expanding — an assumption that has not yet been fully validated.
市场有风险，内容仅供研究参考，不构成投资建议。