Goldman Sachs: AI Power Demand Equivalent to Building Another Japan, Supply Constrained by Seven Major Bottlenecks
nashnova research
Goldman Sachs projects new data-center power demand through 2030 will match Japan's entire national electricity consumption — the world's fifth-largest power user; the real uncertainty has shifted from whether AI needs power to whether equipment, workers, transmission, and politics can keep up.
How big is the demand, exactly?
From early 2024 through 2030, AI-driven data centers will add power demand equivalent to all of Japan's annual electricity consumption. This means → AI is not adding load at the grid's margin — it demands an entire industrial nation's power system built from scratch.
Goldman raised its U.S. power demand CAGR forecast from 3.2% to 3.5% through 2030. Actual load growth in 2025 already exceeds 4%, running ahead of the revised projection.
U.S. data-center power forecasts rose from roughly 83 GW to 108 GW by 2030; global data-center power growth was revised up from ~117% to ~170% over 2025 levels.
Why hasn't efficiency improvement saved any electricity?
Model efficiency is improving, but the savings are consumed by three forces: more compute calls, larger agentic workloads, and higher hyperscaler budgets. In plain terms = the models run leaner, but everyone uses them more — net power consumption still rises.
Hyperscaler capex forecasts were raised in tandem: 2027 from $1.2 trillion to $1.7 trillion, 2029 from $1.5 trillion to $2.1 trillion.
This reflects a pattern where big tech reinvests every efficiency gain into larger-scale compute expansion rather than cutting spend.
Where exactly are the supply bottlenecks?
Goldman's assessment: the question has shifted from "is there enough power?" to "can it be built fast enough?" The bank distills constraints into "Four Ts" — Turbines, Transformers, Transmission, Tradespeople.
Analyst Brian Singer extends this into a "Seven Ps" framework: Pervasiveness, Productivity, Price of power, Policy, Parts, People, Physical environment. In plain terms = power plants can be designed, but the equipment, labor, wiring, and permits are all stretched thin.
Parts shortages are acute: one major gas-turbine manufacturer expects half its capacity through 2031 to be sold out by year-end. On labor, U.S. power load barely grew over the past decade, leaving electricians and high-voltage welders in short supply with long training cycles.
Where will the power come from during the transition?
Expected data-center power mix: roughly 60% natural gas, 40% renewables including storage. Near-term relies on simple-cycle gas turbines plus renewables-and-battery; mid-term shifts to combined-cycle gas; long-term hopes rest on nuclear, including some reactor restarts.
Behind-the-meter gas generation — power plants built by data centers themselves, not connected to the public grid — is viewed as the current "relief valve." Goldman raised its 2030 behind-the-meter gas forecast to roughly 30 GW, covering about 20% of total data-center demand.
This means → behind-the-meter gas is a bridge, not an end state. Large customers still prefer grid connection long-term; some projects will start as islands and connect later.
Why might water and politics be even harder to solve?
Goldman describes water as a constraint that is "harder and more expensive than adding new power," citing West Texas as a prime example.
The U.S. and China are taking opposite approaches: U.S. data centers prioritize water savings at the cost of lower electrical efficiency (worse PUE, better WUE); Chinese regulators lean toward saving power and accepting higher water use. In plain terms = water and power sit on a seesaw, and each country chose a different end.
On the political front, grid interconnection is flagged as the "single most important issue" facing U.S. utilities. Some governments are trying to set rules preemptively to prevent permitting from freezing for years. Major data-center markets have seen vacancy rates drop from 2%–7% to 1%–2%, with Goldman projecting a recovery to about 3% by 2030.
What does this mean for investors?
The core signal: the scale of AI power demand is no longer in dispute. The real uncertainty is whether equipment, workers, transmission, and politics can keep pace.
This means → the pricing logic has changed. The bet is no longer "does AI need power?" but "who can clear supply bottlenecks fastest?" Valuations of utilities, power equipment, and natural-gas infrastructure assets will directly track how tight or loose these bottlenecks remain.
Regional divergence matters: the PJM grid is already under strain; ERCOT is projected to tighten around 2028 as load climbs. Goldman does not forecast a nationwide U.S. shortage, but regional stress will intensify significantly.
市场有风险，内容仅供研究参考，不构成投资建议。