BofA: AI Capex Raised to $1.7 Trillion, Five Opportunities Emerge in Power Supply Systems
Miles Bennett
Bank of America raised its 2030 global AI capex forecast from $1.4 trillion to over $1.7 trillion, with power demand set to triple — electricity infrastructure is now the binding constraint on AI deployment speed.
Where does $1.7 trillion go?
Four structural drivers lifted the forecast: accelerating compute migration, hyperscaler spending increases, sovereign AI infrastructure build-out, and enterprise AI adoption.
China's share over the same period is projected at $327 billion, with a 2026–2030 CAGR of 24%.
This means → AI capex is not a one-off pulse but a multi-track, long-cycle expansion with several drivers stacking simultaneously.
Why does power demand triple?
Global data-center installed capacity is expected to rise from roughly 100 GW in 2025 to nearly 300 GW by 2030; electricity consumption from about 500 TWh to 1,208 TWh.
The core driver is rack power density: traditional cloud deployments run at 10–15 kW, the current Blackwell architecture at 100–120 kW, and by 2030 Nvidia's Feynman-era racks are forecast to exceed 1.5 MW — roughly 100× a standard CPU rack.
In plain terms = a single rack used to draw as much power as a home air-conditioner; now it draws as much as a small building's central cooling; soon it will match a small factory.
What are the five power-supply investment themes?
Transformers & the grid: global transformer supply remains tight; Asian grid-equipment makers benefit from rising Chinese grid capex and China–Korea export demand.
Gas turbines: BofA expects roughly 120 GW in global orders over the next three years, with supply tightness lasting through 2030.
Diesel generators: Chinese manufacturers have strong penetration in the U.S. AI data-center backup-power market. Energy storage (BESS) — large-scale battery storage systems: global new installations at a CAGR of 23% through 2030, with AI data-center-related BESS demand growing at 27%.
Power-supply systems: China's AI data-center power-supply market is forecast at a CAGR of roughly 25%. This means → all five themes point to one logic: the pace of compute expansion is ultimately gated by the pace of power-infrastructure build-out.
How large is China's power advantage?
Electricity prices are 30%–60% lower than in the U.S. and EU; reserve capacity stands at roughly 30%, above the U.S. (under 25%) and the EU (about 15%).
Transmission and distribution equipment averages under 20 years old, versus over 40 years in the U.S. and Europe; 46 ultra-high-voltage lines already form a mature cross-regional transmission network.
This reflects a structural moat built on four layers — price, redundancy, equipment age, and grid reach — not a single cost advantage.
Why is cooling growing even faster than power?
BofA projects China's data-center cooling market will reach RMB 70 billion by 2030, at a CAGR of 36%.
Liquid cooling — using fluid instead of air to cool chips — demand is forecast to rise from 1.4 GW to 9.5 GW, a CAGR of 47%, with penetration climbing from 30% to 70%.
Put simply = rack power has risen 100×, so air cooling can no longer keep up — cooling growth outpaces power growth precisely because the thermal bottleneck arrives before the electrical one.
What is the binding constraint on this cycle?
AI capex forecasts keep being revised upward, signaling the market is still repricing the length and intensity of this cycle.
Whether power infrastructure can keep pace with compute expansion will determine the actual deployment speed of AI data centers.
This means → investing in AI is not just about buying compute — it hinges on whether power and cooling can keep up. Infrastructure is the ceiling, not a sideshow.
Content is for reference only, not financial advice.