AI clusters consume electricity at industrial scale. A single rack of Blackwell GPUs can draw over 100 kW, and hyperscalers are planning campuses that need hundreds of megawatts of clean, reliable power delivered through transformers, switchgear, bus bars, and cooling systems that did not exist at this density five years ago. This theme tracks the companies that sell into the physical power and thermal infrastructure layer of the AI build-out.
Why it matters
- Power is the binding constraint. Chip architects can design faster silicon, but the data center cannot use it if the power and cooling infrastructure cannot keep up. Electrical and thermal delivery are the actual bottleneck at scale.
- Lead times are long. Utility-grade transformers, switchgear, and gas turbines have multi-year order books. Companies with installed capacity and backlog visibility trade at a premium for a reason.
- Content per rack is rising. Every generation of GPU server needs more power conversion, more precise thermal management, and more dense electrical distribution — driving dollar content higher even if unit counts plateau.
Roster
- VRT — Vertiv — power distribution, thermal management, and services inside the data center. Direct beneficiary of rack-density growth.
- NVT — nVent Electric — enclosures, liquid cooling, and electrical connection infrastructure for data centers and industrial facilities.
- MOD — Modine Manufacturing — thermal management solutions including liquid cooling for data centers and EV powertrains.
- GEV — GE Vernova — gas turbines, grid equipment, and electrification solutions. $150B+ backlog across generation and transmission.
- POWL — Powell Industries — custom switchgear and bus duct systems for data centers, LNG, and industrial power infrastructure.
- ETN — Eaton — electrical components and power management across data centers, utilities, and industrial end markets.
- VICR — Vicor — high-density power conversion modules for AI servers, networking gear, and advanced electronics.
- AGX — Argan Group — power plant and industrial facility construction, including data center infrastructure projects.
What to watch
- Rack power density trends — 40 kW/rack was high-end two years ago; 100 kW+ is becoming the baseline for AI. Every step up drives content.
- Liquid cooling adoption — air cooling hits physics limits above ~40 kW/rack. Watch NVT, VRT, and MOD for rear-door and direct-to-chip liquid bookings.
- Transformer and switchgear backlogs — POWL and ETN order books are a proxy for how much new power capacity is being committed.
- GEV gas turbine slot reservations — the upstream indicator that hyperscalers and utilities are locking in generation capacity.
- Hyperscaler capex guidance — any pause or acceleration in Meta/Microsoft/Amazon/Google spending changes the demand curve for all of these names.