The AI Power Buildout
Why electricity, transformers, interconnection, and thermal infrastructure may define the next phase of AI buildout.
AI infrastructure is turning into a physical buildout story. Compute still matters, but the capex stack around it is increasingly shaped by power generation, transmission, transformers, switchgear, cooling, and interconnection timing.
- Demand: larger clusters, denser racks, and faster data-center expansion.
- Bottlenecks: transformers, grid queue backlogs, substations, switchgear, and thermal capacity.
- Winners: electrical equipment, EPC and grid services, utility-linked enablers, and thermal infrastructure.
The power issue is no longer theoretical. The IEA says electricity demand from data centres worldwide is set to more than double by 2030 to around 945 TWh, and its broader Energy and AI work shows data-centre electricity demand continuing to climb through 2035.
That means AI is becoming a real-world infrastructure buildout, not just a software cycle. The practical constraint is not only megawatts on paper, but deliverable capacity, interconnection timing, redundancy, and the time needed to deploy equipment at scale.
- Interconnection queues: grid studies and upgrade costs can delay projects by quarters or years.
- Transformers and switchgear: long lead times and constrained manufacturing capacity.
- Transmission and substations: land, permitting, and skilled construction labour.
- Cooling and power density: higher rack density raises facility-level electrical and thermal demands.
Think in layers. AI demand can help the entire chain, but not every layer benefits equally or at the same time.
Representative U.S.-listed names by angle:
- Electrical equipment: ETN, HUBB
- Grid and construction services: PWR
- Generation and utility leverage: CEG, NRG
- Machinery and heavy equipment: CAT
- Facility and thermal beneficiaries: VRT, GEV
These are examples of exposure, not recommendations. Verify current business mix and backlog quality before treating them as pure-play proxies.
- Timing risk: capex cycles slip, reforms move slowly, and demand may arrive unevenly.
- Execution risk: project delays, cost overruns, and supply chain issues.
- Valuation risk: infrastructure names can price in a lot of optimism early.
- Efficiency surprises: better model efficiency could reduce power intensity per unit of compute.
- Start with the bottleneck: transformers, interconnection, switchgear, generation, or cooling.
- Prefer backlog and pricing power: this is an industrial cycle, not a one-quarter trade.
- Watch regional concentration: grid constraints and utility exposure vary by geography.
- Size for cyclicality: some names are resilient compounders, others are pure capex beneficiaries.
Use this theme as a stack, not a single-stock story. Build baskets around equipment, services, and power-linked infrastructure. Then refine by lead times, backlog visibility, and margin structure.