THS/02 · AI Infrastructure Thesis
AI load is structural, not cyclical.
Hyperscale and inference compute represent a permanent re-pricing of US grid demand — not a transient capex cycle. Underwriting this as a bubble materially mis-prices infrastructure risk.
Context
Even under aggressive efficiency assumptions, training and inference workloads imply a multi-decade step-change in industrial electricity demand. The question is no longer whether the demand materializes — it is which corridors can physically deliver power on the operator's timeline.
Positioning
- Concentrate exposure in jurisdictions with credible new generation pipelines (TX, OH, GA, AZ, VA).
- Discount markets where utility load forecasts have repeatedly understated actuals.
- Price land options against multi-scenario power-availability outcomes, not single-point estimates.
Implications
IMP/01
Tier-1 hyperscale corridors are functionally sold out through 2028.
IMP/02
Tier-2 and Tier-3 markets are repricing rapidly as Tier-1 deliverability tightens.
IMP/03
Behind-the-meter nuclear, gas peaking, and on-site storage become structurally embedded in the AI stack.
Watch
- › Hyperscaler capex guidance vs. actual deployments through 2027.
- › PJM and ERCOT large-load study queues.
- › Federal posture on co-located generation and grid-supportive interconnection.
Pressure-test this thesis against your portfolio or pipeline.