The hyperscalers are spending $400B+ on AI infrastructure in 2026, yet the consensus narrative treats this as a winner-take-all story between three or four well-known names. That framing is wrong, and the wrongness is creating the kind of dislocation I look for.
Three things the market is getting wrong
- Power, not chips, is the binding constraint. GPUs are abundant relative to substations.
- The second-order beneficiaries are bigger than the first-order ones. Cooling, switchgear, transformers, gas turbines — these are real businesses with real cashflows being repriced.
- The duration of capex matters more than the magnitude. A five-year buildout supports very different multiples than the twelve-month sugar high the market is currently extrapolating.
Where I’m looking
I’m focused on the picks-and-shovels names with real backlogs, real margins, and a customer base that extends beyond the top three hyperscalers. Specific names and entry points are in the full deep-dive PDFs — but the framework here is the lens through which I’m reading every print this quarter.