If you accept that hyperscaler capex is heading toward $750B and that nobody yet knows when it earns its return, the interesting question isn't whether to own AI. It's where in the chain you get paid on a schedule you can see.
The structure of the trade
| Layer | Gets paid when | Exposed to |
|---|---|---|
| Hyperscalers | AI revenue arrives | monetization timing |
| Chipmakers | orders are placed | capex cuts |
| Power / grid | capacity is contracted | construction schedule |
| Cooling / electrical | racks are built | construction schedule |
| Memory / storage | shipment | pricing cycle |
Why the middle of the chain is the interesting part
The hyperscalers carry the monetization risk — that's what July 23 repriced. The chipmakers carry order risk: if capex is cut, orders vanish first. But power contracts, grid equipment, transformers, electrical gear and cooling are consumed by the act of building, and the buildout is already committed years out. Those suppliers get paid whether the finished data center earns a good return or a poor one.
That's the toll booth. It doesn't have the upside of owning the winner. It has the property of getting paid during the phase when nobody knows who the winner is.
The mechanism
A gigawatt of data center capacity is a power problem before it's a compute problem. The binding constraint on AI expansion in 2026 isn't GPUs — it's interconnection queues, transformer lead times and available generation. When a constraint binds, pricing power moves to whoever relieves it. That's why electrical equipment and power names have re-rated on the same news flow that punished the spenders.
The risk
This is not a free lunch and I don't want to sell it as one. Three things break it.
First, these are cyclical industrial businesses being valued on a structural story — the same mistake people made with pipeline and mining suppliers in past capex booms. Second, much of the good news is already in the price; several of these names have re-rated hard, so you are no longer buying a boring utility multiple. Third, if capex is genuinely cut rather than merely questioned, the construction schedule shortens and the toll booth closes too — later than for the chipmakers, but it closes.
And there's a regulatory angle: utilities serving data centers face real political pressure over who pays for grid upgrades, consumers or hyperscalers. That's a live risk to the economics.
My take
I'd rather own the constraint than the ambition during a construction phase. Power and electrical infrastructure is where I'm comfortable being early, because the demand is contracted rather than forecast. I'd size it as a cyclical industrial position, not a secular growth one, and I'd avoid paying secular multiples for it — which is getting harder, and is the main reason I'm not adding aggressively at these levels.
Bottom line: during a buildout nobody can yet underwrite, the suppliers of the binding constraint get paid on the construction schedule. Power, grid and cooling are the toll booths — cyclical businesses, currently priced with a lot of optimism already in.
This is analysis, not investment advice.