A week ago the tape looked pretty clean. AI semis working, connectivity working, power starting to work. Felt like the AI capex cycle was escaping the GPU rack and pulling the physical economy with it. Compute, networking, power, electrical and eventually machinery and finance. That was the Selective Capital Formation thesis.
Today the screen looks different. Not weaker in a dumb risk off way either. Breadth actually improved. The problem is what is leading. A bunch of the AI physical capex names disappeared like Marvell and got replaced by tankers like Delek, shipping, oil and existing logistical capacity. Basically the market went from buy the guys building the solution to the bottleneck to buy the guys who already own the bottleneck and that is a pretty meaningful change.
This is where scarcity gets nasty. Early on, scarcity is bullish capex. Price goes up, returns look great, everybody builds. But scarcity can last long enough that it starts breaking the investment response itself. Oil stays expensive, refined products stay tight, freight stays stupid, inflation refuses to fully die, rates stay high. Now the project gets hit twice: construction and operating costs are rising while the hurdle rate is rising too. That is Scarcity Crowding-Out.
And AI infra is basically the perfect victim because this whole thing needs absurd amounts of capital. Data centers, power, GPUs, networking, cooling, transformers, land, fiber, all while the 10y refuses to behave. The bear case is the next 100bn of capex starts looking worse than the last 100bn. Marginal ROIC down, WACC up. Eventually somebody says maybe we dont need to build every single campus tomorrow.
And markets will front run that way before earnings break. MRVL or ALAB or VRT can still print great numbers and get murdered if investors go from pricing 5 more years of accelerating capex to 2 or 3. These stocks arent just trading current earnings. They are trading the assumed duration of the buildout.
The really interesting part is software is still hanging around. DT, FFIV, DDOG, OKTA, ZS, FTNT, PLTR type stuff. Also, Meta is performing well, perhaps because Muse is showing good early indications. That opens the weird outcome where AI keeps winning while AI infrastructure stocks stop winning. The GPUs already built dont disappear. Maybe the trade simply starts moving from build the intelligence to actually monetize the intelligence. Less capex duration, more utilization.
But this is also where the setup gets asymmetric. If the bears are getting hyped because oil, refined product scarcity and high rates are choking the buildout, what happens if those conditions reverse before hyperscalers actually blink? Hormuz normalizes, freight collapses, oil falls, inflation pressure eases, 10y finally comes down… and MSFT/META/GOOG/AMZN are still saying yeah we need a stupid amount of compute.
Now the whole bear equation flips at once. Energy costs down. WACC down. Marginal projects start clearing again. Capex guidance never broke. Meanwhile the stocks already got derated because everyone convinced themselves the cycle was dying.
That would be the second leg.
The first leg was basically AI demand is bigger than we thought.
The second leg could be the macro constraint disappeared and these companies never stopped building.
So this is what I care about now. Not whether NVDA is green tomorrow. Watch whether the market stays in own existing scarcity mode or rotates back toward finance/build the solution to scarcity. If tankers/oil keep leading while AI physical infra like Denison keeps falling out of the screen, Scarcity Crowding-Out is getting real.
If energy/logistics normalize and MRVL/MTSI/ALAB/semicap/power infrastructure suddenly come roaring back, then this was just the interruption between leg one and leg two.
Leave a Reply