The five largest cloud companies are on course to spend about $805 billion on capital projects in 2026, up from about $449 billion in 2025, by Morgan Stanley’s count. Most of it is for AI. The question every investor is asking is whether that is a build-out or a bubble, and after working through the strongest version of the bull case I think the honest answer is that it is two different things stacked on top of each other.

The bull case was put well in September by Jon Gray of Blackstone, in a talk titled “Where’s the Beef?”. His argument: demand for AI computing is real and running ahead of supply. The limit is physical, in chips, power, data centers and permits. And unlike the railroad boom of the 1870s, this build-out is being paid for under contract by companies with strong balance sheets and little debt.

I checked the claims I could. Several hold up. GE Vernova’s gas turbines are reported sold out through 2030, which is what a real power shortage looks like. The usage figures from the model companies are enormous and rising.

Others were stretched. Utility capital spending was described as nearly doubling. The Edison Electric Institute’s figures point to an increase closer to 30 percent, about $1.1 trillion over five years. That is still a large number. It is not a doubling.

And some things the talk left out. Blackstone owns data centers and finances the chips that fill them, so it is describing a market it sells into. The combined revenue run rate of the two largest model labs was around $105 billion in July, against capital spending many times that. A railroad lasted fifty years. A graphics processor is written off in three to five. The comparison flatters the present.

So here is how I sort it. The layer I find credible is the physical one under long contracts: grid equipment, power generation, the utilities that serve it. Demand there is visible in order books. The layer I find fragile is the newest: lightly capitalized computing providers borrowing against chips that lose value quickly. If AI revenue arrives later than the spending assumes, that is where it will show first. Korean equities fell almost 19 percent in the third quarter as some of these trades unwound, a reminder that a low price-to-earnings ratio on a memory maker can mean the market expects earnings to peak.

I also have a local stake. I live in Loudoun County, which hosts one of the largest concentrations of data centers in the world, and the fights over where the next ones go, and what power lines they need, run straight through the rural west of the county where I ride and where I am trying to keep land open.

This is how I think about my own portfolio, not advice for anyone else’s.