Capex Is Not Capacity

The number everybody quoted this month is a trillion dollars. Global data center capex for 2026 crossed it, the top four US cloud providers raised their combined spend something like 78 percent year over year, and roughly three-quarters of all of it is pointed at AI. The headlines treat that figure as a scoreboard — as if dollars committed were the same thing as compute delivered.

They aren't. And the gap between them is the only number worth watching right now.

Capex is a commitment. It's a board approving a budget, a hyperscaler signing a purchase order, a press release with a big round figure in it. What it is not is a running rack drawing power and serving tokens. Between the commitment and the running rack sits a chain of physical steps — land, permits, a utility interconnect, transformers, switchgear, the actual grid capacity to feed a campus that pulls like a small city. Every one of those steps has a lead time measured in years, set by physics and permitting, and none of them move faster because you wrote a bigger check.

So you get the split that defined this quarter. The capex line went vertical. At the same time, the buildout data says something close to half of the data centers expected to come online in 2026 will be delayed or canceled outright. Dozens of projects got shelved last year. PJM — the largest grid operator in the country — is projecting a six-gigawatt shortfall against its own reliability requirements by 2027. The Department of Energy quietly stood up a simulation platform this month just to model what happens to the grid when these campuses ramp from idle to full draw in seconds, because a hyperscale AI site behaves less like a data center and more like an aluminum smelter that switches on without warning.

Read those two stories together. Record money going in. Delivery slipping. That's not a contradiction. It's the signature of a fast variable racing a slow one.

The fast variable and the slow one

Capital is the fastest variable in this whole system. You can raise it, allocate it, and announce it inside a single earnings cycle. Silicon is nearly as fast now — GPU supply has actually loosened, and the Rubin ramp is on schedule for the back half of the year. The things money buys quickly are no longer the bottleneck.

The bottleneck is everything money can't hurry. A new utility interconnect is an 18-to-36-month negotiation. High-voltage transformers have lead times that turned into a running joke and then stopped being funny. Grid-scale generation doesn't get permitted on a product roadmap. These are slow variables, and they were slow before AI showed up — the buildout didn't create the constraint, it just drove a freight train into it.

When a fast variable chases a slow variable, the fast one piles up in front of the slow one. That pile is what a trillion dollars of capex looks like right now: capital stacking against a physical substrate that cannot absorb it on the timeline the spreadsheets assume. The money isn't wrong. The implied schedule is.

What the number actually tells you

Here's the part that should change how you read the headlines. Capex tells you about ambition. It's a real signal — nobody commits that kind of money on a whim, and the direction of travel is genuine. But it tells you almost nothing about capacity, and people are using it as a proxy for capacity because it's the number that's easy to get. It's printed in the 10-K. The megawatt figure isn't.

Strip it to first principles. What does an AI buildout produce? Not dollars spent — useful compute, energized and utilized. That's the output. Capex is an input, and a leading one by a couple of years. Grading the buildout by capex is grading a restaurant by how much it spent on the kitchen instead of how many meals it served. The spend is necessary. It is not the result, and treating it as the result is how you end up surprised when the result arrives late.

So if you actually want to know how the AI infrastructure story is going, stop counting dollars announced and start counting two things the announcements bury: gigawatts energized, and utilization on what's already standing. Those are the numbers that survive contact with physics. Everything upstream of them is intent.

None of this says the buildout fails. The capital is real, the demand is real, and most of those megawatts arrive eventually. It says the timeline is governed by the slow variable, not the fast one — and anyone planning against the capex headline is planning against the wrong clock. The dollar figure is the easy number to print. The megawatt figure is the one that's true.

A trillion dollars buys you a plan. The grid decides when it ships.

— Dustin