America Is Building Shells It Cannot Fill

The AI buildout has a dirty secret: the concrete is faster than the chips.

Jefferies ran satellite imagery across American soil and found datacenter construction accelerating on every axis — cranes, pads, transformer bays. The bottleneck isn’t permits or power, though both bite. It’s advanced chip packaging. TSMC’s CoWoS lines can only move so fast, and that constraint, according to Jefferies, caps US AI capacity deployment in 2027 at somewhere in the low 20s of gigawatts. Some of the more exuberant 2028 forecasts assume two to three times that. Nobody’s publicly updating the models.

Everyone says this is a supply problem. The opposite is closer to true: it’s a commitment problem wearing a supply costume.

Here’s the mechanism. Hyperscalers and co-location operators locked in construction contracts in 2024 and 2025 when chip delivery timelines were opaque and capital was cheap enough that overbooking felt rational. The buildings go up on 18-month cycles. The packaged GPUs arrive — or don’t — on TSMC’s schedule, which answers to physics, not capex plans. So you end up with gigawatts of shell: conditioned space, fiber, power interconnects, cooling loops, and no accelerators to fill them. The meter runs on the real estate. The revenue doesn’t.

KKR flagged exactly this dynamic in its credit markets warning this week, using words like “overexposure” and “concentration” — the bloodless language of a firm that has lent into the boom and is now watching the math. When a private credit giant starts hedging its own positioning in public, that’s a signal, not a hedge. They don’t publish warnings about sectors they’re comfortable holding through.

The deeper issue is what this does to the narrative. The AI capex supercycle has been the single most important story in global equity markets for three years. It rests on a chain of beliefs: demand is infinite, supply will catch up, and every dollar of construction spending eventually converts to inference revenue. The packaging bottleneck breaks link two. If capacity deployment stalls in the low 20s of gigawatts through 2027 while demand compounds, you get a two-year period where the hyperscalers can’t actually sell the compute they’ve promised — and the buildings they’ve built sit dark.

That’s not a catastrophe. It’s a timing mismatch. But timing mismatches are where the credit stress lives.

The inflation data this morning — core PCE up 0.2% month-over-month in August, consumer spending rising at the fastest pace in over a year — means the Fed is not riding to the rescue with rate cuts. The capital costs on those empty shells don’t compress. They compound.

Here’s the inversion the market hasn’t priced: the AI infrastructure buildout’s biggest short-term risk isn’t that demand fails to materialize. It’s that demand is real and the supply can’t meet it — and the financing was sized for a world where it could. The buildings don’t produce revenue while the chips are in queue. The debt service doesn’t pause.

Satellite imagery tells you how fast the shells go up. It doesn’t tell you when TSMC finishes packaging the brains.

The market rewards being early and still solvent when the world catches up. The question right now is whether the operators who poured the foundations in 2024 structured their debt to survive a two-year gap between the building and the machine. Some did. Some didn’t. The ones who didn’t are about to find out that concrete is not collateral when the tenant is a GPU that hasn’t shipped.