Vacancy is not the constraint

This week

1. Nvidia disclosed residual-value guarantees of up to $105B for OpenAI's Ohio campus, covering 4.25 GW initially with an option taking it to about 8 GW on a 20-year lease. First 800 MW in 2028.
2. PJM is weighing ride-through rules after nearly 3,800 MW of Northern Virginia data center load shifted to onsite generation on 22 July. Largest such event in PJM history. Recovery took nine minutes.
3. H.R. 10102 would impose a 1 cent per kWh excise tax on data centers above 1 MW, including power generated onsite.
4. North American vacancy sits at 1%, but Uptime estimates under 4% of US data centers can support 100 kW racks today.
5. PJM's first-ready cycle qualified 715 projects totalling 201.5 GW.

The read

Signal 1 will be read as 1999 all over again. Vendor financing, circular flows, Lucent and Nortel. That comparison is everywhere and it misses the part that matters.

In 1999 the collateral outlived the credit. Fiber and optical transport ran 15 to 25 years. The asset survived the borrower, which is why the overbuild was eventually absorbed.

Now it inverts. The lease runs 20 years. The silicon turns over every two to three.

The guarantee sits on land, power and shell, not on compute. Read that way, the rest stops being separate stories. Empty shell is worthless without density. Connected load is worthless if the grid operator cannot predict its behavior under fault. Generating your own power no longer exempts you from the bill. And a 200 GW queue is a statement about how long energized capacity stays scarce, not about how fast it arrives.

The scarce asset this decade is not GPUs or megawatts. It is energized megawatts, in shell that supports density, with predictable behavior under disturbance.

What it changes

Stop comparing capacity on $/kW. Three questions decide the real value of the contract: does the shell support the next density generation or only this one, what grid behavior is contracted under disturbance and who pays if regulation changes it, and is there an exit clause that survives a shift in rack density.

The question

How many of your AI infrastructure commitments guarantee the asset that lasts, and how many guarantee the one that will be obsolete before the contract is half done?