Section 6, Approval is not allocation: "A shell stays fungible only if its specification…"

[dave-dev650]

A shell stays fungible only if its specification (power per rack, cooling type) is written against the high scenario

From Section 6, Approval is not allocation · section anchor #6-2-supply-is-ordered-long-before-it-arrives-and-bought-through-contracts

While the entire document makes sense in the context of CPU, Storage or GPU planning, this sentence is transformatively interesting based on the past 2 years and will continue to be extremely important. In the past there was a power to space ratio that would give you reasonable fungibility and the growth curves for compute and storage worked well enough that you could make some good guesses more than a decade out. In order to accelerate GPU focused locations so the shell can land more quickly some very extreme positions are being taken around density while at the same time power consumption of GPU and CPU is exploding. The fungibility of these buildings in the future may be extremely limited and particularly awful at supporting storage, so we should expect a whole net set of constraints to show up.

[anacay]

Thank you, this is a real gap in that sentence. Here’s the proposal from this thread.

Change: add a caveat to Section 6.2. Extreme power density and liquid cooling may make new buildings far less reusable than the old open-hall designs, especially for storage. When the building shapes what can go inside it, the shell’s long lead time starts to shape the machines too.

Why: the old steady power-to-space ratio is what let a shell stay useful for whatever came next. Density built for speed trades that away.

Evidence: argued (practitioner review). Next, I’ll research public sources and cite what I find here before the caveat lands. If you know of building specs, reuse cases or operator disclosures, please point us to them.

Open for comment until 23 November: 14 days, doubled while there’s one editor. I’ll call it then and record the decision here.

[anacay]

The sources promised here, and what they changed.

The public record. For about two decades most racks drew 5 to 10 kW, and a shell designed once could be reused for whatever came next (survey figures as reported: https://mgrid.org/2026/02/13/data-center-pue-plateau-liquid-cooling-19-percent-adoption/). Current accelerator racks ship near 120 kW, and the vendor roadmap shows 600 kW racks for 2027 (https://www.datacenterdynamics.com/en/analysis/nvidia-gtc-jensen-huang-data-center-rack-density/). Retrofitting an air-cooled hall for liquid is possible but constrained — power, floor loading, pipework, live-site risk — and one analyst model finds it pays only with committed tenants and a price premium (https://stlpartners.com/articles/data-centres/liquid-cooled-data-centre-retrofits-economics/). The clearest single datapoint: one large operator cancelled a nearly-complete European expansion in late 2022 because the design could not be retooled for its AI-era needs, and rescoped its other projects in flight (https://www.datacenterdynamics.com/en/news/exclusive-after-meta-cancels-odense-data-center-expansion-other-projects-are-being-rescoped/). Above the building, grid connection queues add years of their own (https://www.iea.org/reports/electricity-grids-and-secure-energy-transitions/executive-summary).

What follows, argued. A machine order's real lead time is the lead time of the slowest non-fungible thing it needs. If the right hall type exists, that is weeks or quarters. If it does not, the order inherits the building's clock, and the building inherits the grid's. The commit-point ladder gains rungs above the ones the paper walks, and the top rung dominates.

Candidate caveat for 6.2 (for the 23 November call; wording open to this thread):

The AI era is spending shell fungibility fast. Racks that drew 5 to 10 kW for two decades now ship near 120 kW, with 600 kW on vendor roadmaps [sources above]. At least one large operator has cancelled a building rather than retool it [source above]. Where the right hall type does not exist, the building's commit point becomes the machine order's real lead time.

This grew past a caveat. The question now has a working note on the site: https://cpp.anacay.org/notes/buildings-at-ai-density/ — a draft, open for argument, and this thread is where it gets argued. The note applies results the paper already cites — chained flexibility [77], postponement [73], the option value of waiting [76] — one layer up, at the halls themselves. It also takes the storage point here seriously: if AI halls and storage halls diverge hardest, the chain breaks first at its quietest link.

Evidence status: the linked sources are published; the inheritance claim and the note's mechanisms are argued, not measured.