Appendix A1 · Context for §2 of the main paper · May 2026
The asset, at four scales
Section 2 prices two assets on one site: a durable shell and a wasting interior. This note shows the interior at four scales, and why the durable perimeter, not the interior, is the investable layer. Nothing here changes the decision.
The stack, from die to hall
Four scales of the same machine. Each row is built from the one above it, and the multipliers compound: eight chips to a server, eight servers to a rack, eighty or more racks to a hall.
| Scale | Composition | Power | Cost (USD) | Economic life |
|---|---|---|---|---|
| Chip · GPU | One accelerator; matrix arithmetic in parallel | ~700 W | 25–40 k | 3–4 yrs on training work |
| Server · node | Eight GPUs in one chassis, liquid-cooled at the chip | ~10 kW | 250–400 k | Turns over with the chips |
| Rack · cabinet | Eight servers; 64 GPUs per cabinet | ~80 kW | 2.5–3.5 m | Turns over with the chips |
| Hall · building | Eighty or more racks; grid feed, cooling water, fibre | 50–200 MW | 0.5–2 bn | Shell, substation and grid right do not age |
Source: AI_Datacentres_v3.docx Ch.6 (the stack, from die to hall); figures identical to main paper §6.B. Cooling capacity, not floor area, is what binds how many racks a hall can hold.
Main paper · The compact form of this table: §6.B
What the chips do, in two modes
The same hardware earns in two different shapes of work.
| Mode | Shape of the work | What binds it | Where it sits |
|---|---|---|---|
| Training | One cluster, one model, weeks to months. Tens of thousands of GPUs in one fabric, working as one machine | Network speed inside the building decides whether training converges | Frontier build; the highest capex per chip |
| Inference | Many small jobs, milliseconds each. The trained model copied across many smaller clusters, serving live queries | Proximity and availability | The production fleet, distributable across sites; where the operating revenue lives |
Source: v3 Ch.6. The four-scale stack and the two modes hold at any facility size; nothing in this note is specific to the boutique.
The investor read
An AI datacentre is a life-support system for chips. The compute layer ages on a three-to-four-year economic life; the shell, the substation and the grid connection right do not. The durable asset is the building and the connection. The wasting asset is what sits on the floor.
Section 2 of the main paper draws exactly this line between the two assets on the site, and prices Gowings into the durable side only.
Main paper · The two assets, and who owns which: §2