Appendix D
Numbers Provenance & Forecast Register
A number with no date, no source, and no caveat is not a fact — it is a rumor with a decimal point; this appendix is the live, date-stamped register that turns every volatile headline figure in the guide into one you can re-check before you bet a slab, an interconnection slot, or a financing structure on it.
What you'll decide here
- How much weight to put on any single figure — read its as-of date and caveat first, and treat anything market/economic and more than ~6 months old as a direction, not a level.
- Whether a figure that drives one of your irreversible forks is a contested figure — if so, run the decision across the plausible range rather than picking a point estimate.
- Where to re-pull a live number before committing capital: open the live register, filter by category, and follow the source link to its primary publication.
- Which class a figure belongs to — durable (physics/standards), semi-durable (shipping hardware), or volatile (market/economics) — and therefore how it may be used.
- When to escalate: if you are underwriting an irreversible decision on a contested or stale figure, stop and re-verify against the primary source before signing.
This guide is full of numbers that move. Rack power, capex totals, GW of grid demand, $/GPU-hour, depreciation lives, TWh forecasts — these are the figures readers actually act on, and they are exactly the figures that go stale fastest. A printed reference freezes them at press time and quietly misleads everyone who reads it a year later. This appendix is the opposite discipline: a live, date-stamped provenance register in which every volatile headline figure in the guide is logged with its value, its as-of date, its source and a direct link, the scenario it assumes, and an honest caveat on how contested or fragile it is. The differentiator over frozen print: every figure that drives a decision is one click from its origin, so you can re-check it before you act.
Unlike the rest of the appendices, Appendix D is not primarily prose — it is realized as an interactive tool in the site: a searchable, category-filterable table backed by the same provenance dataset that powers the inline figures throughout the guide. This page explains what the tool does, how to read an entry, the provenance discipline behind it, and the one static reference table that belongs here permanently: the contested depreciation bound carried over from Chapter 1.8. For the live data, open the tool.
What the tool does
The register is a single source of truth for the volatile figures in this guide. Each entry is one metric — a value, the date it was true, the source, a link, the scenario it assumes, and a caveat. The tool lets you do three things at the bench:
- Search by metric, value, source, or caveat text — type "depreciation," "interconnection," or "HBM" and the table narrows live.
- Filter by category — Market & capex, Power & energy, Cooling & water, Compute & memory, Networking & optics, Economics, Reliability.
- Verify — every row carries an as-of date and a source link, so re-confirming a number against its primary publication is one click, not a research project.
The same dataset feeds the keynumbers blocks and inline citations throughout the chapters, so a figure you read in Part 4 and the same figure in the register are the one record. Update it once, and the guide stays internally consistent.
How to read an entry
Every entry carries six fields. Read them in this order, because the order encodes how much you are allowed to trust the number:
| Field | What it tells you | How to use it |
|---|---|---|
| Metric | The named quantity (e.g. 'GPU useful life, book vs economic') | Confirm it is the quantity you actually need, not a near-neighbor |
| Value | The figure, with units and any range | Prefer the range to the midpoint when the spread is wide |
| As-of | The vintage — the date the value was last true | If older than ~6 months and the figure is volatile, treat it as a direction, not a level |
| Scenario | The assumption set behind the value (base case, accelerated, announced-not-built, etc.) | A number is only valid inside its scenario — never lift it out |
| Source | The primary publication and a direct link | Click through before underwriting anything irreversible on it |
| Caveat | The honest note on fragility, contestation, or scope | A 'CONTESTED' flag means model the range, do not pick a point |
The provenance discipline
The register enforces one rule before a figure is allowed to drive a decision: classify it, then constrain how you may use it.
- Durable (physics, standards, thermodynamics) — design against the level. Air saturates near 41 kW/rack regardless of the year; ASHRAE envelopes are stable. → Chapter 0.3.
- Semi-durable (shipping hardware specs) — use the level, but confirm the part actually ships on your timeline. A roadmap rack power is not a procurable one.
- Volatile (market, capex, rental rates, forecasts) — use the direction, and re-pull a live number before committing. Cross this with the reversibility lens: a volatile figure driving a reversible decision is fine; a volatile figure driving an irreversible one is the trap.
Three classes of figure are flagged CONTESTED in the register because reasonable analysts disagree and the disagreement is itself load-bearing: GPU economic-versus-book life, HBM4 per-stack pricing, and hyperscaler depreciation policy. When a contested figure drives an irreversible fork, the correct move is not to pick a point estimate — it is to run the decision across the plausible range and confirm the branch is robust to where the number actually lands. → the discipline is set out in Chapter 1.8.
Static reference: the contested depreciation bound (from Chapter 1.8)
One figure-set is important enough, and contested enough, to pin here permanently rather than leave it to drift in the live table: the GPU useful-life debate bound out of Chapter 1.8. The accounting world books AI servers over a 5–6 year life; frontier economics observes a 2–3 year economic life after which a chip is displaced from the highest-value workloads. On the 1 GW reference model this single assumption swings annual TCO by ~70% — larger than any cooling, fabric, or siting optimization in the entire guide. The honest posture is to model both lives and disclose which one your covenants assume.
| Lever | Book / long-life view | Economic / short-life view | Consequence |
|---|---|---|---|
| IT useful life assumed | 5–6 yr accounting (some hyperscalers to 6 yr) | 2–3 yr frontier-economic | Sets the depreciation schedule — the dominant TCO lever |
| 1 GW reference annual TCO | ~$8.5B/yr at 5-yr (~$7B at 7-yr) | ~$12B/yr at 3-yr | ≈70% swing in annual cost from one assumption (Epoch AI) |
| GPU residual after 3 yr | Supports extended life (cascade-to-inference) | ~20–40% residual; rental rates −64–75% from peak | Residuals underwrite the long-life defense — and are themselves contested |
| Hyperscaler policy moves | Meta 4→5.5 yr (+$2.9B income); MSFT/Google 4→6 yr | Amazon 6→5 yr (−$700M); short-sellers argue obsolescence understated | Opposite moves by peers — no industry consensus |
| Honest engineering posture | Model reported earnings on book life | Model cash flows on economic life | Watch the gap — that is where stranded-asset risk hides |
Why a date-stamp beats a footnote
A conventional reference cites its source and stops. That tells you where a number came from but not when it was true — and in this market, when is everything. A capex forecast, a rental rate, or a queue length is a snapshot of a moving system; the same source publishes a different value six months later. The register's discipline is to make the vintage a first-class field, equal in prominence to the value itself, so a reader can instantly distinguish a figure that is still a level from one that has decayed into a mere direction.
The second move is the explicit caveat. Most numbers in this domain carry a scope condition (a scenario, a precision, a 'announced not built' qualifier) that, when stripped, turns a defensible figure into a misleading one. By forcing every entry to carry its caveat, the register makes the scope travel with the number. The CONTESTED flag is the strongest form of this: it is a standing instruction not to point-estimate. → the figures that bind to this register are surfaced in Chapter 1.8, with macro framing in Chapter 16.4 and 2030 scenarios in Chapter 16.5.