Assumption Register — Appendix (Release Asset #13)
Companion to “Why Are They Spending Trillions on AI?” (Valto Loikkanen, CC BY 4.0)
What this appendix is. Every cost-curve, per-token, per-hour, per-member, and per-rack figure in this paper’s five scenario workbooks — Release Asset #7 (Global Baseline), #8 (Localized EUR/Finland), #9 (AI Working-Capacity Conversion), #10 (Token-Factory Scenarios), and #11 (Humanoid Working-Capacity) — rests on a small set of visible, editable inputs. This appendix pulls every one of those inputs into a single table so a reader can see, in one place, every lever they could change and re-run the model with their own number. (Release Asset #12, the Investment-Thesis Notes, is not a separate source of new assumptions — it explicitly cites Asset #7’s figures rather than re-deriving them, so it is covered here by reference, not with duplicate rows.)
What this appendix is not. It is not a ranking of which assumption is “right,” not a forecast, and not investment, legal, tax, procurement, or policy advice. Every row below is labelled SCENARIO ASSUMPTION per the paper’s five-class evidence standard — a visible, editable parameter used to explore a possible case, not a market figure — even in the handful of rows (marked ⚠ below) where the underlying workbook flags the number as a stand-in for missing public data rather than a deliberate “what-if” choice. Both kinds are assumptions in the sense that matters here: change the number, change every downstream result.
Evidence-class legend (for cross-reference; only SCENARIO ASSUMPTION rows appear in this appendix)
| Code | Meaning |
|---|---|
| OBSERVED FACT | Primary documentation/official spec or pricing, independently checked against a live/primary source on or immediately before 2026-08-12. |
| ATTRIBUTED STATEMENT | A named person/org’s on-record claim; not independently proven merely because it was said. |
| DERIVED CALCULATION | Transparent arithmetic from cited OBSERVED/ATTRIBUTED inputs; formula always shown. |
| SCENARIO ASSUMPTION | A visible, editable parameter, not a market figure. Every row in this register. |
| INTERPRETATION | A labelled explanation of how the above may connect; never presented as fact. |
How to use this table
- Find the assumption you want to change (by category or by which workbook/tier it feeds).
- Look up its current value and the formula context in the cited workbook section.
- Re-run that workbook’s formula (every formula is shown explicitly, never just a result) with your own number.
- Where a “suggested range to test” is given, it is drawn either from a sensitivity table the source workbook already built, or from the workbook’s own stated real-world range (e.g., “$0.03–$0.40/kWh exist across real global markets” — Asset #7’s own framing) — it is not a prediction of where the true value lies.
⚠ = the source workbook explicitly flags this number as a placeholder standing in for a public price/spec that could not be found (Professional-tier hardware, hyperscale rack capex, and the illustrative humanoid robot price), not a deliberate policy-style “what if” — treat these with lower confidence than the other rows, and replace with a real quote before using this tier for anything beyond illustration.
Consolidated Assumption Register
| # | Category | Assumption | Value(s) used | Workbook & location | What it affects | Suggested range to test |
|---|---|---|---|---|---|---|
| 1 | Financing | Down payment % | 20% for Home/Cooperative/Humanoid tiers (sensitivity: 0% / 40% tested); 0% (100% financed) for Professional and Hyperscale tiers, corrected 2026-08-13 — an earlier version of this register incorrectly stated 20% applied to all four tiers, but #7’s actual Professional (§4.1) and Hyperscale (§5.4) formulas finance the full capital amount with no down payment subtracted | #7 §2.1 (Home only); #7 §4.1, §5.4 (Professional/Hyperscale, 100% financed); #11 Parts 2–3.3; #8 §7a/7c | Split between financed vs. cash-paid capital → annual capital+financing cost on the tiers where a down payment actually applies (Home, Cooperative, Humanoid) | 0–50% for Home/Cooperative/Humanoid; 0% is the baseline (not a sensitivity test) for Professional/Hyperscale unless a reader chooses to model a down payment there |
| 2 | Financing | Financing interest rate | 8% APR (flat, illustrative) | #7 (all four tiers); #8 §1 (explicitly “not a verified Finnish loan rate”); #9 (inherited from #7); #10 §1 and §4.1; #11 Part 3; #12 (inherited) | Annual loan payment on every hardware/robot capital line, therefore every $/M-token and $/hour figure in every workbook | 4–15%; or your actual quoted rate — for Eurozone readers, #8 §2 suggests (current Euribor reference) + (bank margin, ~1–4pp) instead of a flat APR |
| 3 | Financing | Financing term | 5 years (sensitivity: 3 / 4 / 7 yr tested) | #7 §2.1 and §5.4 (Home, Hyperscale sensitivity tables); #10 §4.1 (industrial sensitivity table); #11 Part 3.2 | Size of the annuity payment — the single biggest capital-side lever after utilization; longer terms lower $/M-tokens but increase total interest and lock in today’s hardware generation longer | 3–10 years |
| 4 | Electricity | Home-tier electricity price (USD baseline) | Low $0.06 / Mid $0.12 / High $0.25 per kWh | #7 §2.2, §2.4 | Home-tier electricity cost line and $/M-tokens (though workbook notes this is a “rounding error” next to financing at home scale) | $0.03–$0.40/kWh (Asset #7’s own stated global range) |
| 5 | Electricity | Cooperative-tier electricity price (USD baseline) | $0.12/kWh (fixed point, reuses Home-tier mid value) | #7 §3.4 | Cooperative $/M-tokens and $/member/month | $0.06–$0.25/kWh |
| 6 | Electricity | Cooperative-tier electricity price (alternative recomputation) | ~$0.20/kWh | #10 §1b | Same DGX Spark cooperative cost model, different workbook — this differs from row 5’s $0.12/kWh, a genuine minor cross-workbook drift within the same hardware tier (both describe DGX Spark cooperatives), unrelated to the now-resolved €42-vs-€19.50–25 question (see row 6a and the cross-workbook notes below) | Re-test at $0.12 (Asset #7’s value) to isolate this variable’s own effect within the DGX Spark tier |
| 6a | Hardware/price placeholder ⚠ | DGX Station-class illustrative capital cost (the €42/member/month scenario’s actual basis) | €100,000 excluding VAT | #10 §1a; #8 §6 | The DGX Station-class cooperative scenario’s entire cost base — financing (§/month), and every downstream €/member/month figure for that tier | RESOLVES the former “€42 vs €19.50–25” entry (see row 19 and the cross-workbook notes below): this was always the real basis of the author’s originally published €42/50-members and €21/100-members figures — not a DGX Spark price, and not an official NVIDIA price, since NVIDIA publishes none for DGX Station (confirmed [FACT], nvidia.com/en-us/products/workstations/dgx-station/, checked 2026-08-13). Get a real reseller quote before using this tier beyond illustration; test against DGX Spark (row 1/#7 §2.1) as the alternative, cheaper hardware tier |
| 7 | Electricity | Professional-tier electricity price | $0.10/kWh (single point; no low/mid/high tested) | #7 §4.3 | Professional-tier $/M-tokens | $0.06–$0.25/kWh (Asset #7’s own Home-tier range, not separately tested for this tier) |
| 8 | Electricity | Hyperscale-tier electricity price | $0.06 / $0.10 / $0.25 per kWh (electricity-only sensitivity, §5.3); $0.10/kWh used as the central value in the main cost-curve table | #7 §5.3, §5.6; #10 §4 (industrial tier, same $0.10/kWh) | Hyperscale $/M-tokens — shown to be well under 1¢/M-tokens even at the high end, i.e. electricity is not the binding cost at this tier | $0.06–$0.25/kWh |
| 9 | Electricity | Home/prosumer electricity price, Finland/EUR localization | €0.15/kWh | #8 §1, §5 | Localized Home-tier €/M-tokens | Explicitly “not a verified current Finnish retail rate” — reader should substitute an actual bill/tariff |
| 10 | Electricity | Industrial electricity price, Finland/EUR localization | €0.09/kWh | #8 §1, §9 | Localized industrial-tier €/M-tokens | Same caveat as row 9 |
| 11 | Electricity | Humanoid-robot flat electricity rate | $0.11/operating hour (≈€0.10/hr) | #11 Part 4(a); #8 §7a/7c | Humanoid cost-per-hour, electricity line, “financing+electricity” columns | Compare against row 12’s spec-derived method — the two differ by roughly an order of magnitude |
| 12 | Electricity | Humanoid-robot spec-derived electricity price | $0.15/kWh, applied to 1X NEO’s observed 0.2105 kW average draw | #11 Part 4(b) | Humanoid cost-per-hour, “spec” columns | $0.10–$0.30/kWh; note the workbook applies this rate (verified only for 1X NEO’s power draw) as an illustrative anchor for the other three robot price points too |
| 13 | Utilization | Home-tier utilization | Low 10% / Mid 40% / High 80% | #7 §2.4–2.5 | $/M-tokens and $/AI-working-hour — the single biggest lever in the Home tier per Asset #7’s own read (≈7x swing from low to high) | 10–80%; test above 80% for near-continuous personal/small-team use |
| 14 | Utilization | Cooperative-tier utilization | Low 20% / Mid 50% / High 85% | #7 §3.4–3.6 | Cooperative $/M-tokens, $/member/month | 20–85% |
| 15 | Utilization | Cooperative-tier utilization (alternative) | 70% (“plausible steady-state,” not verified) | #10 §1b; #8 §6 | Same DGX Spark cooperative cost model, alternative assumption within that tier — unrelated to the now-resolved €42/DGX-Station-tier question | Sits inside row 14’s own 20–85% band |
| 16 | Utilization | Professional-tier utilization | Low 25% / Mid 60% / High 90% | #7 §4.3–4.4 | Professional-tier $/M-tokens | 25–90% |
| 17 | Utilization | Hyperscale-tier utilization | Low 25% / Mid 60% / High 90% (main tables); 90% used in facility-power calc | #7 §5.6, §5.8; #10 §4.1 | Hyperscale $/M-tokens, $/AI-working-hour | 25–90%+ |
| 18 | Utilization | Humanoid-robot annual productive hours | 2,000 / 4,000 / 6,000 / 8,000 h/yr | #11 Part 6; #8 §7a/7c | Cost-per-productive-hour — dominates the result more than any other single variable in the humanoid model (≈3.5–4x swing from 2,000→8,000 h/yr) | 1,000–8,760 h/yr (single-shift to fully continuous) |
| 19 | Overhead/Admin | Cooperative admin/overhead (coordination, basic networking/switch gear) | Low $1,000 / Mid $3,000 / High $6,000/yr | #7 §3.3–3.4 | DGX Spark cooperative $/M-tokens, $/member/month | $0–$15,000+/yr; note this row no longer needs to “explain” the former €42-vs-€19.50–25 gap (see row 6a — that gap was resolved as a different-hardware-tier question, not an overhead question) |
| 20 | Overhead/Admin | Hyperscale/industrial opex (facility, cooling beyond raw electricity, staffing, networking, replacement reserve) | Low $200,000 / Mid $500,000 / High $1,000,000 per rack/yr | #7 §5.5–5.6 | Hyperscale $/M-tokens | $100,000–$2,000,000/rack/yr |
| 21 | Overhead/Admin | Humanoid maintenance/repair reserve | 10% of purchase price/yr | #11 Part 5; #8 §7a/7c | Humanoid “incl. maintenance” cost-per-hour columns; not a manufacturer-quoted service contract for any of the four cited platforms | 5–20% of purchase price/yr |
| 22 | Overhead/Admin | “Full operating infrastructure” support/software/space overhead | 5%/yr of capital | #8 §8 (economic-layer illustration, Layer 4) | The “full operating infrastructure cost” layer in the localized Home-tier walk-through | 0–15%/yr of capital |
| 23 | Hardware/price placeholder ⚠ | Professional-tier (HGX B300 node) capital cost | Low $250,000 / Mid $350,000 / High $500,000 | #7 §4.1 | Entire Professional-tier $/M-tokens and $/AI-working-hour curve | No public NVIDIA/reseller price exists for this SKU as of the check date — the workbook’s own stated open item is to replace this with a real reseller quote before using this tier beyond illustration |
| 24 | Hardware/price placeholder ⚠ | Professional-tier power draw | Low 8kW / Mid 11kW / High 15kW | #7 §4.1 | Professional-tier electricity cost line | Scaled illustratively from GB300 NVL72’s 135kW/72-GPU ratio; no primary source — replace with vendor spec if available |
| 25 | Hardware/price placeholder ⚠ | Hyperscale rack hardware-only capex | $4,000,000/rack (→ $16M for the 4-rack/288-GPU cluster used in #10 and #12’s Tier 4) | #7 §5.4 (Reference Point B); #10 §4; #12 Tiers 3–4 | Hyperscale/industrial-tier capital+financing line, $/M-tokens, $/AI-working-hour | No official NVIDIA rack price found; test against the implied full-infrastructure figure of $6.75M–$8.1M/rack derived in #7 §5.4 (Reference Point A) from Huang/Fink’s on-air “$50–60B/GW” attributed statement, to see the spread between hardware-only and full-buildout cost |
| 26 | Hardware/price placeholder ⚠ | Illustrative humanoid robot price (author’s original worked example) | €25,000 (≈$27,000) | #11 Part 1; #8 §7a | Base-case humanoid capital layer, financing, and every downstream cost-per-hour figure | Not a real product price for any named platform — compare against the real observed/attributed market range assembled in #11 Part 1/7b: $13,500 (Unitree G1, OBSERVED) to ~$250,000 (Agility Digit, ATTRIBUTED) |
| 27 | Currency/FX | USD→EUR FX rate (three separate, non-identical values used) | 1 USD = 0.92 EUR (#7 §3.5); 1 USD = 0.93 EUR (#8 §1, §3, §9); 1.08 USD per 1 EUR, i.e. ≈0.926 EUR per USD (#11 Part 1, §7b) | #7 §3.5; #8 §1/§3/§9; #11 Part 1/§7b | Every EUR-denominated conversion in the cooperative (§3.5), localized (all of #8), and humanoid (#11 Part 1/7b) workbooks | The three values sit within ~1% of each other (EUR/USD ≈1.075–1.087 across all three), so this is a minor inconsistency rather than a large one — but none is stated as a live-checked rate on a specific date; pick one current, dated FX quote and use it consistently across every workbook if reconciling EUR figures precisely matters |
| 28 | Workload | Usage-intensity ladder (tokens/AI-working-hour) | Chat/advisor 10K–30K (mid 20K); Active coworker 60K–120K (mid 90K); Delegated single agent 200K–600K (mid 400K); Heavy multi-agent orchestration 1M–12M+ (mid 5M) | #7 §6 (ranges only); #9 Part B (adds explicit low/mid/high); #10 §6 (reuses #7’s ranges); #12 (reuses) | Every $/AI-working-hour figure in every workbook that has one — the conversion from token cost to hourly cost | Replace with your own organization’s measured token logs (#9 Part G’s own top recommendation) rather than the illustrative bands |
| 29 | Workload | Retail-API input/output token blend ratio | 30% input / 70% output | #9 §C.0 | The blended $/M-tokens figure used for each named retail model in the Retail-API comparison column (Floor $0.90 / Mid $9.00 / Ceiling $38.00) | Shift toward more input-heavy for chat-style workloads, more output-heavy for generation-heavy agentic workloads |
| 30 | Workload | Illustrative blended token rates used in the token-cost→working-hour bridge | ~$6/M tokens (retail, Sonnet-5-style blended); ~$1/M tokens (owned-production blended) | #10 §6 | The token-cost-to-working-hour example table in #10 §6 | Substitute the actual blended rate of whichever specific model/tier you are costing |
| 31 | Human-labor comparison | Entry-level/routine labor fully-loaded $/hour | $15–$35 | #9 §E.2 | Human-vs-AI cost-per-hour comparison table | Replace with your own local, role-specific fully-loaded cost (wage + benefits + overhead) — explicitly not a wage-survey figure |
| 32 | Human-labor comparison | Skilled professional fully-loaded $/hour | $50–$150 | #9 §E.2 | Same comparison table | Same as row 31 |
| 33 | Human-labor comparison | Senior specialist/expert consultant fully-loaded $/hour | $150–$500+ | #9 §E.2 | Same comparison table | Same as row 31 |
| 34 | Human-labor comparison | Standard full-time month, for methodology demonstration only | ~160 working-hours/month | #9 Part F | The illustrative “≈1,696x” conversion of Altman’s attributed “three months of work in seventeen minutes” line — explicitly a methodology demo, not a validated productivity multiplier | Not intended to be re-tested as a market figure; if reused, keep Part F’s explicit non-endorsement language attached |
| 35 | Energy-monetization comparison | ASIC efficiency bands (Bitcoin mining) | 25–38 J/TH (older fleets); <14 J/TH (efficient current fleets) | #10 §5 | Gross Bitcoin-mining revenue/MWh in the Bitcoin-vs-AI-tokens comparison table | Substitute an actual fleet’s spec sheet; these are illustrative generational bands, not a specific verified ASIC checked in this project |
| 36 | Energy-monetization comparison | AI-token-side hardware assumption for the comparison | 240W continuous draw, 50 tok/s output | #10 §5 | Tokens-per-MWh figure feeding the “AI tokens sold at $X/M” rows of the same comparison table | Within DGX Spark’s OBSERVED throughput envelope (28.3–51 tok/s, per #7 §2.3) — test across that full range |
| 37 | Energy-monetization comparison | AI-token retail sale price points used in the comparison | $1 / $5 / $10 / $25 per million tokens | #10 §5 | Gross revenue/MWh rows in the same comparison table | Swap for whichever retail tier (#7 §7 / #9 §C.0) you actually want to test — the workbook stresses the whole comparison is sensitive to this single input |
| 38 | Modelling choice | Per-rack throughput derivation method for GB300 NVL72 | 623,578 tok/s (even 4-way split of the 288-GPU/2.49M tok/s aggregate benchmark) used as the workbook’s primary number, vs. 707,112 (offline per-GPU-derived) or 580,608 (server per-GPU-derived) | #7 §5.2 | Hyperscale tokens/kWh and $/M-tokens | Test the full 580,608–707,112 tok/s bracket (~14% spread) rather than treating 623,578 as a fixed figure |
| 39 | Modelling choice | Down-payment amortization treatment | Spread evenly across the financing term and added back to the annual cost (chosen convention); treating it instead as a one-time year-0 cost is flagged as an equally legitimate alternative | #7 “How to adapt this workbook,” formula note | Every “capital+financing per year” figure across every tier and workbook | Compare annualized-spread vs. year-0 lump treatment for your own cash-flow purposes |
| 40 | Modelling choice | Hardware-price localization basis | US MSRP mechanically FX-converted (e.g., $4,699 → €4,370 “floor reference”), explicitly not a real EU/Finland reseller quote | #8 §3 | Every EUR-denominated capital figure in the localized workbook (Home §5, Cooperative §6, industrial §9) | Get a real EU reseller quote — the workbook’s own instruction is to expect a higher landed price than the FX-converted floor, not an equal one, once VAT/import/logistics/margin are added |
Cross-workbook consistency notes (INTERPRETATION)
- Most consistent lever: financing rate. 8% APR is used, unchanged, across every tier, every workbook, and both currencies (rows 2). This is the one input a reader can compare “apples to apples” across the whole paper without adjustment.
- Least consistent lever: electricity price. Nine different electricity-price values are in active use across the five workbooks (rows 4–12). This remains true of the DGX Spark cooperative figures specifically (rows 5 and 6, both describing that same hardware tier) — a reader comparing those two should still reconcile them before treating the outputs as directly comparable. It is not related to the separate, now-resolved DGX-Station-vs-DGX-Spark tier question below.
- CORRECTED (2026-08-13) — the former “€42-vs-€19.50–25/member/month cooperative discrepancy” is resolved, not merely narrowed. It was never a single measurement disagreement to attribute to admin, utilization, or electricity assumptions. €42/member/month (50 members) and €21/member/month (100 members) are [CALC] outputs of an illustrative €100,000 DGX-Station-class capital base (row 6a) — a real hardware category NVIDIA sells but does not publish a price for — while €19.50–25/member/month describes a completely different, cheaper hardware tier (10× DGX Spark units at the verified $4,699 each). See the Token-Factory Scenario Workbook (#10) §1a/§1b and the Localized Scenario Workbook (#8) §6 for the full corrected treatment. Rows 14–15 (utilization) and row 19 (admin/overhead) remain genuine, separate levers within the DGX Spark tier — they no longer need to explain a cross-tier gap that turned out not to exist.
- The FX-rate near-miss (row 27) is a minor but real inconsistency: three workbooks each picked their own EUR/USD rate independently rather than sharing one dated, live-checked figure. The values are close enough (within ~1%) that it does not materially change any conclusion, but it is exactly the kind of silent drift this register exists to surface.
- The five ⚠-flagged rows (23, 24, 25, 26, 6a) are qualitatively different from the rest of this table: they are not “what if we tried a different scenario” levers but placeholders for numbers that simply could not be found as an official price in any primary source as of the 2026-08-13 cut-off (no public HGX B300 price or power spec; no official NVIDIA rack price; no official NVIDIA DGX Station price, confirmed directly on NVIDIA’s own product page; no confirmed market price for the illustrative €25,000 robot). Any reader planning to act on the Professional, Hyperscale, or DGX-Station-class cooperative tiers, or on a specific humanoid-robot cost claim, should treat these five rows as the highest-priority items to replace with a real quote before relying on the workbook’s output.
Boundary note, restated per this project’s methodology rule: nothing in this appendix, or in any of the five workbooks it summarizes, is investment, legal, tax, procurement, or policy advice. Every row above is an editable illustration built on a stated, visible assumption — changing any one of them will change every downstream number in the corresponding workbook, which is the entire point of publishing this register rather than only the workbooks’ finished output tables.