# Demand Shock Model calculation contract

The useful result is a transparent stress experiment, not a claim that a particular company becomes insolvent in six to nine months. The supplied model supports conditional arithmetic. Its deck's headline probabilities are not reproducible from the supplied original model, and the companion's newly specified model answers different questions.

## Price, useful work and dollars

`revenue index = accepted-job index × tokens per accepted job index × net effective price / starting net price`.

Default source is `Consumption!C12:C16`: jobs 1.26440071556; intensity 1.1; total tokens 1.39084078712; $0.55/$0.68 price ratio; potential revenue index 1.12494475429. The job index comes from `(adoption/adoption0) × per-adopter task growth multiple × unblocked-share ratio × verified-success ratio`. This is a scenario construction; no population task count was measured.

For a simple graph use independent total-token volume and price controls. `R/R0=(Q/Q0)×(P/P0)`. A 35% price decline needs 1.53846× volume merely to keep revenue flat. A 65% decline needs 2.85714×. Four times as many tokens with a 90% price cut produces only 40% of original dollars (`Price_Regimes!E14`).

An optional elasticity curve uses `Q/Q0=independent_volume×(P/P0)^(-epsilon)`. Epsilon is the user's assumption: 0 means fixed volume, 1 preserves revenue, above 1 grows revenue when price falls. This does not estimate Jevons' response. Do not simultaneously interpret both a manually entered total-volume result and the elasticity result as independent additive growth.

The price control is net: caching, model routing and discounts already affect it. No second price haircut. For a separate rate-card explanation, `effective $/M = input_share×[(1-cache_of_input)×input_price + cache_of_input×cache_price] + output_share×output_price`. “90% of total tokens cached” and “90% of input tokens cached” have different denominators. Token price is not a GPU rental price, an electricity price, a matched-quality price index, or cost per accepted economic outcome.

Tokens per paid accepted output may fall with better models/retries/caching, or rise with reasoning. Changing this intensity and changing accepted work are distinct controls. Failed attempts, human review, integration and residual error costs must be included consistently in a genuine cost-per-accepted-output calculation. A benchmark score alone supplies none of those costs.

## Compute, capacity and utilization

`full-load capacity index = (1 + energized capacity growth) × performance per watt gain`.

Default is 1.4×1.5=2.1 (`Production!C9`). `required utilization = starting utilization × busy-compute index / capacity index`. At starting utilization 72% and compute 1.188430085, required utilization is 40.7461743% (`Production!C13`). The separately assumed 72% utilization is not multiplied into capacity again. This 31.25 percentage-point inconsistency is a model diagnostic, not proof of actual spare GPUs.

Cap served utilization at 100%, but preserve `unserved compute=max(0,compute−capacity/starting_utilization)`. The supplied bull case requires 685.96% utilization: its 14.697× revenue index is potential demand, not deliverable or recognized revenue. No capacity allocation among training/inference is modeled, so do not invent an exact capacity-capped revenue figure. Show “demand exceeds modeled capacity; realized sales not determined.”

Workbook compute conversion is tokens × relative model-tier energy × relative cache/output energy, then 70% inference and 30% training at the starting mix. These energy coefficients are assumptions. Independent busy-compute and token-revenue sliders can intentionally illustrate decoupling; label them as separate inputs, not a fitted causal system.

## Cash, commitments and runway

`net external funding need = cash capex + principal due − revenue×OCF margin` (`CoreWeave!B31:D31`). Defaults are no-shock $7.68bn, blip $14.58bn, full-shock $21.88bn. Revenue/margin/capex are scenario inputs; the $6.18bn principal input is inherited from the supplied June 2026 maturity schedule and is not a current roll-forward.

OCF means operating cash flow, including existing operating cash interest. Subtracting total interest again would double count it. A gross-contribution alternative instead uses `receipts×cash contribution margin−fixed operating cash costs−cash interest`; do not also apply an OCF margin in that route. Revenue recognition, invoice collection, prepayments and OCF can differ materially.

`cash[t]=cash[t−1]+OCF[t]+available executed funding[t]−cash capex[t]−principal[t]−incremental financing costs[t]`.

“Executed” funding means available under its conditions in that period. Optional financing, hoped-for parent support and backlog are separate and contribute zero until availability is established. Long-term contract face value cannot be divided arbitrarily by the stress horizon and called a fixed cash obligation. Use the actual payment schedule and avoid counting operating leases/cash interest already inside OCF again.

Show opening available cash, cash floor, financing dates and cash headroom. A negative floor result is a scenario liquidity gap; it is not a covenant breach, legal insolvency date, or actual six-to-nine-month runway. Unknown actual inputs are missing, not zero. `Quarterly_Bridge!B20:E22` shows that annual financing can balance at year-end yet arrive too late. Under its assumptions, Q1 floor breach occurs before Q2 negative cash; worst headroom is −$2.223060686bn in Q3.

For an optional all-in financing extension, the supplied workbook assumes every dollar of maturing principal is refinanced. Incremental refi interest is principal×(new rate−old rate)×half-year. New growth funding solves its own financing costs: `(capex−OCF+incremental_refi+refi_fees) / [1−debt_share×(new_rate×half_year+fee)]`, floored at zero. Gross funding equals principal plus growth funding. Blip becomes $15.257414248bn. Do not call this market funding secured. It does not price the whole debt stock at today's secondary yield.

## Power and valuation bridges

`quadrillion tokens/year = revenue $bn/year / $ per million tokens`; `IT TWh/year = quadrillion tokens×kWh/million tokens`; `average IT GW = TWh/8.76`; `facility GW = IT GW×PUE`. One quadrillion tokens at one kWh per million is one TWh, not one GW. At PUE1.4 the deck's 13–31% electricity share mixes facility load with a server-only denominator. Matched IT shares are 9.12–21.86% (`Token_to_Power!J7:J9`), before correcting revenue perimeter, training and idle power. None is a global demand estimate.

The $217.0867bn revenue requirement is `(32,563bn GDP×10% illustrative equity anchor)/15× sales`. The $117.2bn comparator adds $50bn OpenAI run-rate to four times Anthropic's reported $16.8bn quarter. Different accounting bases and vintages remain visible. It is not consolidated observed AI revenue, sell-side consensus or a cash-flow valuation.

## Macroeconomic scale and capture

An explicitly illustrative $30tn annual nominal GDP base ×10–20% yields $3–6tn incremental annual value added. At a separately assumed 10% provider capture share, $0.3–0.6tn of modeled value accrues to providers. Captured value is not automatically their revenue or profit. Final value added cannot be computed by adding customer sales, lab sales, cloud revenue and chip revenue. Consumer surplus can be real without appearing in measured GDP. Do not require all social gains to become provider revenue.

The supplied workbook's $32.563tn GDP anchor is preserved as an inherited source claim, separate from this $30tn illustration. The current-evidence worker reports an official reference: 2025 annual nominal GDP $30.8613tn, revised, in CEA/JEC Economic Indicators September2026 (data as of October7), PDFpage3/printedpage1, https://www.govinfo.gov/content/pkg/ECONI-2026-09/pdf/ECONI-2026-09.pdf. That baseline yields$3.08613–6.17226tn at10–20%. It is a later-vintage source than the accepted September27 historical layer. `macro_bridge.official_reference` preserves attribution and scope; this audit lane did not fetch that source independently.

## Probability display

Deck: 62% demand constraint, 40% joint shock, 12% systemic crash. Label these attributed V2 judgments; original complete model is absent. Reconstructed 250,000 draws: mechanical constraint 41.094%, joint proxy7.948%, thin-bridge proxy2.1816%; score proxy57.1544%. This is an alternate event definition, not a correction establishing the true probability. Tail proxy does not mean systemic crash.

The same assumed future price distribution yields14.6544% probability of at least35% decline from the current$0.68 reference, but93.604% from the historical$1.15 peak. The page must expose the reference baseline.

If displaying intervals, preserve their type: finite-simulation Wilson intervals, conditional draw quantiles, sensitivity across128 weight/threshold specifications, or deck's unreplicated ranges. None is empirical forecast coverage. Do not allow unrelated economic sliders to imply recomputed calibrated probabilities.