# The Great Expansion: sources, measurement and reproducibility

Evidence revision R1. Evidence cutoff: 18 September 2026. First publication: 19 September 2026. Unified article and presentation revision R5: 20 September 2026.

Presentation revision R2 combines the previously separate data and research articles into one publication with Story, Research and Audit views. The evidence and full research manuscript remain at R1. Context photographs have a separate license and provenance register; they are not evidence about the audited firms' facilities.

Presentation revision R3 retains that structure and the same evidence snapshot. It adds fuller explanatory panels, optional motion with static/reduced-motion alternatives, two complementary accounting summaries and a linked D3 capital laboratory. The six original evidence figures remain available; the new forms do not replace their time paths or change the empirical classification.

Presentation revision R5 replaces the separate reading modes with one integrated article. The canonical authored source is `article.json`; the editable article Markdown is generated from the same paragraphs, concept definitions and figure interpretations. The original R1 research manuscript remains a labeled historical research source, not a second active article. Theory, actual financial observations, company episodes and source limits appear in one argument.

## What this release measures

This release combines three distinct evidence objects:

1. Annual company financial disclosures: cash purchases of property and equipment and reported revenue, in nominal US dollars.
2. A purposive audit of dated industry-participation episodes and infrastructure milestones.
3. Original research on the economics of organization and measured uses of AI.

They are not merged into a single integration index. A financial observation is a firm-fiscal-year; an industry case is an episode observed at its documented date; a study reports an outcome under its own research design. None is a substitute for the others.

## Source selection and retrieval

The supplied research dossier, revision 1 of 18 September 2026, provided the starting question and candidate sources. The pilot spreadsheets mentioned in that dossier were not supplied. This release therefore constructs a new evidence file from independently retrieved source material rather than recreating the dossier's unsupplied tables.

Preference is given to annual reports, regulatory releases, first-party announcements about the announcing entity, original research and institutional paper repositories. Company announcements establish what was announced; they do not independently establish subsequent commercial operation, market success or a causal effect.

The source register retains stable IDs, URLs, publishers, publication dates, retrieval dates and access notes. Each record cites one or more source IDs. Downloaded reports are not included simply because they were accessible: the public data contain factual extracts and original analytical coding, not a redistributed library of company reports or journal articles. Source-specific rights remain with their owners.

The cutoff applies to the underlying event or financial reporting period. A historical source may be retrieved after that cutoff. The latest full common fiscal-year comparison is distinguished from later company-specific disclosures. Calendar 2026 is not annualized.

## Financial series

The four-firm comparison covers Alphabet, Amazon, Microsoft and Meta. The selected measure is cash purchases of property and equipment, as defined in the cited cash-flow statement or accompanying disclosure. It is not management's prospective spending guidance, total purchase commitments, a lease-inclusive investment measure, or a measure of AI-specific capital.

Microsoft's fiscal year ends in June. The three peers' fiscal years end in December. A sum for fiscal year 2025 therefore combines overlapping but nonidentical twelve-month windows. It is labeled a fiscal-year cohort total throughout. It must not be used as a precisely calendar-aligned aggregate.

Amounts are retained in nominal USD. No inflation adjustment, currency conversion or seasonal adjustment is applied. The public CSV includes the period end, exact dollar amount, source identifiers and accounting caveat for every observation. Company-specific definitions, proceeds and financing arrangements need checking before interpreting differences as economic rather than accounting differences.

The disclosures describe the consolidated companies, including international activity. They are not US domestic investment or national-accounts observations. The industry episodes also span different geographic markets; a launch in one jurisdiction does not establish a global operating footprint.

Capital intensity is calculated as:

`100 * cash purchases of property and equipment / reported revenue`

Both numerator and denominator refer to the same company's same fiscal year. Positive revenue is required. A missing observation is not replaced with zero. The generator rejects duplicate firm-year keys and does not report a complete-cohort sum when a member is missing.

No structural-break coefficient, p-value or causal AI effect is estimated. An apparent change in the spending trajectory would not by itself identify the effect of AI, let alone a change in legal ownership boundaries.

### Retained financial calculations and historical companions

The accounting bridge begins with the FY2018 four-firm cash-PPE total and adds each firm's FY2025 minus FY2018 change. The four contributions sum to $293.395 billion and reconcile the starting $64.113 billion with $357.508 billion at the endpoint. Intermediate levels depend on display order; the chart does not assign causality or accumulate all spending during the eight-year observation window.

The paired endpoint summary compares each firm's cash-PPE/revenue ratio at FY2018 and FY2025. Its change labels are percentage points. The complete 32-observation panel remains available in the R5 capital-path instrument and exact-value table; the R3/R4 small-multiple and endpoint exports are preserved in their versioned packages.

The optional D3 laboratory places all 32 firm-year observations in a revenue-versus-cash-PPE coordinate system. Ratio guides satisfy `cash PPE = ratio * revenue`; they are exact accounting identities, not fitted lines, forecasts or uncertainty bands. Firm/year controls and the linked year brush select reported observations. A selected-firm trace, where shown, is only a chronological reading aid between annual records.

Financial selection is kept separate from the seven-firm industry audit and from the illustrative teaching model. Their query keys are independently validated and preserved in the article URL. A financial evidence-card export includes its selected record, period end, ratio and source identifiers; the boundary card retains its separate case-audit scope.

## The R5 comparative-cost teaching model

This is illustrative comparative-cost arithmetic, not a fitted economic model, company forecast, welfare calculation or estimate of an AI effect. The financial panel does not identify its inputs. All amounts below are arbitrary cost units per equivalent project, not dollars.

The internal arrangement has production cost 60, initial coordination cost 30, verification cost 10, expected delivery/liability cost 10 and an adjustable incremental fixed cost `F`. The market arrangement has purchase cost 70, initial contracting cost 25, verification cost 10, expected delivery/liability cost 5 and fixed cost 5. Non-coordination components are held at those stated assumptions, except for the explicit overhead control.

For assumed percentage reductions `s_I` and `s_M`:

```text
C_I = 60 + 30 * (1 - s_I / 100) + 10 + 10 + F
C_M = 70 + 25 * (1 - s_M / 100) + 10 + 5 + 5
```

Both savings controls accept integers from 0 to 100. `F` accepts integers from 0 to 30. The displayed private-cost rule favors making when `C_I < C_M`, buying when `C_M < C_I`, and records a tie at equality. The formulas do not estimate whether legal, financing, capacity, quality or strategic constraints permit the cheaper arrangement.

At `s_I=40`, `s_M=0`, `F=10`, costs are 108 and 115. At `s_I=40`, `s_M=40`, `F=10`, they are 108 and 105. With the latter savings but `F=0`, the costs are 98 and 105. The direction therefore depends on a comparison, not on whether internal coordination improved in isolation.

The default exact-value grid evaluates the two savings dimensions at 5-percentage-point increments: 21 by 21, or 441 deterministic model states. It is not a sample of 441 companies. Every exported row carries its illustrative truth status and unit. The break-even line follows the same equation; regions are not probabilities or confidence intervals. The implementation uses integer hundredths before conversion to cost units to avoid spurious floating-point inequality at indifference.

Model assumptions and results are exported separately from empirical records. A model evidence card records the selected inputs, all cost components, result and limitation. A case or company is never placed on this surface as if its coordination costs had been measured.

## Industry-participation coding

Seven focal firms are included: Alphabet, Amazon, Microsoft, Meta, NVIDIA, OpenAI and Anthropic. These firms were selected because they make the research question consequential, not by random sampling. The episodes were selected for their ability to distinguish organizational arrangements. Unequal coverage makes company rankings inappropriate.

Each episode retains its activity, sector description, event date, event type, stage interval, legacy/core flag, event-time status, any documented exit, sources, coding rationale and limitation.

| Stage | Meaning | Required distinction |
| --- | --- | --- |
| 0 | Technology supplier | Another business supplies the final industry service. |
| 1 | Task augmentation | A professional or organization retains responsibility while using the tool. |
| 2 | Intermediation | The platform connects, routes or facilitates a transaction. |
| 3 | Production | The firm produces a substantive industry output. |
| 4 | Operating participation | An owned business operates in the industry; entity structure and legal responsibility still matter. |

The categories are ordered descriptions, not equally spaced quantities and not an inevitable sequence of development. Four assistants are not treated as equivalent to one operating business. A lower and upper stage record unresolved classification. They do not describe a statistical confidence interval.

For a threshold `k`, the lower count is the number of selected cases whose lower stage is at least `k`. The upper count is the number whose upper stage is at least `k`. The denominator remains the selected episode collection as the threshold changes. These counts are not a funnel, conversion rate, sample estimate of industry prevalence, or active operating stock.

The interface can exclude episodes flagged as established legacy/core activity. That is a coding sensitivity check, not automatically a post-2022 comparison: legacy status and event date are separate fields. A failed venture remains a dated episode in this audit. Its closure is visible; it is never silently treated as continuing operation.

The firm's ownership boundary, the party producing an output, the intermediary directing a transaction and the entity bearing legal responsibility may differ. In particular, ownership of a services organization does not settle ownership of regulated professional entities, and a transaction interface does not make the platform the merchant.

## Milestones and venture exits

A milestone carries its own relationship type and status. Chip design is not fabrication; commissioning is not announcement; procurement is not ownership; minority investment is not automatically control. The selected milestones cannot reconstruct annual active ownership stocks.

The exit exhibit preserves documented dates without treating unlike events as comparable lifetimes. Stadia and Mixer enter the audit at their exit announcements; Sora enters at product launch. The later date is the documented end of the named service. These first-event meanings are labeled explicitly. Selected ventures are not the denominator of all attempted entries, and elapsed time is not compared as a survival or performance statistic.

## NVIDIA revenue composition

Data Center revenue and total revenue refer to NVIDIA's disclosed fiscal periods. The derived share is Data Center revenue divided by total revenue. The remainder is calculated by subtraction, and the generator rejects a component larger than the total.

Data Center is a company-defined revenue platform, not a NAICS industry, an ownership share of customers, a value-added share or a direct measure of monopoly power. A concentrated revenue mix is used as a counterexample to the necessity of broad industry operation, not as proof of narrow technological capabilities.

## Literature and causal inference

The literature register separates theory, evidence reviews, observational studies, randomized experiments and working papers. Effects measured in customer support, product-development teams or a national labor market are not pooled as though they estimate a common parameter.

The central unresolved causal claim is whether more upstream integration produces later entry into additional industries. Cash, distribution, managerial capability, anticipated demand and a pre-existing expansion strategy can affect both. Lagging an integration variable does not remove anticipation. Seven selected firms and a few post-2022 years do not support confident population or causal inference.

The integrated article proposes a historical firm-sector risk set, observed entries and exits, frozen definitions and explicit alternative explanations. It does not claim that this research design has already been estimated.

## Reproduce and inspect

Within the repository, run:

```powershell
npx tsx scripts\great-expansion\generate.ts
npx vitest run tests\great-expansion
```

The generator reads the reviewed evidence snapshot, checks its SHA-256 against the source manifest, validates schema and provenance, and writes deterministic JSON, CSV, exact-value tables, audit totals, visual assets and a public checksum manifest. Re-running against the same reviewed inputs must produce the same bytes. `--data-only` builds the analytical files without visual exports and marks that output as a data-only build, not a completed publication release.

For a clean R2 media rebuild, first run `npm run great-expansion:media`. That command verifies the reviewed photograph hashes and licenses, then produces fingerprinted landscape and portrait WebP crops. The main data generator subsequently rebuilds the matching figure and cover exports. R1 visual exports are retained rather than silently relabeled as R2.

The reviewed snapshot is assembled by `npx tsx scripts\great-expansion\assemble-evidence.ts` from three versioned factual-extract files. That step converts USD millions to USD, reconciles Meta's gross and net presentations, checks all eight independently tabulated fiscal-year totals, and deduplicates identical source URLs. Changing a factual extract changes the reviewed-input hash and requires another editorial review; a same-input reassembly must reproduce the accepted bytes.

For an independent arithmetic check, download `evidence.json` and `reproduce.mjs` into the same directory and run `node reproduce.mjs` with Node 22 or later. It requires no packages and prints the fiscal-year totals, revenue intensities and stage-threshold counts.

To reproduce the separate teaching example, also download `model-assumptions.json` and run `node reproduce.mjs --model`. Its labeled model output is added separately from the empirical totals; no string formula is evaluated as code.

Every quantitative exhibit has an exact-value download. The portable evidence card includes the selected population, threshold, specimen, revision, cutoff and measurement boundary. A deep link restores the same state; it does not fetch a newer data vintage.

## Corrections

A correction requires source reconciliation and a new reviewed input hash, regenerated assets and updated revision metadata. A newly announced project must not overwrite an earlier observation as if operation had been continuous. Unsupported claims are withheld rather than filled with illustrative values.

This publication is an economic research interpretation, not investment, legal or medical advice.