A RAAK.work project
Capex Watch← The Trillion-Dollar Capex Bet
Updated: 2026-10-13
What this page is for. Capex Watch makes claims with numbers in them: that five companies must eventually earn a certain amount to justify what they are spending now, that a gap of hundreds of billions of dollars a year has opened between the two, that labour's share of income is the lowest since 1947. This page shows where each of those numbers came from and how it was worked out. It is here so you can check the site instead of trusting it.
Every figure is one of three kinds, and each entry says which. Reported means it comes from a company's own filing or from an official statistics release, and the entry names the document and the date. From the BIS means it comes from the Bank for International Settlements' Annual Economic Report 2026, the report this whole site is built around, and the entry cites the page or graph. Worked out here means our own arithmetic, and the entry shows the sum so you can redo it. Where our method differs from the BIS's, the entry says so and says why, and the site repeats the warning wherever that number appears.
The site stays inside the BIS's model. That report tracks five companies: Alphabet, Amazon, Meta, Microsoft and Oracle. Others appear on the site as evidence beside the model, never inside its totals, and their entries say so.
The numbers the charts run on are checked by machine. They are listed as data at the foot of this page. The site cannot be published while that list and the live pages disagree, so a figure cannot quietly drift away from the entry that explains it.
If you think something here is wrong, this page is what lets you say which number and where the reasoning breaks. That is why it is published.
BIS citations below are to the Annual Economic Report 2026: "Ch I p. N" (chapter pagination), Graph 11 (pp. 22–23, notes p. 35), Box C (pp. 20–21, Graph C1).
| Page claim | Source & status |
|---|---|
| Five companies; over a trillion dollars of AI capex across 2025–26 | BIS Ch I p. 22: the five largest hyperscalers are set to spend over $1T on AI-related capex from 2025 through 2026. The five are named in the Graph 11.A notes (p. 35): Alphabet, Amazon, Meta, Microsoft, Oracle. Fixed 2026-07-30: the page previously said "a trillion dollars a year" — the BIS figure is two years combined. 2026 guides alone total ≈$840B/yr (see §7 split; Amazon raised 30 Jul), so the run-rate approaches but has not reached $1T/yr. |
| Spending outpaces earnings and free cash flow; gap covered by debt | BIS Ch I p. 22 and Graph 11.A (capex-to-revenue ratio + debt issuance). Sourced. |
| Canal mania / railway mania / dot-com parallels | BIS Ch I pp. 22–23 and Graph 11.C. Sourced. (Graph 11.C notes date the AI boom's capex trough to 2023 — that date anchors the §7 actuals baseline.) |
| "The five is the BIS's frame, not a census" — SpaceX as a sixth builder (added 2026-08-07) | Named scope note, stated on the page. The BIS names five hyperscalers; this page keeps that model intact and adds no sixth slice to the §7 ramp, gap-chart split or capex total. SpaceX is carried as evidence only, in the §1 note, the §5 circuit, the §3 third strip's worked case, the strain-chart marker and one watch-list row — every one of them labelled "not one of the BIS's five". The capex-to-revenue comparison is a page construction from the two companies' own releases: SpaceX Q2 2026 capex $18,369M ÷ revenue $7,814M = 2.35×; Oracle Q4 FY26 capex $16.5B ÷ revenue $19.2B = 0.86×, the highest of the five. Same ratio as BIS Graph 11.A (capex-to-revenue), applied to a firm the BIS did not model. |
| Page claim | Source & status |
|---|---|
| Floor = capex/life + $250B opex + required return × capex | Page construction. The BIS's own sustainability test (Graph 11.B, Rungcharoenkitkul 2026b contest model) is different machinery: net economic surplus = revenue − capex − debt service, under "AI delivers" vs "AI disappoints" scenarios. Our floor annualises the same committed capex and adds an investors' required return instead of subtracting realised debt service — a stricter test (vindicated, not merely survived). Both ask the BIS's question: can revenue justify the committed capex? Annotated on the page. |
| Capex slider: $3–4.5T committed 2025–30, default $3.5T | BIS Graph 11.B's x-axis is cumulative AI capex committed 2025–30, range $2.0–4.5T, with the Nvidia-CEO 2030 projection ($3–4T) marked. Default $3.5T = that projection's midpoint. The page's slider starts at $3.0T, not the BIS's $2.0T: ≈$1.3T is already spent or guided for 2025–26 alone (BIS's >$1T plus the 2026 guides), so a sub-$3T outcome by 2030 would require the build-out to nearly stop — we cut the dead range. Named deviation, justified. |
| Newer forecasts run above the slider's range (sentence under the §2 slider, added 2026-10-05, owner's decision): "analysts expect the five to spend about $1.1T in 2027, and spending that only held at that level would pass $4.5T in 2029" | The slider is unchanged and keeps the BIS's range; the sentence reports one attributed forecast and one page construction beside it. BIS check, read at source: Ch I says the five "are set to spend over a trillion US dollars on AI-related capital expenditure from 2025 through 2026", and Graph 11.B's x-axis is "AI capex committed (cumulative 2025–30, USD trn)", running 2.0 to 4.5, with note b marking the Nvidia chief executive's 2030 projection of $3–4trn. The graph gives no path by year, so the BIS itself does not say when the top of its range would be reached. Attributed: $1.1T for 2027 is the analysts' consensus as cited by Goldman Sachs ("More AI capex, more volatility", 25 Sept 2026), confirmed twice in the §8 row for FT Alphaville's piece. It is the five's capital spending as a whole, which is also what the page's $840B for 2026 counts. Page construction: the page's own "about $1.3T already spent or planned for 2025–26" (previous row) plus $1.1T a year from 2027 is $2.4T, $3.5T, $4.6T and $5.7T at the ends of 2027 to 2030, so the slider's default is reached at the end of 2028 and its top passed in 2029. Holding spending level is the cautious reading: Goldman's own forecast is above consensus and the consensus chart Apollo published on 21 Sept rises each year to 2030. Registered as spendPath; the checker recomputes the two years from it and cpxSlider and holds the sentence, the §8 entry and its chart to them. Why the range is not widened: it is the BIS's, and the page's model is the BIS's; widening it would be a named deviation that moves the headline floor (each extra $1T adds $300B a year), on the strength of forecasts for years that no company has yet guided. The sentence says the floor is kind to the bet on this count, as §3 and §4 say of the buyers' bill. |
| $250B/yr fixed operating costs | Page assumption (electricity + operations, order of magnitude). Not a BIS figure. Held constant across the slider range; stated in the §2 legend. |
| Floor is static; the BIS's model is amplified (audit note, 2026-07-31) | Named limitation, stated on the page (§2 note): the floor is a linear bill — a 20% revenue miss reads as a 20% shortfall. The BIS's contest model (Graph 11.B, with cross-firm/circular financing and debt) is a financed system in which a miss repricing debt and withdrawing financing deepens the miss — outcomes are nonlinear, and the delivers/disappoints gap widens with committed capex. The same wiring runs upward in booms (equity-stake mark-ups booked as profit — Amazon's $53.4B Anthropic gain). The page keeps the legible static test and points to §5 for the amplified dynamics. |
| §2 floor expressed as a multiple of Microsoft's Office franchise (added 2026-09-03) | Page construction, and a scale comparator only — it divides the §2 floor and feeds nothing. Denominator: Microsoft's Productivity and Business Processes segment (Microsoft 365, Office, LinkedIn, Dynamics), trailing twelve months to 31 Mar 2026 = $135.3B, built from Microsoft's own reported figures as nine months FY26 ($102,149M, quarter ended 31 Mar 2026) + Q4 FY25 ($33,112M) = $135,261M. Confirmed twice per the standing rule: Microsoft's FY26 Q3 press release of 29 Apr 2026 (segment revenue $35,013M for the quarter, +17% yoy, with the nine-month and prior-year columns) is the primary source, and Futurum's independent write-up of the same print agrees at $35.0B and +17%; geoveu holds no record of a Microsoft segment breakout, so the primary plus one independent report stands in, as the rule provides. At the page's default sliders the floor is $1,300B/yr, so the multiple is 1300 ÷ 135.3 = ≈9.6×. Computed live in floorCalc() from OFFICE_TTM_B, so it tracks the sliders rather than going stale; the segment figure itself moves only when Microsoft reports. Chosen as a comparator because the objection it answers is not "can a software business succeed" but "has one ever been built at this rate" — the franchise is the most successful of its kind and took four decades to reach this run-rate. |
| Floor vs annuity cross-check (audit note, 2026-07-31) | The floor charges the required return on gross cumulative capex; a textbook annuity earns it on the declining balance: annuity-equivalent floor at the defaults = $3.5T × [0.10 ÷ (1 − 1.10⁻⁵)] + $0.25T ≈ $1.17T/yr versus the page's $1.30T — the page's construction is ≈11% stricter. Neither is the BIS's formula (the BIS tests net economic surplus, Graph 11.B). Kept deliberately on the strict side; recorded here so the bias is visible. |
| Hardware life 3–6 years | Page slider. Consistent with BIS Ch I's depreciation discussion (AI hardware earns over a few years, not decades); exact range is ours. |
The buyers' bill: a second measure beside the floor (added 2026-09-30, owner's decision; on its own page, bill.html, since 2026-10-03) | Page construction, beyond the BIS's model. bill = floor × [ d + (1 − d) ÷ (s × (1 − m)) ], where the floor is the §2 floor (on bill.html it is held at the essay's slider defaults through FLOOR_DEFAULTS, as debt.html's is), d is the share of the builders' computing sold straight to the final customer, s is computing as a share of a product firm's costs, and m is the product firms' profit per dollar of sales. The builders receive the floor either way; the product firms' customers pay rent ÷ (s × (1 − m)) for the (1 − d) share that is rented. At the defaults since 22 Oct 2026: $1,300B × [0.17 + 0.83 ÷ (0.75 × 0.90)] = $1,300B × 1.400 = $1,819.5B ≈ $1.82T a year, split as $1,300B reaching the builders, $360B of product firms' other costs (0.83 × 1,300 × (1 ÷ 0.75 − 1)) and $160B of product firms' profit (10% of their $1,599B of sales). Against the §3 software pool: $1.82T ÷ $1.2T = 1.5×. (From 3 to 21 Oct the direct share was 16%, giving $1,826B ≈ $1.83T; Meta's narrowed 2026 plan moved it, see the row on d.) (From 30 Sept to 3 Oct the margin default was 30%, giving $2,288B ≈ $2.29T and 1.9×.) With s = 100% and m = 0 the bracket is 1 and the bill equals the floor, so the floor is this measure with the middle layer removed. BIS check (Ch I pp. 22–25, Graph 11.B and its notes p. 35, Graph 13.C): the BIS tests the hyperscalers' own revenue against capex and debt service; AI labs appear as counterparties in cross-firm (circular) financing and as debt issuers, never as a layer that marks up the builders' revenue on its way to a final buyer. The measure therefore adds to the BIS's model and replaces nothing in it. It uses no company outside the five. |
| Buyers' bill: the three sliders and their defaults | s = 75% (slider 50–100%) is Goldman Sachs's, from the application-layer table in "More AI capex, more volatility" (25 Sept 2026) as reproduced by FT Alphaville on 28 Sept 2026: "compute represents 75% of application layer operating expenses", with EBIT margins of 10–40% across the columns. m = 10% (slider 0–40%) is the bottom of that range, the default since 3 Oct 2026 (owner's decision; it was 30%, the margin the FT's account reads first). The defaults are chosen to be cautious, meaning towards the end of the evidence that gives the smaller bill, now that the bill feeds the headline, §3 and §4: the bill shrinks as s rises, as m falls and as d rises. The one lab whose accounts have been reported points both ways (see the next row): its computing share, 58%, would raise the bill to $2.30T, so 75% is the cautious value; its reported Q2 2026 margin, about 5%, would lower it to $1.74T, but that is one forecast quarter excluding stock-based pay, so the page keeps 10%. Goldman's note is single-sourced (it is not public), labelled as Goldman's on the page, and on sliders for that reason. d = 17% (slider 0–50%; 16% until 22 Oct 2026) is a page construction that follows Goldman's treatment: its table applies no margin to Meta, so the default is Meta's share of the five's announced 2026 capex from the registry, $140B ÷ $842.5B = 16.6%, rounded to the slider's whole percent: 17% since 22 Oct 2026, when Meta narrowed its plan to $135–145B (it was $137.5B ÷ $840B = 16.4%, so 16%, from 3 Oct). The method check that follows was run at 16% and is left as run; at 17% each figure moves by under 1%. The checker fails if that share no longer rounds to the slider's default, so a change to the capex split moves the default or stops the deploy. Method check against Goldman's own figures: with its $308B break-even requirement and this page's 16% sold direct, the formula gives $433B, $480B, $542B and $624B at margins of 10%, 20%, 30% and 40%, against the table's 0%-return row of $425B, $475B, $525B and $625B. The largest difference is $17B, on a table that appears to be rounded to $25B. That shows the formula is the right shape. Extended to all sixteen cells on 2026-10-05: dividing each cell by the formula's bracket (16% direct, 75% computing, that column's margin) gives one builders' requirement per row to within the table's rounding: about $303B, $522B, $741B and $956B for returns of 0%, 10%, 20% and 30%, with at most $13B between the four cells of a row. So the formula fits the whole table. The rows step by about $218B for each ten points of return, twice the chart's $109B ($308B, $417B, $526B, $636B), which is what the two exhibits' titles say: the chart charges the return on "average annual" capex in 2026 and 2027 and the table on "cumulative" capex in those years. A 10% return therefore means $417B a year for the builders in the chart and about $520B in the table, and the §8 entry now says so. The step sizes imply bases of about $1.09T and $2.18T if the return is charged on capex one-for-one, a little above the $1.0T the §8 row works out from Bloomberg's report of Goldman's 2027 forecast; that difference is not resolved here. Goldman's dollar figures are not used on the page: its chart is sized on the average of 2026 and 2027 capex and its table on the two years together, where the floor is sized on 2025–30. |
| Buyers' bill: what it feeds (decision record, 2026-10-03 morning, replacing that of 2026-09-30; itself replaced that evening, next row) | The headline, §3 and §4, through a switch; never the tracking page or the debt page. On 30 Sept the owner decided the bill should feed nothing. On 3 Oct the owner reversed that, for two reasons. §3 and §4 set AI revenue against what customers spend (the software pool, the wage pool), and the capture rate in §4 is the share of the saving kept by whoever sells to the customer, so for those sections the floor understated the requirement. And Anthropic's draft prospectus, as reported by Reuters (not an official filing), gave a second view of the middle layer, beside Goldman's note (next row). The bill is now the default measure. One function, requiredT() in index.html, tells the stat strip, §3's drawing and §4's starting value which measure is in force: the bill at its default (BILL_T = 1.83, i.e. $1,826B rounded as drawn; §4's slider starts at 1.8, its step being 0.1, giving $7.2T of wages, 33% of the $22T pool and 90M jobs-equivalent at $80k, where the floor gave $5.2T, 24% and 65M), or, in the BIS model, the floor (FLOOR_T = 1.3). A switch at the head of §2 chooses; ?model=bis opens the page in the BIS model; the browser remembers the last choice. §3 and §4 use the fixed default values, not the live §2 sliders, as §3 always has. The floor stays on the tracking page (required ramp, gap chart, clocks, strain chart) and the debt page, because both measure the builders' own revenue, cash and debts; holding those against what final customers pay would compare two different things and break comparison with every print recorded since July. The checker fails a deploy that brings the bill's code to either page. The 30 Sept objections are met as follows: the headline no longer rests on one bank note alone (the S-1 figures are a second, labelled source); the two questions keep separate names and the switch keeps them apart; the requirement the prints are measured against is unchanged. The design not taken: (1) a frozen copy of the BIS tool on its own page, bis.html, with the front page moved to the bill. It is the most literal way to keep the original model, but it copies the floor's code, and two copies drift (as fifteen pages' copies of the palette did before 11 Sept); a switch keeps one copy and makes the BIS view a setting. (2) The bill everywhere, tracking page included: one number site-wide is simpler to explain, but it measures the five's revenue against a requirement on their customers. |
| Buyers' bill: its own page (decision record, 2026-10-03 evening, replacing the morning's) | Nothing on the front page, the tracking page or the debt page; it lives on bill.html. The owner's decision, the same day the bill became the default: with two measures, a switch and a long panel, the front page had become harder to follow than its argument. The front page is the BIS model again: the stat strip, §3's drawing (FLOOR_T = 1.3) and §4's starting value (1.3) run on the floor, and §4 reads $5.2T, 24% and 65M again. The bill panel moves whole to bill.html: the three sliders and their defaults (unchanged), the cautious-defaults note, the method note, the Anthropic check and what the bill leaves out. That page adds the jobs arithmetic run on the bill, beside the floor (next section). The front page keeps one sentence in §2 stating the bill at its defaults rounded to $0.1T ("about $1.8T a year for the builders to receive $1.3T"), and one pointer each in §3 and §4 saying the floor is kind to the bet because the pools and the capture rate are about what customers pay. The checker holds that sentence to measuresT, holds bill.html's FLOOR_DEFAULTS to the essay's sliders, and now fails a deploy that brings the bill's code to index.html as well as to tracking.html and debt.html. The switch, requiredT(), BILL_T and ?model=bis are removed; requiredT in the registry becomes measuresT, with the same two values. What it costs: the morning's reason still stands. §3 sets the floor against pools of customer spending and §4's capture rate divides what customers pay, so on the front page both understate the requirement by the bill's multiple, 1.40× at the defaults. The page says so in plain words where it matters rather than drawing the larger figure. The design not taken: keep the bill on the front page, folded into a closed box below the floor, with the floor as the only measure and no switch. It touches fewer files and needs no new page. It lost because it removes the switch but leaves the panel and its long notes in the front page's code and length, and a reader who opens the box is back to two measures on one page, which is what the owner wanted gone. The morning's design not taken, a separate bis.html, is not this one: that copied the floor's code to a second page, where this keeps one floor on the front page and puts only the mark-up elsewhere, reading the floor's defaults through the same checked FLOOR_DEFAULTS pattern as debt.html. |
| Anthropic's draft prospectus: the check on the bill's settings, and the tracking page's Anthropic row (added 2026-10-03) | Reported, not filed: a named exception to the two-confirmation rule (owner's decision, 2026-10-03). Anthropic's S-1 is a confidential draft; on 3 Oct 2026 SEC EDGAR held no public registration statement from Anthropic. Every figure below comes from Reuters' report of 28 Sept 2026, which says Reuters saw the document; Fortune, CNBC, InvestmentNews and Capital Brief repeat Reuters and did not see it. That is one confirmation, so each use on the site carries the label "From Anthropic's draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing." The exception lapses when the public S-1 is filed; its figures then replace these. Figures: 2025 revenue $4.59B (2024: $386M); computing and infrastructure $7.33B; total costs $12.65B ($4.59B − $12.65B = −$8.06B, which is the reported operating loss, so the $12.65B includes cost of revenue); operating loss $8.06B (2024: $2.98B); net loss about $42B, of which about $34B is a fair-value charge on financing convertible into shares; Q2 2026 revenue $11.5B, the second quarter in a row with an operating profit (amount not reported). Commitments $518B: Google $111.1B (Apr 2026–Jul 2033), Amazon $110B (May 2026–Apr 2036), Microsoft $31.4B (Nov 2026–May 2033), Broadcom equipment leases $161.2B, xAI up to $84.5B (mostly cancellable on 90 days' notice), AMD up to $20B; the parts sum to $518.2B, which agrees with the total. Page arithmetic: computing share of costs = 7.33 ÷ 12.65 = 58%; to the BIS five = 111.1 + 110 + 31.4 = $252.5B, 49% of $518B, "less than half". The 58% includes training as well as serving customers, and some went to suppliers outside the five, so it is an approximation of s, not a measurement. Separately sourced, not from the prospectus: a Q2 2026 operating profit of $559M on $10.9B of revenue, excluding stock-based pay, a company forecast reported by the Wall Street Journal in May 2026 (≈5% of sales); labelled on the page as the WSJ's. Bill at the alternative settings (floor $1,300B, d = 17% since 22 Oct 2026): s = 58%, m = 10% → $1,300B × [0.17 + 0.83 ÷ (0.58 × 0.90)] = $2,288B ≈ $2.29T; s = 75%, m = 5% → $1,300B × [0.17 + 0.83 ÷ (0.75 × 0.95)] = $1,735B ≈ $1.74T. (At d = 16% until then: $2,300B and $1,741B.) |
| Page claim | Source & status |
|---|---|
| Global enterprise software ≈ $1.2T/yr | Industry estimate (Gartner-scale worldwide enterprise software spending). Not a BIS figure; cited in the footer. Registered 2026-09-30 as softwarePoolT: it was load-bearing in §3's geometry all along and now also divides the §2 buyers' bill, so the checker holds SOFTWARE_T to it. |
| White-collar wage pool: slider $20–35T, default $25T (revised 2026-07-31; rebased 2026-10-03) | Page derivation: advanced-economy GDP $69.6T (IMF NGDPD 2025, October 2026 round, the aeGdpT registry value; see "Output base" below) × labour share ≈ 60% (BIS-consistent baseline) = $41.8T labour income, "about $42T" on the page. The white-collar fraction is definition-dependent — ≈50–55% counting occupations, ≈60–70% weighting by wages — so the pool is on a slider ($20–35T ≈ 48–84% of labour income) with the default $25T ≈ 60% ($41.8T × 0.6 = $25.1T), the middle of the estimate range. The October 2026 round moved the base from $69.4T to $69.6T, which leaves the default at $25T. Before 3 Oct 2026 the base was $60T, the pool $36T, the slider $18–30T and the default $22T; the slider's proportions are unchanged. The page launched at a fixed $30T (83% — the range's most generous end for the thesis: a bigger pool makes displacement look smaller); the audit flagged it and the slider replaced it. The BIS publishes the labour share, not a white-collar split. §4's share readout uses the same slider. |
| §3 drawn-to-scale geometry (bug fixed 2026-07-31) | The software-pool rectangle had been drawn at 136×26 px — the width from an area-correct 136×64 shape, but with the height of the strip layout — i.e. ≈0.49T-equivalent area instead of 1.2T. That exaggerated the floor-vs-software comparison ≈2.7× (the true margin of the $1.3T floor over the $1.2T pool is thin). All §3 shapes are now computed from the pool slider at one px-per-$T scale, so "drawn to true scale" is enforced by the code rather than by hand geometry. |
| The floor must draw on one pool or the other, or on growth (third source added 2026-10-04) | Page framing of the BIS's tool-vs-substitution distinction (Box C contrasts tool-like general-purpose technologies with AI competing directly with human cognition, Ch I p. 21). The two pools are money that exists today: software budgets and wages. Since 2026-10-04 the page names a further source, growth: if AI raises what people produce, AI revenue can be paid from output that would not otherwise exist, with no wages displaced. This is the BIS's own "bounded productivity boost" scenario (Box C: trend growth permanently higher, labour share flat), which the page's earlier wording ("only two pools") left out. Why the link from spending to replacement is the page's, not the BIS's: Chapter I's capex analysis (pp. 22–23, Graph 11) tests whether returns cover the investment and never mentions wages; Box C's scenarios differ by the scope of automation, with growth and the labour share both outcomes. The inference "vindicated spending means replaced work" is Adam Tooze's and this page's §3–§4 arithmetic. Tooze states the either/or on Ones and Tooze, "The AI Arms Race" (Foreign Policy, 10 July 2026): either the investment is vindicated and the labour market is upended, or it is not and investors take losses; Chartbook 456 (5 July 2026) puts the same link in writing, that market wealth rides on a vision in which the balance between capital and labour is upended. Growth is the outcome that framing omits; the next row sizes it. |
| §3's growth figures: customers must gain $5.2T a year by 2030; about 3.4% growth a year against 2.1% | Restated from the Growth Saviour's growth-or-wages chart (see its rows below): $1.3T floor ÷ 25% capture = $5.2T; found entirely from output above today's pace by 2030, the advanced economies must grow 3.44% a year, "about 3.4%" (October 2026 round; 3.2% against 1.8% before). The checker holds both figures to that chart's arithmetic. |
| Third strip: the builders' own 2026 AI capex, $0.84T/yr (added 2026-08-07) | Page construction, drawn at the same px-per-$T scale as the other two shapes and registered as aiCapexPoolT. The value is the same $840B that splits the §7 gap band — the five's announced 2026 capex (see the §7 row) — expressed as a pool, so the two figures cannot drift: the checker fails if AI_CAPEX_T × 1000 ≠ capexShareTotalB. Scale check: $0.84T ÷ $1.2T = 70% of the global enterprise-software market. Its status is deliberately not a third pool the floor can be paid from — capex is the bill the floor services, so revenue drawn from it is the bet paying itself; the page says so in the same breath as it draws the strip. It earns its place because it is where a rising share of disclosed AI revenue now demonstrably comes from (§5's circuit), and because omitting it would let the tool-vs-worker dichotomy imply that today's revenue must already be from one pool or the other. This is the revenue-side counterpart to the funding-side circular financing the BIS documents (Ch I pp. 22–23, Graph 11.B notes) — the extension to the revenue side is this page's, not the BIS's. |
| SpaceX as the worked case for the third strip (added 2026-08-07) | Company figures from the SpaceX Q2 2026 earnings release (SEC EX-99.1, 4 Aug 2026): AI-segment revenue $2,561M, +247% yoy, of which $2,194M is "AI solutions & infrastructure"; $14.1B of Cloud Services Agreement contracted sales, contributing $1.6B of incremental Q2 AI infrastructure revenue; nameplate compute 1.4 GW, from 0.4 GW a year earlier. Contract terms are press-reported, not filed, and labelled as such on the page: Anthropic at ≈$1.25B/month with ≥300 MW at the Memphis Colossus 1 site (announced May 2026), Alphabet at ≈$920M/month from October 2026 to June 2029 (announced June 2026), reported by CNBC/TechCrunch and Reuters via Yahoo Finance. Page arithmetic: $1.25B + $0.92B = $2.17B/month = ≈$26B/yr, against SpaceX's own AI-segment revenue annualised from Q2 of ≈$10.2B. GPU counts differ between reports (≈220k vs ≈325k for the Anthropic deal) and are therefore not used on the page; megawatts and monthly payments, which agree across sources, are. |
| Page claim | Source & status |
|---|---|
| Wages displaced = required revenue ÷ capture rate | Page construction — an identity given the capture-rate definition; every input on a slider. |
| The capture rate: 25% default, and the evidence for it (added 2026-10-04) | Page assumption, now set beside evidence; each figure confirmed twice. The rate is the AI seller's revenue as a share of the customer's gain (a saved wage or new output). Prices where AI is sold by the job done: Salesforce Agentforce launched at $2 per conversation (Salesforce press release, 12 Sept 2024, "standard volume discounts apply"; Salesforce Ben, 6 Nov 2024); Intercom's Fin AI agent is $0.99 per outcome (Intercom's pricing page, read 4 Oct 2026; independent pricing guides agree). Cost of the human alternative: live service channels (phone, chat, email) cost an average of $8.01 per contact in Gartner's 2019 Customer Service and Support Leader poll (Gartner press release, 25 Sept 2019; DestinationCRM, 26 Nov 2019). Page arithmetic: 2 ÷ 8.01 = 25.0%; 0.99 ÷ 8.01 = 12.4%, "about 25% and 12%". Caveats, stated in the FAQ: the cost figure is from 2019, so today's ratio is probably lower; a contact is not the same unit as a resolved query; list prices before discounts; Salesforce added per-action pricing in 2025; one use case, customer service, chosen because per-job pricing makes the comparison possible. History: William Nordhaus, "Schumpeterian Profits in the American Economy: Theory and Measurement", NBER Working Paper 10433, April 2004: for the US non-farm business economy 1948–2001, innovators captured about 2.2% of the total present value of the social returns to innovation, from a first-year appropriability of 7% (paper; NBER Digest, Oct 2004). Reading: 25% sits at the top of the observed range, so it is generous to the AI sellers, and every figure built on it (wages displaced, growth needed) is more likely understated than overstated. Sensitivity: at 10%, the floor needs 1.3 ÷ 0.10 = $13T a year. The three prices are registered as captureEvidence; the checker recomputes the two ratios and the $13T wherever the pages state them. |
| Required-revenue slider starts at the floor, $1.3T | Page construction. At the defaults (25% capture, $25T pool, $80k a job): 1.3 ÷ 0.25 = $5.2T, 5.2 ÷ 25 = 21%, $5.2T ÷ $80k = 65M. (With the $22T pool before the output base moved to $69.4T on 3 Oct: 24%.) The checker holds the default to measuresT.floor. History: on the morning of 3 Oct 2026 the slider started at the buyers' bill, 1.8 (BILL_T 1.83 rounded to the slider's $0.1T step), giving $7.2T, 33% and 90M; that evening the bill moved to bill.html and the slider returned to the floor. The text beside it says the floor is kind to the bet, because the capture rate divides what customers pay, and points to bill.html, which runs the arithmetic on the bill. |
| China-shock benchmark ≈ 1–2M US manufacturing jobs | Standard literature estimate (Autor–Dorn–Hanson range). Not BIS. |
| Page claim | Source & status |
|---|---|
| Bust mechanism: repricing, circular-financing unwind, capex cut as recession mechanism | BIS Ch I pp. 22–23; Graph 11.B notes define the cross-firm/circular financing and the financial-unwind scenario. Sourced. |
| Labour shock: labour share ~60% → ≈20% by around 2060 | Box C Graph C1.B, transformative-AI path (simulation runs 2030–2060); Box C text says the share "falls towards zero" in the long run — 20% is the graph's end-of-horizon reading. Fixed 2026-07-30: page previously said "over four decades"/"by ~2065"; the BIS horizon is 2060, ≈ −11 pts per decade. |
| "Euthanasia of the labouring classes" (attribution corrected 2026-10-04) | The phrase is William Nordhaus's (2021), adapting Keynes's "euthanasia of the rentier", and Adam Tooze cites it: Chartbook 451, 6 June 2026 ("euthanasia of the labouring classes", attributed to Nordhaus 2021), and Ones and Tooze, "The AI Arms Race", 10 July 2026, where he attributes it to Nordhaus beside the BIS's 20% scenario. Until 4 Oct the page attributed the phrase to Tooze himself, as "euthanasia of the working class", citing Chartbook 456; that post carries his gloss of the BIS's 20% scenario but not the phrase. Attributed external commentary, not BIS. |
| Demand-bottleneck twist: displaced workers are lost consumers; growth ends below trend | Box C pp. 20–21: under the demand bottleneck, output growth rises before falling below its historical trend; the consumer base erodes as automation diverts wage income. Sourced (page wording is a close paraphrase). |
| Fork chart paths | Stylised after Box C Graph C1.A (labelled stylised on the page). 2% historical trend is the BIS's business-as-usual growth rate. |
| The circuit diagram — circular financing with figures on every arrow (added 2026-08-07) | Page construction illustrating a BIS mechanism, annotated as such on the page: the BIS defines the cross-firm and circular financing (Ch I pp. 22–23; Graph 11.B notes) but publishes no such diagram. Arrow provenance, left to right: (1) Amazon's equity stake in Anthropic and the $53.4B mark-up within its $62.6B Q2 2026 net income — Amazon's Q2 release, already sourced in the §7 row; (2) Anthropic → SpaceX at ≈$1.25B/month for ≥300 MW, press-reported terms (see the §3 row), not a filing; (3) Alphabet → SpaceX at ≈$920M/month from October, same status; (4) SpaceX → Nvidia, $15,828M of AI-segment capex in Q2 2026, from the SpaceX release — the destination of that capex is not itemised in the release, and Nvidia's position as the supplier is the page's existing framing of its watch-list row, not a disclosed SpaceX vendor breakdown. The return leg is deliberately drawn dashed and labelled "not cash but a price": the loop closes through a mark-to-market, not a payment. The soft-backlog point uses the release's own Note 2, which defines contracted sales as covering only the non-cancellable, enforceable period — which is why $14.1B of booked contracted sales sits far below the reported headline value of the same agreements. |
| Page claim | Source & status |
|---|---|
| Dial machinery (rebuilt 2026-10-05): what customers must gain by 2030, split between growth and wages | Page construction on the bet's own numbers; no new assumption. The same arithmetic as the Growth Saviour's second chart (its rows below carry the derivation and the BIS check), on the front page and with §4's sliders as its input. What customers must gain, V, is §4's figure: the floor ÷ the capture rate, $1.3T ÷ 25% = $5.2T at the sliders' starting values. Output in 2030 at growth g is 69.6 × (1 + g)⁵; at today's pace, the IMF's 2.1% for 2026 (October 2026 round; it was 1.8% in July), it is 69.6 × 1.021⁵ = $77.22T. Growth pays for min(V, output − 77.22); the rest comes out of wages. Workers' pay in 2030 = 60% × output − the wages replaced; their share is that ÷ output; "total pay, change a year" is (pay ÷ 41.76)^(1/5) − 1, where 41.76 = 60% × 69.6. At today's pace that is −0.3% a year, and the share is 53%. Workers keep 60% of the larger economy, following Box C's bounded-productivity-boost scenario, in which the labour share does not move. All growth above today's pace is credited to AI, which is generous to growth, and the page says so. All dollar figures are in today's prices (constant prices): the growth rate is real (after inflation), so the levels it compounds are real too. The front page rounds output to $69.6T where the Growth Saviour uses its data's $69.605T; no displayed figure differs. The slider runs from today's pace, 2.1%, to 4%, and starts at today's pace (registry growthDial). The gauge's zones, its two ticks and the key under it are computed from the three rates in the row after next, and move with §4's sliders. |
| The dial runs to 2030, on the bet's numbers (decision record, 2026-10-05, owner's decision) | Replaces the fifteen-year dial built on a doubling of capital's income. That dial fixed capital's income at twice today's after fifteen years (on a slider since 4 Oct) and gave workers the rest. Its readings (workers' share 49% at 3% growth, wages keeping pace at 3.2% growth for fifteen years, each point of growth worth 7–8 points of share) followed from the doubling, and the doubling was this page's illustration: nothing in the spending set it, and Box C has no dollar amounts to take it from. The floor falls due in 2030, so a reader who set the dial beside §2 to §4 was comparing two clocks, and the fifteen-year charts suggested that growth must stay high for fifteen years, where the bet needs a level of output by 2030. Removed with it: the slider for capital's multiple and its registry entry (capitalMultiple15), the four fifteen-year sparklines, the FAQ item on the doubling, the "horizon bridge" between a fifteen-year dial and a 2030 floor, and the Growth Saviour's button that extended its chart fifteen years at 3% a year. The design not taken: keep the fifteen-year dial and label it more firmly as an illustration. It lost because the label was already there, since 4 Oct, and the charts still misled: an illustration that carries a section's headline claim is read as the claim. A record of 4 Oct had turned down rebuilding the dial on the $5.2T, because on a fifteen-year horizon it would have duplicated the Growth Saviour's chart with a horizon the floor does not have. On the 2030 horizon the overlap is intended: the front page's argument now ends on the trade the Growth Saviour sets out in detail. The cost is one piece of arithmetic written on two pages; the checker holds their constants equal and recomputes both pages' figures. Also turned down on 4 Oct, and still: recalibrating the doubling to 1.76× against the BIS's 20% by 2060, which paired the BIS's most extreme scenario with today's growth, as neither the BIS nor Tooze does. |
| Output base: today's economy, $69.6T (owner's decision, 2026-10-03; October 2026 round since 2026-10-22) | IMF-sourced; the registry's aeGdpT. The IMF's 2025 GDP for the advanced economies in current dollars at market exchange rates (WEO October 2026, ADVEC NGDPD = $69,605.3B; April 2026 had $69,386.7B, $69.4T), the same figure economies.html shows; scripts/check-derivations.py holds aeGdpT, the §6 code's Y0 and the §3 prose to the economies page's data, so a refresh of that data fails the deploy until the front page follows. BIS check (Box C, Ch I pp. 20–21): the BIS's model has no dollar level of output; it works in growth rates and shares, with the pre-AI labour share "matched to data". So no BIS figure moves, and the base is a page construction on an IMF input. Since 2026-10-05 the dial sets a dollar figure from the spending ($5.2T) against this base, so its rates depend on the base: a larger economy makes the same sum a smaller share of output. The design not taken: the economy before the boom, 2022, which the BIS's narrative dates from ("following the release of generative AI tools in late 2022", Ch I). In today's prices that is $65.7T (2022's output in 2025 dollars, from the economies page's data). It lost because every use would need to say "2022 in today's prices", while §3 and §4 compare today's revenue requirement with today's pay and §6 starts "today". Keeping $60T was not an option either way: it is 2022 output in 2022 dollars, beside figures in 2025–26 dollars. |
| Readouts at the dial's start (g = 2.1%, October 2026 round): growth pays for $0.0T, wages for $5.2T (65M jobs' worth), workers' share 53%, total pay −0.3% a year | Output 77.22; pay = 0.6 × 77.22 − 5.2 = $41.13T; share 41.13 ÷ 77.22 = 53.3%; (41.13 ÷ 41.76)^(1/5) − 1 = −0.30% a year; 5.2 ÷ $80k = 65M. (At July's 1.8%: share 53.1%, −0.64% a year.) ✓ |
| The three rates: total pay stops falling at 2.2%, grows as it used to at 2.9%, and growth pays for all of it at 3.4% (October 2026 round; 2.0%, 2.7% and 3.2% before) | While some of V still comes out of wages, pay = 0.6 × out − (V − (out − 77.22)) = 1.6 × out − V − 77.22, so the output at which total pay grows w a year is (41.76 × (1 + w)⁵ + 5.2 + 77.22) ÷ 1.6. w = 0: out = $77.61T, g = 2.20%. w = 2%, the long-run trend on Box C's business-as-usual path, taken here as total pay's old pace: out = $80.33T, g = 2.91%; there growth pays for $3.11T and wages for $2.09T (26M jobs' worth), and the share is 57.4%, "near 57%", "about two and a half points" down (the checker words the drop to the nearest half point). All of V from growth: ((77.22 + 5.2) ÷ 69.6)^(1/5) − 1 = 3.44%, 1.34 points above today's pace, the claim's "about 1.3 percentage points". The level reached ($82.4T against $77.2T) and what a tenth of a point is worth (4.7M jobs' worth) are the Growth Saviour's rows. ✓ |
| "Why 2.7% is seductive" | At 2.73% growth total pay grows 2.0% a year while workers' share falls from 60% to 57.9% and $1.67T a year, 21M jobs' worth, comes out of wages (row above). "Outside the rebound after Covid, the advanced economies have not reached it in any year since 2016": in the reality chart's data the group's rate is at most 2.6% (2017) outside 2021 (6.1%) and 2022 (3.1%). The reading that a loss of this kind would go unnoticed is this page's own. |
| No-substitution scenarios: labour share never moves at any growth rate | Box C: business-as-usual and bounded-productivity-boost scenarios keep labour the binding input. Sourced. The right-hand end of the dial (growth pays for all of it, the share stays at 60%) is the bounded-productivity-boost case. |
| OPEN QUESTION: the 60% labour-share intercept (raised 2026-09-03) | §4 derives labour income as GDP × 60%, and §6 starts workers at 60% of output; both start from the same intercept (the GDP base, $69.4T since 2026-10-03, is registered as aeGdpT; the 60% is not in the machine-checked registry — it lives as LS0 in the §6 dial and in §4's prose note). Two findings sit against it. (1) It is a model parameter, not a measurement: Box C says the model economy's labour share before transformative AI is matched to data and established estimates, so the BIS publishes no advanced-economy labour-share series for this page to track. Earlier ledger wording called it a "BIS-consistent baseline", which is defensible; describing it as the BIS's published figure would not be. (2) On the level-comparable whole-economy measure (see the AMECO row in §7) no developed bloc reads 60% — euro area ~57%, EU ~56%, US ~54%, Germany ~60% alone. Moving the intercept would shift every §4 pool readout and the whole §6 dial, so it is the owner's decision, not a print's arithmetic (BIS-fidelity rule, point 3). Left at 60% pending that decision, with the observed series carried beside the model instead. If it stays, this row is the justification: the page's job in §6 is to mirror Box C's model, not to re-estimate it. |
| Horizon: the dial stops in 2030; the long run is the forty-year chart (2026-10-05) | The dial and the §2 floor now share one date, the end of the BIS's capex window, so the seam this row named from 2026-07-31 (a fifteen-year dial beside a 2030 floor) is gone. The report's two horizons are both still on the page: Chapter I's boom to 2030 is the dial, and Box C's long transformation is the forty-year chart, presented as resting on a much stronger assumption (the transformative branch, where AI keeps taking over new tasks). "At worst about seven points": 60% − 53.1%. |
| Reality-check chart: a decade of AE growth, US vs rest (added 2026-07-31) | US and advanced-economies annual real GDP growth 2016–2026 from the IMF World Economic Outlook datamapper (series NGDP_RPCH, codes USA and ADVEC, October 2026 round, fetched 2026-10-22; 2026 is the WEO projection. The July 2026 vintage it replaced had AE 2026 at 1.8%; October has 2.1%, and revised 2021, 2022, 2024 and 2025 by 0.1–0.2 points). The rest-of-bloc line is derived: row = (ae − w·us)/(1−w) with w = 0.37 (since 2026-10-03, owner's decision), the US share of the group on purchasing-power weights, IMF series PPPSH, 2025: 37.1%. Method from the BIS, number from the IMF: the IMF averages the group's growth on purchasing-power weights, and so does the BIS wherever it averages countries (Graph 1.A and 1.C notes, "GDP-PPP weighted average"; also Graphs 8.A and 15.C), so undoing the IMF's average needs the IMF's weight; the BIS itself never states a US share of the advanced economies, and Box C has no countries. The share moved only between 36.1% and 37.3% over 2016–2026, so one weight serves the decade (sensitivity ≈±0.05pp on the derived line). The checker holds aeUsGrowthWeight to economies.html's usPppSharePct. Until 2026-10-03 w was 0.45, described here as the PPP weight; no measure supports it (market rates give 0.44). Results at 0.37 on the October 2026 round: rest of the group 2023–2026 between 1.0% and 2.0% (July: 1.0–1.8%), 2026 1.98%, "about 2.0%" (July: 1.5%); 2022 3.35%. Reference lines (since 2026-10-05): the dial's two rates at §4's starting values, 2.9% (total pay grows as it used to) and 3.4% (growth pays for the bet; July: 2.7% and 3.2%), typed in the page as REALITY_REFS and recomputed by the checker; before that they were the old dial's anchors (2 / 3.2 / 3.5). The group as a whole is above 2.91% only in 2021 (6.2%) and 2022 (3.0%), and above 3.44% only in 2021. The rest of the group is above the lower line in 2021 and 2022 only (6.14%, 3.35%; 2017: 2.66%). The US is above 2.91% also in 2018 and 2024 (3.0% each; 2023 is 2.9%, not above), and above 3.44% only in 2021. Growth paying for the bet with the rest at its 2026 rate: 1.98 + (3.44 − 1.98) ÷ 0.37 = 5.9% US growth, "about 6%". A page construction. Refresh with new WEO vintages, not per print. |
| The IMF's October 2026 round (folded in 2026-10-22) | IMF-sourced; page construction on it. The round was published on 13 Oct 2026 and caught nine days late, because the detector carried it as an undated reminder; it is now a dated entry in scripts/check-prints.py. What moved: the advanced economies' 2025 output, $69.4T → $69.6T (aeGdpT); their 2026 growth, today's pace, 1.8% → 2.1% (growthDial, BET.basePct); the rest of the group's 2026 growth, 1.5% → 2.0%. The US purchasing-power share stays 37.1%, and the wage pool's default stays $25T. Downstream, by the rows above and below: growth that pays for the bet 3.2% → 3.4%, its gap over today's pace 1.4 → 1.3 points, total pay keeps its old pace from 2.7% → 2.9%, and the level needed by 2030 $81.1T → $82.4T against $77.2T at today's pace. BIS check: nothing the BIS published moves; Box C works in rates and shares, not a dollar level, and "today's pace" was always the IMF's 2026 forecast. One thing to watch: the IMF's 2.1% for 2026 is a one-year forecast and its 2027 forecast is 1.8%, so the dial now starts from a year that may be unusually strong. The rule (today's pace = the IMF's forecast for the current year) is unchanged; changing it would be the owner's decision. The design not taken: holding today's pace at 1.8% until the 2027 figure becomes the current year, which keeps the dial's start steadier but would set the page against the IMF's own figure for this year. |
| 40-year chart: BIS line to ≈20% by 2060, then toward zero | Box C Graph C1.B reading (see §5 row). Fixed 2026-07-30 — endpoint moved from 2065 to 2060; slope label −11 pts/decade; the line continues declining past 2060 per Box C's "towards zero". Historical line: stylised advanced-economy average, ≈ −1 pt/decade post-1980 (labelled stylised). |
Moved 2026-09-11. The scoreboard — the three clocks, the gap chart, the strain chart and the watch list — lives on its own page, tracking.html ("Tracking the bet" in the site nav); the front-page essay ends at §6 with a pointer. The section number is retired on the front page rather than reused, as §8 was, and this ledger keeps the heading for its rows. One construction changed in the move: the gap chart's required ramp scaled the §2 floor as the sliders set it; on the tracking page it is drawn at the sliders' defaults ($3.5T over 2025–30, 5-year life, 10% return, $250B fixed opex → $1.3T/yr). check-derivations.py holds those four defaults equal to this registry and to index.html's slider attributes, and reads the GAP data from tracking.html.
| Page claim | Source & status |
|---|---|
| Three clocks | Market patience: BIS Ch I pp. 23–25 (valuations, risk premia, financial stability). Depreciation wall 2027–29: page inference — 3–5-year hardware life applied to the 2024–25 purchase wave. Debt rollover 2028–30: page inference — typical tenors on the 2025–26 issuance the BIS documents (Graphs 11.A, 13.A); the BIS documents the issuance, not the maturity wall. Constructions labelled as inference. Pointer corrected 2026-10-05. microsoft.html said "the front page's model treats 2027–29 as the years when a second round of spending has to be paid for". That is the second clock, which has been on the tracking page since 11 Sept 2026. The sentence now reads "the second clock on the tracking page marks 2027–29" and links to the clocks. "What the front page calls the depreciation wall" became "what this site calls", because the front page's text no longer uses the phrase. No figure is involved. The design not taken: naming the clock by its label on the chart, "Chips need replacing", and dropping "depreciation wall". That is closer to what the reader finds, but it is a larger rewrite of a page written as of its own results. Not a systemic change: two phrases and one link. |
| The fourth clock: the spend counter (added 2026-09-19) | Page construction, and arithmetic on one registered constant rather than a new one. The counter on tracking.html runs at the five's announced 2026 capital spending, capexShareTotalB = $842.5B since 22 Oct 2026 ($840B before Meta narrowed its plan), spread evenly over 365 days (2026 is not a leap year). Derived figures, all printed beside it: $842.5B × 10⁹ ÷ (365 × 24 × 60 × 60) = $26,715 a second; $842.5B ÷ 365 = $2.31B a day (from $26,636 and $2.30B). The counter itself shows elapsed time since the page was opened × that per-second rate, computed from the clock rather than accumulated per tick, so a throttled background tab cannot make it drift. Status of the shape, not just the figure: this is a rate, not a measurement — capex is lumpy, landing on completions and deliveries, and the page says so in the panel. It states no new quantity; it restates the §7 split-share total, which is why it takes no registry entry of its own. check-derivations.py holds the JS constant equal to capexShareTotalB and the three printed figures equal to the arithmetic above — they are static text (the panel must read correctly with scripting off), and static text is what Amazon's $200B → $220B raise left stale in two places in July. Scope: the BIS five only. Nvidia is the supplier and SpaceX the sixth builder, both outside the model and outside this figure, per the closed-roster rule. |
| Required slope ≈ +40%/yr in the watch-list | Page derivation: identifiable AI-revenue run-rate entering 2026 ≈ $250B/yr (order of magnitude across disclosed run-rates and lab revenues); floor $1.3T by end-2030 ⇒ (1300/250)^(1/5) − 1 ≈ 39%/yr. Order-of-magnitude, hence "~40%". Where the print pages say it is stated (2026-10-05). The 40% is stated in the watch list, which moved to the tracking page on 11 Sept 2026. The front page's text no longer states it. Five print pages still credited it to the front page. alphabet.html and microsoft.html now read "The tracking page puts the growth the sector needs at roughly 40% a year". amazon.html reads "By the tracking page's arithmetic". spacex.html reads "The growth rate the tracking page requires of the whole sector, and treats as hard to keep up"; it used to say "the one it calls demanding", a word no page uses now. oracle.html keeps its words, because the revenue-floor argument it names is on the front page, and its "about 40% a year" now links to the watch list. No figure changed. The design not taken: "this site works out", with no link. It is true wherever the figure is shown, but a reader who wants to see the 40% in use gets no pointer to it. Not a systemic change: four phrases and four links. |
| Required slope is a band, not a line (audit note, 2026-07-31) | The 40% inherits the base estimate's uncertainty: a $200B base ⇒ 45%/yr, $300B ⇒ 34%/yr. Sparkline verdicts within a few points of the line (Azure +43 "clears", AWS +37 "below") are judgment calls inside a ≈35–45% band; the colors mark the point estimate, this row records the band. |
| Gap chart: required ramp | §2 floor × the share of the BIS 2025–30 committed-capex window in place at each date. Anchors (share of the $3.5T default): end-2025 ≈ 13% (2025 spend ≈ $0.45T — the BIS's >$1T for 2025–26 minus ≈$0.84T of 2026 guides); end-2026 ≈ 36%; linear build to 100% at end-2030. Fixed 2026-07-30: the ramp previously accumulated from 2024 inside the window — the BIS window opens in 2025; prior-year spend is excluded, matching Graph 11.B's definition. |
| Gap chart: five-way split | Announced 2026 capex: Microsoft $190B (Apr 2026 call), Amazon $220B (raised from $200B on the 30 Jul Q2 call), Alphabet $195–205B (mid $200B), Meta $135–145B (mid $140B; narrowed from $130–145B, mid $137.5B, in its Q3 release of 21 Oct 2026), Oracle $90–95B FY27 gross guide (mid $92.5B — Oracle gives no CY2026 figure; the FY27 guide is the nearest announced annual number). Total $842.5B ($840B until 22 Oct 2026). Meta's change moves the ramp's end-2026 anchor by under 0.001 of the window, so the 0.36 anchor is kept, as it was for Amazon's raise. Amazon's raise also moves the ramp's end-2026 anchor from 0.363 to 0.369 of the $3.5T window — immaterial at the ledger's stated precision, so the 0.36 anchor is kept. Fixed 2026-08-27: the raise moved the data, this row and the registry on 30 July, but not the two places that restate the total in prose — the page's split-shares comment ("AMZN 200 ... = $820B") and the ramp-anchor row above ("≈$0.82T of 2026 guides"). Both now read $840B / $0.84T. check-derivations.py gained a check on the split-shares comment, mirroring the CY2023-baseline one: the share: fields were always guarded, the sentence documenting them was not, and on a page whose ethos is inspectability a stale comment is a wrong number. |
| Gap chart: actuals | Trailing-twelve-month reported total revenue minus each company's calendar-2023 total (Oracle on fiscal quarters ending Feb/May/Aug/Nov). 2023 baseline = the BIS's AI-boom trough (Graph 11.C notes). Quarterly revenue via stockanalysis.com, cross-checked against captured prints (Microsoft Q2'26 $90.0B ✓, Meta Q2'26 $60.8B ✓, Amazon Q2'26 $200.6B ✓ — SEC 8-K, 30 Jul, captured by the geoveu pipeline; Amazon TTM 775.7 − baseline 574.8 = +200.9). CY2023 baselines ($B): Microsoft 227.6, Alphabet 307.4, Amazon 574.8, Meta 134.9, Oracle 51.6. Deliberately over-credits AI (all growth counts); stated on the page. Oracle point appended 2026-09-10 for its FY27 Q1 (quarter ended 31 Aug, mapped to calendar Q3 2026): TTM revenue = FY26 total 67,357 - Q1 FY26 14,926 + Q1 FY27 19,345 = $71,776M, minus the 51.6 baseline = +$20.2B/yr. Both comparatives from the 10 June 2026 release, the current quarter from the 10 Sept 2026 release. This makes the series ragged by construction, and that is now a documented rule rather than an accident: Oracle's offset fiscal quarters land its August reading a month before the other four report calendar Q3, so GAP.x carries the longest series while drawGap() holds the aggregate at aggN = the shortest v array. Appending a company past that point therefore moves its own line only. The aggregate actual line and the 'reading as of' note did not move on this print, and the page says the reading is as of the last quarter all five have reported. |
| Gap chart: trend-adjusted toggle (added 2026-07-31) | Second crediting mode: expected TTM(t) = CY2023 base × (1+g)^(t−2024), residual = actual − expected, with g = each company's CY2022→CY2023 revenue growth. CY2022 totals ($B): Microsoft 204.0 (Mar/Jun 2022 quarters via stockanalysis.com; Sep 2022 $50.1B and Dec 2022 $52.7B from Microsoft's own FY23 Q1/Q2 releases), Alphabet 282.8, Amazon 514.0, Meta 116.6, Oracle 46.1 (offset fiscal quarters). Growth rates: 11.5% / 8.7% / 11.8% / 15.7% / 12.1%. Conservative for the AI residual by construction: Meta's 15.7% is a rebound from its 2022 revenue decline, and Oracle's 12.1% includes the Cerner acquisition's first full year — inflated trends shrink the residual. At Q2 2026 the aggregate residual is ≈ +$149B/yr versus ≈ +$553B/yr on ceiling crediting; the two readings bracket the unobservable true AI figure, and the page says so. |
| Oracle row in the watch-list | Oracle Q4 FY2026 press release, 10 Jun 2026 (investor.oracle.com): total revenue $19.2B +21% yoy; OCI (IaaS) +93% to $5.8B; RPO $638B, +$85B in the quarter; FY26 revenue $67.4B +17%; FY26 capex $55.7B (exceeding free cash flow); FY27 gross capex guide $90–95B; shares −10% on the print. Added 2026-07-30 — the BIS's fifth hyperscaler had been missing from the watch-list. Sparkline slope uses OCI +93% vs the +40% required slope. Companion page added 2026-08-27 (oracle.html), the last BIS-five member to get one, written from tools/companion-template.html. Figures re-confirmed against the release itself, Oracle being outside geoveu's capture: FY26 operating cash flow $31,977M, capital expenditures $55,663M, free cash flow −$23,686M, all from the release's trailing-four-quarter table. That table also independently confirms this page's Oracle strain series — TTM capex ÷ operating cash flow runs 41%, 53%, 72%, 102%, 127%, 159%, 205%, 174% across the eight quarters to May 2026, matching the chart's peak-above-200-then-ease-to-174 shape; and the implied Q4 step (TTM Q4 55,663 − TTM Q3 48,250 = 7,413) matches the chart's own 16.5 − 9.1 = 7.4. Corrected 2026-08-27: the row previously said "shares −10% on the spend". Sources disagree on magnitude and sampling — TradingKey records more than 7% down after hours recovering to −4.85% at $191.49, Sherwood 5%, others "as much as 10%", and the following morning's session −12.6% — so the row and the page now state the range (−7% after hours, −12.6% next morning) rather than a single unsupported figure. FY27 gross capex guide $90–95B and ≈$70B net are call figures, not in the release, labelled as such on the page; note the gross guide sits at or above the $90B FY27 revenue guide the release confirms. Refreshed 2026-09-10 for Oracle FY2027 Q1 (quarter ended 31 Aug), the row rewritten and moved to the top of the table; the 10 June figures above are retained as the prior reading. All figures from Exhibit 99.1 to the 8-K filed 10 Sept 2026 (SEC EDGAR accession 0001193125-26-387905), Oracle being outside geoveu's capture: total revenue $19,345M +30% (prior-year quarter $14,926M); total cloud $11.6B +62%, inside Oracle's own +58-64% guide; OCI (IaaS) $7.4B +121%, against +93% the prior quarter — the sparkline slope figure; cloud applications $4.2B +10%; software $5,550M -3%; RPO $664B, +$209B year on year and +$26B on the $638B of the prior quarter, on more than $30B of new AI cloud contracts; GAAP EPS $1.56 +55%, non-GAAP EPS $1.92 +30% against a ~$1.74 street; operating cash flow $23,103M +184%; capital expenditures $28,499M; free cash flow -$5,396M (stated on the page as -$5.4B); net cash outlay for capital expenditures $17,966M after subtracting $11,363M of customer prepayments and adding back $830M of short-term financing; a completed $20B at-the-market equity sale; 850MW of data-centre capacity and over 300,000 GPUs delivered since the May quarter; FY27 revenue guide raised to at least $90B with non-GAAP EPS $8.10, Q2 revenue growth guided 30-34%. The FY27 gross capex guide of $90-95B is unchanged, so the gap chart's $92.5B Oracle share and the $840B split total do not move. Page construction in the row: capex-to-revenue for the quarter, 28,499 / 19,345 = $1.47 per dollar of revenue, up from $0.86 in the May quarter. Press-reported, not filed: the share move. Sources disagree — StockTitan records +4.13% after hours at $159.26 after a session close of $152.94 that was itself down 5.38%, other reports run to +7% — so the row states the session fall and an approximate after-hours recovery rather than a single figure, the same treatment the June print received. Link added 2026-10-05. oracle.html's "about 40% a year" now links to the watch list on the tracking page, where the figure is stated. The words are unchanged. The note and the design not taken are in the row "Required slope ≈ +40%/yr in the watch-list". |
| Non-BIS rows carry a not-in-model note | Policy (2026-07-30): watch-list rows outside the BIS's five hyperscalers say so inline — Nvidia (the supplier the five's capex flows to), Apple (the control group, added 31 Jul at the owner's request), Anthropic and OpenAI (labs, tracked as pure-play revenue evidence, not capex), SpaceX (added 7 Aug — a sixth builder, tracked as the outer bound of the capex-to-revenue gauge). The US-jobs row is macro, not a company. Non-BIS rows get no slope sparkline and feed no gap-chart slice. SpaceX's AI revenue grew +247% yoy, which would clear the required slope several times over; it is stated in the row's caption text and deliberately not drawn, because a sparkline against the +40% line would imply membership of a model it is not in. |
| Apple row in the watch-list | Apple Q3 FY2026 press release, 30 Jul 2026: revenue $109.4B +16% (record June quarter); quarterly capex $3.4B / ≈$11B annualized (PrimeXBT preview comparison, cited on apple.html); shares −8% after hours on soft September-quarter guidance (CNBC live coverage). Control-group framing per the apple.html analysis; no sparkline — Apple makes no AI-capex bet, so there is no required slope for it to clear. Pointers corrected 2026-10-05. apple.html called the patience clock and the jobs test the front page's. Both moved to the tracking page on 11 Sept 2026 with the rest of the scoreboard. The page now says "the patience clock on the tracking page" and "the tracking page's test", and links to the clocks and to the jobs chart. "Lands on 7 August" stays, because the page is written as of the results of 30 July. No figure is involved. The design not taken: saying "this site" with no link. It stays true wherever the scoreboard lives, but it sends the reader nowhere. Not a systemic change: two phrases and two links. |
| Amazon row in the watch-list | geoveu capture of the 30 Jul 2026 SEC 8-K: revenue $200.6B +20% yoy (guide $194–199B); operating income $27.5B (guide $20–24B); AWS +37% yoy at a $169B annualized run rate, fastest in 18 quarters; AWS backlog $496B and the 2026 capex raise to $220B from the call (CNBC earnings coverage, 30 Jul); net income $62.6B including a $53.4B non-operating gain, primarily the Anthropic stake. Sparkline uses AWS +37% vs the +40% required slope. Full-page analyses (amazon.html, apple.html) carry their own source lists in their footers. Pointers corrected 2026-10-05. amazon.html called three things the front page's: the test set for Amazon, the gap chart and the jobs test. All three moved to the tracking page on 11 Sept 2026. The page now says "the tracking page" in each place and links to the watch list, the gap chart and the jobs chart. Its sentence on the required 40% is covered in the row "Required slope ≈ +40%/yr in the watch-list". "Lands on 7 August" stays, because the page is written as of the results of 30 July. No figure is involved. The design not taken: "The test this site set for Amazon", the wording alphabet.html opens with. It needs no link and cannot go stale, but it gives the reader no way to the row. Not a systemic change: three phrases and three links. |
| Nvidia row in the watch-list (reported 2026-08-26) | Not one of the BIS's five (it is the supplier); kept as the capex-side thermometer, labelled as such on the row. Q2 FY2027 figures from the 8-K filed 26 Aug 2026 (SEC EDGAR, EX-99.1 q2fy27pr.htm, accession 0001045810-26-000073): revenue $96,221M, +18% q/q and +106% y/y; Data Center revenue $89.0B, +18% q/q and +117% y/y; GAAP gross margin 75.0%; GAAP diluted EPS $2.46, non-GAAP $2.22; ~$26.0B returned to shareholders in the quarter, ~$99.0B remaining under the repurchase authorisation; Q3 FY2027 outlook $108.0B ±2%, explicitly assuming no Data Center compute revenue from China; announced financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500B of third-party capital, stated in the release as subject to definitive agreements. Beat comparators: its own Q1 guide of $91.0B ±2% (Q1 FY2027 release, 20 May 2026) and a street consensus of ≈$91.9B revenue / ≈$2.08 diluted EPS (the consensus figure is the geoveu capture's, not a filing figure, and is load-bearing for nothing). Sequential base stated in the row from 2026-08-27: Q1 FY2027 revenue $81.6B (+20% q/q, +85% y/y) and Data Center revenue $75.2B (+21% q/q, +92% y/y), from the Q1 FY2027 release of 20 May 2026 — the prior quarter's actuals, shown so the +18% q/q step has a visible base rather than only a percentage. The Q2 release rounds Data Center growth to +18% q/q, which against $75.2B implies $88.7B; the row uses the release's own $89.0B, not the back-computed figure. Correction folded in with this print: the pre-print row described the $91B figure as a data-centre guide — it was the guide for total revenue. Non-BIS row, so no slope sparkline, no gap-chart slice and no contribution to the capex total or required ramp, per the standing policy row above; the pre-print placeholder's promise to draw a slope after the print predated that policy (commit 229711d, before 2026-07-30) and is retired rather than honoured. Nvidia's revenue is the five's capital spending seen from the supplier's side of the invoice, so it enters neither the strain chart nor the gap chart, which would double-count the same dollars. |
| nvidia.html analysis page (added 2026-08-27) | Page construction on the companion page, every input from the Q2 FY2027 earnings release (SEC 8-K exhibit 99.1, filed 26 Aug 2026 — income statement, balance sheet and cash-flow statement). Cash conversion: operating cash flow ÷ net income = 24,077 ÷ 59,688 = 40.3%, against 15,365 ÷ 26,422 = 58.2% a year earlier. Receivables build ÷ revenue = 22,346 ÷ 96,221 = 23.2%. Days sales outstanding = AR ÷ quarterly revenue × 91 days (a 13-week fiscal quarter): 63,059 ÷ 96,221 × 91 = ≈59.6 days, against 38,466 ÷ 68,100 × 91 = ≈51.4 days at FY2026 year-end (Q4 FY2026 revenue $68.1B per the Q1 FY2027 release). Equity gains ÷ net income = 7,771 ÷ 59,688 = 13.0% for the quarter and 23,707 ÷ 118,010 = 20.1% for the half-year. Equity stakes held = marketable equity securities + non-marketable securities = 42,783 + 51,157 = $93,940M, against 12,886 + 22,251 = $35,137M at FY2026 year-end, an increase of $58,803M in six months. H1 equity purchases ÷ H1 own capex = 42,404 ÷ 4,434 = ≈9.6×. H1 equity purchases ÷ H1 revenue = 42,404 ÷ (96,221 + 81,615 = 177,836) = ≈23.8%, the page’s "roughly 24 cents per dollar of revenue". Capex ÷ revenue on BIS Graph 11.A's gauge = 2,677 ÷ 96,221 = ≈0.028×, the lowest on the watch list by an order of magnitude and the reason the supplier's own spending is immaterial to the model. Long-term debt 7,469 → 32,366 over the half-year, with $24,896M raised in Q2; buybacks $19,732M and dividends $6,047M in Q2 (summing to the release's stated ≈$26.0B returned). Q3 guide step-up = 108,000 ÷ 96,221 − 1 = +12.2% q/q; Data Center share of revenue = 89,000 ÷ 96,221 = 92.5%. Non-filing items, attributed on the page and load-bearing for nothing: the after-hours share move (down ≈0.5% on the release, then up ≈5% to record highs during the call), the "extreme pricing conditions in memory" remark, the fiscal-2028 comment and the cloud-industry backlog figure of >$2 trillion — all from Yahoo Finance's coverage of the call, not from the release. The backlog is the cloud industry's as reported, not Nvidia's own; an earlier secondary summary described it as Nvidia's backlog and was not used. A hyperscaler-capex forecast (≈$800B for 2026, ≈$1.3T for 2027) was also attributed to the call by secondary sources but could not be verified against a readable transcript, so it is deliberately absent from the page. Per the roster rule, none of this enters the constants registry, the gap chart, the strain chart or the capex total: Nvidia's revenue is the five's capital spending seen from the supplier's side of the invoice. Two strain readings dated 2026-10-05. "The cash behind the profit" quotes two readings from the strain chart (row "Strain chart"): Amazon about 107% and Oracle about 174%. Both were the chart's readings when Nvidia reported on 26 Aug 2026: Amazon 173.0 ÷ 161.4 = 107%, Oracle 55.6 ÷ 32.0 = 174%. The sentence gave them in the present tense and called the chart the front page's. Oracle's results of 10 Sept moved its reading to 161%, and the chart moved to the tracking page on 11 Sept, so by 5 Oct the sentence was wrong on both counts. It now opens "When Nvidia reported" and links to the chart on the tracking page. No figure changed. The two readings are not updated at later prints, and the checker does not read this page (row "Strain chart: the sentences that quote a reading"). The design not taken: changing Oracle's figure to 161% and adding this page to the checker. The page would then need an edit at every Oracle print and every Amazon print, the checker would need a new rule for Amazon's reading, and Oracle's 161% would need the caveat it carries everywhere else (about 244% without customers' prepayments). That is a lot of upkeep for a page that reports one set of results. Not a systemic change: one sentence reworded. Jobs link moved 2026-10-05. "Next test" linked the August US jobs report to the front page and called the jobs test the front page's. The jobs row and the jobs chart moved to the tracking page on 11 Sept 2026 with the rest of the scoreboard, and this sentence was missed. The link now goes to the jobs chart on the tracking page, and the sentence says "the tracking page's test". "Arrives on 4 September" stays, because the section is written as of the results of 26 Aug. No figure is involved. The design not taken: linking to the jobs row in the watch list instead. The row moves down the table as newer results land, and the chart stays where it is. Not a systemic change: one link and one phrase. |
| SpaceX row in the watch-list (added 2026-08-07) | SpaceX Q2 2026 earnings release, 4 Aug 2026 (SEC EX-99.1): revenue $7,814M, +92% yoy; segments Space $962M / Connectivity $4,291M / AI $2,561M; AI capex $15,828M against total capex $18,369M; H1 capex $28,476M and H1 net cash from operating activities $3,466M; net loss $541M; Adjusted EBITDA $3,538M; nameplate compute 1.4 GW; $14.1B of contracted cloud sales; IPO closed 15 Jun 2026 with ≈$85.7B net proceeds (Nasdaq: SPCX, trading from 12 Jun), plus a $25B bond issue on 26 Jun in five tranches maturing 2031–2056 at a weighted-average 5.855%; cash and marketable securities $100B. Share reaction −8% after hours to ≈$125 against the $135 offer price: TechCrunch, 4 Aug 2026 (not in the release). Non-BIS row, so no slope sparkline and no gap-chart slice, per the standing policy row below. Page arithmetic in the row: capex ÷ revenue 18,369 ÷ 7,814 = 2.35×; AI capex ÷ AI revenue 15,828 ÷ 2,561 = 6.18×; AI revenue annualised 2,561 × 4 = ≈$10.2B/yr. |
| §2 floor arithmetic applied to SpaceX's quarter (added 2026-08-07) | Page construction — the §2 floor run on one company's single quarter at the page's own defaults (5-year hardware life, 10% required return), with the $250B fixed-opex term omitted because it is a sector-level figure, so the result is a floor on the floor. Q2 AI capex $15.828B ⇒ 15.828/5 + 0.10 × 15.828 = 3.166 + 1.583 = ≈$4.75B/yr of permanent bill from one quarter's spending; four such quarters = ≈$19.0B/yr against an AI revenue run-rate of ≈$10.2B/yr. (H1's $23.551B on the same formula ⇒ ≈$7.07B/yr.) The same known limitation applies as in §2: the bill is static and linear, while the system that must pay it is financed. |
| spacex.html capex-to-revenue bar chart (added 2026-08-07) | Page construction on the companion page, same gauge as BIS Graph 11.A applied per company to the June 2026 quarter, each from its own release: Amazon 54.2/200.6 = 0.27×, Alphabet 44.9/119.8 = 0.37×, Microsoft 35.8/90.0 = 0.40×, Meta 30.1/60.8 = 0.50×, Oracle 16.5/19.2 = 0.86× (offset fiscal quarter ending May; Oracle's next quarter, reported 10 Sept, reached 28,499/19,345 = 1.47×, and the §1 note and the SpaceX watch-list row were dated on 2026-09-10 so the 0.86 is no longer read as current — the chart itself stays pinned to the June quarter, which is what makes it like-for-like with SpaceX's), SpaceX 18,369/7,814 = 2.35×, SpaceX AI segment 15,828/2,561 = 6.18× (drawn truncated — the bar would need 2.6× the chart's width; the label carries the figure). Numerators are the quarterly capex figures already anchored in the strain-chart row above; cash capex, finance leases excluded. Companion-page target-slope arithmetic, same status: $1T by end-2030 from ≈$31.3B/yr (Q2 × 4) over 4.5 years ⇒ (1000/31.3)^(1/4.5) − 1 ≈ 116%/yr; the page states ≈138%/yr over four years (end-2026 → end-2030), the stricter and more natural reading of "reaching $1T in 2030". Both readings dwarf the sector's ~40% required slope, which is the only claim the page hangs on them. Pointers corrected 2026-10-05. spacex.html said in two places that nothing on it feeds the front page's required pace of revenue growth, its split of the spending or its gap chart, and it called the jobs test the front page's. The required pace, the five-way split, the gap chart and the jobs chart moved to the tracking page on 11 Sept 2026. The capital spending total is still the front page's §2, and the page still says so. The other phrases now name the tracking page and link to the gap chart and the jobs chart. Its sentence on the required 40% is covered in the row "Required slope ≈ +40%/yr in the watch-list". "Landed on 7 August" stays. No figure is involved. The design not taken: one sentence at the top of the page saying the scoreboard has moved, with the old wording left below it. That is one edit instead of four, but every later sentence would still point the wrong way. Not a systemic change: three phrases and three links. |
| spacex.html "Should the model's bill go up?" (added 2026-08-07) | Page construction answering the double-counting question explicitly. Capex side: not double-counted — compute rented from SpaceX sits in the renter's operating costs, not its capex line, so SpaceX's machines are in none of the five's capex totals. Model side: the §2 slider is anchored to BIS Graph 11.B's sector-wide cumulative axis ($2–4.5T committed 2025–30), and the five's $840B splits the bill (gap-chart band) rather than setting it — so a sixth builder's spend argues the slider toward the top of its range instead of adding a slice on top, which would double-count against the sector-wide anchor. Arithmetic: AI-segment capex annualised 15.828 × 4 ≈ $63B/yr; sustained mid-2026 → end-2030 (≈4.5 yrs) ≈ $0.3T; slider $3.5T → $3.8T moves the floor by 0.3/5 + 0.10 × 0.3 = ≈$90B/yr at the §2 defaults. Revenue side: would double-count — intra-boom payments net to zero on consolidation; the floor must be paid from outside the build-out. This is the ledger's §1 scope-note and §3 third-strip reasoning, stated on the page where the reader will actually ask it. |
| Strain-chart SpaceX marker (added 2026-08-07) | Named deviation from the chart's own construction, stated in the label and the panel note. Every plotted line is trailing-twelve-month; SpaceX has only two quarters of public accounts, so its ratio is half-year: 28,476 ÷ 3,466 = ≈820%, H1 2026 capex ÷ H1 2026 operating cash flow. Drawn as a top-edge flag rather than a sixth line for two reasons — it is not one of the BIS's five, and plotting 820% on a 220% axis would compress the five into the baseline. Comparator on the chart: Oracle's TTM ratio, ≈174% when the marker was added and ≈161% since Oracle's 10 Sept 2026 print (≈244% net of customer prepayments; both derived in the next row). Corrected 2026-10-05: the panel note's closing sentence still said "Oracle, at 174%" after that print. It now gives 161% and the 244%, and adds the $20B share sale the Oracle row already reports, since "spending its reserves and borrowing" no longer covered how Oracle pays for the difference. No new figure. |
| The five builders' pages: revenue growth against each one's share of the floor (added 2026-10-23, owner's request) | Page construction, restating the gap chart one company at a time; no new constant. Each of alphabet.html, amazon.html, meta.html, microsoft.html and oracle.html draws, from gap.js: (1) the company's share of the floor coming due, share × rampFrac(t) × floorB, where share is its 2026 plan ÷ the five's (capexShareB ÷ capexShareTotalB) and floorB is measuresT.floor × 1000 = 1,300; (2) its revenue growth on the upper-limit reading, the v series (TTM revenue minus CY2023 revenue); (3) the same on the trend-adjusted reading, v − base × ((1 + trend)^(t − 2024) − 1). These are the gap chart's own definitions (rows above), drawn for one company instead of stacked. Worked example, Meta at its latest quarter (x = 2026.75): ramp fraction 0.13 + 0.75 × (0.36 − 0.13) = 0.3025, × 1,300 = $393.3B due across the five; Meta's share 140 ÷ 842.5 = 16.6%, so $65.3B due; upper reading $106.7B; trend reading 106.7 − 134.9 × (1.157^2.75 − 1) = 106.7 − 66.6 = $40.1B. Ahead on the first, behind on the second. Its share by the end of 2030: 0.166 × 1,300 = $216B. The chart's note is computed from the data and names the quarter it reads, because the chart is redrawn with each print while the page's text is written as of the company's own print. The data moved from tracking.html to gap.js in the same change, and check-derivations.py now reads it there (ramp, split shares, baselines, trends, and the new floorB). The design not taken: a script that writes each company's chart into its page as fixed SVG, like scripts/menu.py writes the masthead. It needs no new shared file and draws without scripting, but every print would need a second write step and a check that the five SVGs match the data, and a forgotten step leaves a chart a quarter behind with nothing to show it. With one data file the six charts cannot disagree. The cost of the design taken: the company charts need scripting, and gap.js is cached for an hour apart from the pages, so a visitor can briefly see a chart one quarter behind. |
| Strain chart (added 2026-07-31) | TTM capital expenditure ÷ TTM operating cash flow per company, quarterly raw series Q1 2024–Q2 2026 from the quarterly cash-flow statements via stockanalysis.com (Oracle on offset fiscal quarters). Anchors cross-checked: Oracle's four FY26 quarters sum to its reported $55.7B capex; Microsoft's Sep/Dec 2024 quarters re-verified after a table misread (OCF $34.2B/$22.3B, capex $14.9B/$15.8B); Meta's Q2 2026 (capex $30.1B cash vs $31.1B incl. finance leases in its release) and Microsoft's Q2 2026 ($35.8B cash vs $41B incl. leases) consistent net of leases; Alphabet/Amazon 2024 quarters match reported figures. Cash capex only — finance leases excluded, understating strain (the conservative direction). Named deviation from BIS Graph 11.A: the BIS divides capex by revenue (intensity); this chart divides by operating cash flow (funding capacity) — stated on the page. TTM ratios computed in code from the raw quarters, so a print appends raw data and the chart follows. The sentences that quote a reading do not follow by themselves: see the row "Strain chart: the sentences that quote a reading". Oracle quarter appended 2026-09-10 (FY27 Q1, ended 31 Aug, mapped to calendar Q3 2026): capex $28,499M → 28.5, operating cash flow $23,103M → 23.1, both from the 10 Sept release's condensed cash-flow statement. New TTM: capex 12.0 + 18.6 + 16.5 + 28.5 = 75.6; operating cash flow 2.1 + 7.2 + 14.6 + 23.1 = 47.0; ratio 161%, down from 174%. Page construction, annotated on the page because the fall is an artefact: $11,363M of that operating cash flow is customer prepayments for capacity not yet built, and the prior quarter carried $4,592M of the same. Oracle subtracts both in its own 'net cash outlay for capital expenditures' table. Stripping them from the denominator gives 75,660 / (46,940 - 11,363 - 4,592) = 75,660 / 30,985 = 244%, the figure the chart's construction note and oracle.html both carry. The chart plots the unadjusted statement figures, so the note says to read the fall as funding arriving early rather than strain easing. Two geometry changes in the same commit, both forced by the ragged series: the x-domain was extended from 2026.5 to 2026.75 with a 'Q3 2026' tick, and the per-company end labels now anchor to each company's own last point instead of a shared right edge — which was only ever correct while every series ended together. |
| Jobs row: the divergence test | Page derivation (§4 arithmetic inverted, added 2026-07-30; payroll figures refreshed 2026-10-02 with the September print). If the gap chart's aggregate actual (+$553B/yr at Q2 2026, all five reported) were wage-pool revenue: 553 ÷ 25% capture = ≈$2.21T displaced wages/yr; ÷ $80k = ≈28M jobs-equivalent — visibly absent from payrolls (September: +29k, u 4.2%, average hourly earnings +3.0% yoy, prior-12-month average gain +45k; per the BLS Employment Situation release of 2 Oct 2026 and the geoveu pipeline's capture of the same morning, which agree), so today's revenue is tools + circular financing, not substitution. The same release revised July from +21k to −10k and August from +162k to +133k (−60k combined); July had been −23k as first reported on 7 Aug and +21k after the 4 Sept revision, and the row states each revision rather than overwriting it. Industry detail from the release's table B-1 (seasonally adjusted, September 2026 vs September 2025): information −10k in the month, 2,859k → 2,739k = −120k on the year; computing infrastructure, data processing and web hosting −1.6k in the month, 476.3k → 453.8k = −22.5k on the year; manufacturing +9k in the month and +72k since its December 2025 low (release text, matching geoveu's captured component). The page claims nothing about the white-collar split, which neither the release's captured components nor the row carries; the information-sector losses are reported as the BLS states them, unattributed. The gap chart's aggregate holds at quarters all five have reported (Q2 2026), so Oracle's 10 Sept point did not move it and the dollar figures stand. The 2027–28 boundary is where the required ramp exceeds plausible IT-budget supply (software pool ≈$1.2T/yr total, §3): past it, on-schedule revenue and flat payrolls are mutually exclusive. Figures update with the gap chart; the jobs row's dollar figures should be refreshed whenever the aggregate moves materially (a jobs print itself changes no GAP or STRAIN data). Scope caveat (added 2026-09-04): the row watches one country while the BIS's framing, and §6's $24T→$48T, is across the advanced economies; the page reads US payrolls as the leading indicator rather than the whole test because the capex, the AI-exposed sectors and the BIS's own early evidence of labour-market adjustment (AI-exposed US sectors showing higher productivity gains partly at the expense of employment growth, Graph 10.C, p. 21) are US. That ordering is the page's reasoning, not a BIS claim. |
| Jobs chart: US payrolls since January 2023 (added 2026-10-02) | Page construction, drawn on the tracking page beneath the strain chart. What it plots. Upper panel: the monthly change in US total nonfarm payrolls (BLS series CES0000000001, seasonally adjusted), January 2023 to the latest release, with a line for the average of the twelve months before each month, the BLS's own yardstick in the release text (45,000 a month for September 2026: the average of September 2025 to August 2026, rounded to the thousand). Lower panel: the change in jobs since December 2022 in information (CES5000000001) and manufacturing (CES3000000001). January 2023 is the start because it is where the gap chart measures revenue growth from, so both halves of the divergence test begin on the same date. Sources. The series are the BLS's current figures, fetched from its public data API. They are not a record of what was first published: the BLS revises the two most recent months with every release and the whole series each February, so earlier bars move and the chart is regenerated from the BLS on each print rather than appended to. The newest month is the second confirmation the figures rule asks for: it must equal geoveu's capture of the same release (headline, and manufacturing where geoveu holds it) or nothing is written. On 2 Oct 2026 every figure the release states in words matched the series: July −10k, August +133k, September +29k; the 45k average; manufacturing +9k in the month and +72k from its December 2025 low; information −10k in the month and 2,859k → 2,739k on the year. First estimates. Dashed outlines mark a month's first published figure where it differs from today's. The BLS does not keep these, so only those this site recorded on the day are shown, from July 2026: July −23k (geoveu's capture of 7 Aug 2026, kept in its figure history, and this site's revision note of the same day), August +162k (geoveu, 4 Sept 2026, and the revision note), September +29k (geoveu, 2 Oct 2026). The list is append-only. Scope. United States only, and the chart says so on its face: the BIS frames the bet across the advanced economies. The reasoning for reading US payrolls first is the jobs row's scope caveat above; it is the page's reasoning, not a BIS claim. What it does not show. The scale that would count, about 28 million jobs-equivalent on the jobs row's arithmetic, is not drawn: on the same axis it would flatten every bar. The panel points to the row instead. Nothing on the chart attributes any month's change to AI, and the lower panel's two industries are reported as the BLS states them. The chart feeds no model, constant or other chart. The design not taken. A hand-kept array appended each month, like the gap and strain charts' data. Rejected because payrolls, unlike company results, are revised after the fact: an appended array would keep superseded figures beside current ones with nothing to tell them apart, and the February benchmark would leave the whole history quietly wrong. A chart that fetched the BLS in the reader's browser was rejected too: it would make the page depend on a third party at read time and would skip the second-source check. |
| US labour-share row in the watch-list (added 2026-09-03) | Directly sourced, and evidence beside the model — no slope, no gap-chart slice, no contribution to any total. BLS Productivity and Costs, Second Quarter 2026, Revised, released 3 Sept 2026 (USDL 26-1434): labour share 52.8%, stated by BLS as "the lowest level in the series, which begins in the first quarter of 1947"; labour productivity +1.4%, output +1.7%, hours +0.3%, unit labour costs +1.2%, hourly compensation +2.6%, real hourly compensation −3.3%. The 6 Aug preliminary (USDL 26-1290) read labour share 52.9%, unit labour costs +1.3%, hourly compensation +2.7% — both prints are recorded because the revision moved every one of those figures. Confirmed twice per the figures rule: geoveu captures the series from 3 Sept 2026 (us-productivity-q2-2026-revised, parser guards quarter, vintage and release-dateline) and the BLS release page itself. Sector, which is load-bearing: nonfarm business — excludes general government, nonprofits and households — so this is NOT a whole-economy labour share and is not comparable like-for-like with the AMECO row below or with §4/§6's intercept. |
| US labour-productivity row in the watch-list (added 2026-09-05) | Directly sourced, and evidence beside the model — no slope, no gap-chart slice, no contribution to any total. The third row drawn from BLS Productivity and Costs, Second Quarter 2026, Revised (USDL 26-1434, released 3 Sept 2026), nonfarm business, quarterly changes at seasonally adjusted annual rates: labour productivity +1.4%, output +1.7%, hours worked +0.3%; from a year earlier productivity +2.2% (output +2.5%, hours +0.2%) — Table A1. Table B1 shows the 6 Aug preliminary (USDL 26-1290) productivity figure also at +1.4%, with output and hours unrevised; Table B2 puts Q1 2026 at +0.8% (output +1.5%, hours +0.7%), unrevised. Cycle rates from the release text: 2.1% annualised since Q4 2019 (output 2.5%, hours 0.4%), 1.5% in the previous cycle (Q4 2007 to Q4 2019), 2.1% long-run since Q1 1947. Unit profits of nonfinancial corporations +43.0% annualised is the release's nonfinancial-corporate figure already recorded in §8. Confirmed twice: geoveu's us-productivity-q2-2026-revised record carries productivity +1.4%, unit labour costs +1.2%, hourly compensation +2.6% and real hourly compensation −3.3% as components, agreeing with the release; and the BLS public API series PRS85006092 (nonfarm business labour productivity, % change from previous quarter at annual rate) reads 1.4 for Q2 2026 and 0.8 for Q1 2026, with PRS85006091 (% change from the same quarter a year ago) reading 2.2 for Q2 2026 — independent retrievals of the same revised data. Output, hours and the cycle rates appear in the release only (geoveu's record carries no component for them), so they stand on the release plus the API's agreeing productivity figures, which are the ratio those inputs produce. Scope, load-bearing: the series is output per hour worked, so it presupposes labour in the production; it rises both when workers produce more and when fewer hours produce the same output, and the row reads it only beside its hours line for that reason. It measures neither capture (§4) nor the bill (§2), and the row says so. Vintage handling as for the labour-share row: both prints of a quarter fold in separately and the row names its vintage. |
| Whole-economy labour-share row in the watch-list (added 2026-09-03) | Directly sourced, and evidence beside the model. European Commission AMECO, series ALCD0 "Adjusted wage share: total economy", percentage of GDP, spring 2026 vintage, read via DBnomics 2026-09-03. Adjusted = employee-equivalent compensation imputed to each self-employed person, which is what makes it whole-economy comparable. 2024 actuals: euro area (EA20) 56.34%, EU27 55.70%, US 54.49%, Germany 59.09%. Commission forecasts for 2026: 56.98 / 56.24 / 53.56 / 60.31 — labelled as forecasts on the page, because 2025–27 in this vintage are projections. The direction split the row rests on: EA20 55.39 (2022) → 56.98 (2026f) rising, US 54.89 (2022) → 53.56 (2026f) falling. Refresh with new AMECO vintages (spring and autumn forecast rounds), not per print — same policy as the §6 reality-check chart's IMF WEO series. |
| Strain chart: the sentences that quote a reading (check added 2026-10-05) | Page construction, held by machine. No new constant. The chart computes each company's reading from the raw quarters. The sentences that repeat a reading are typed by hand, so a print that appends a quarter moves the chart and leaves them behind. After Oracle's 10 Sept 2026 print the chart read 161% and one sentence said 174% until 5 Oct (row "Strain-chart SpaceX marker"). check-derivations.py now recomputes the readings from the STRAIN arrays as the chart does (last four quarters of capital spending ÷ last four of operating cash flow, to the whole percent) and fails the deploy on any of three things. (1) The chart's aria-label must give the latest reading of each of the five. (2) Two standing sentences must give Oracle's latest: the note on the SpaceX flag ("Oracle is at 161%") and debt.html ("came to 161% of the cash its business generated"). (3) On tracking.html, oracle.html and debt.html, Oracle's previous reading may appear only as the start of a move ("from 174% to …"), and that sentence must also give the latest. Anywhere else it is a copy the last print left behind, and the checker prints the sentence. Rule 3 lists no sentences, so it catches a copy nobody remembered, and it survives the rewrite the Oracle row and oracle.html get at every print. Readings today: Oracle 75.6 ÷ 47.0 = 161%, and before that 55.6 ÷ 32.0 = 174%; Microsoft 63%, Alphabet 71%, Meta 69%, Amazon 107%. Not read, on purpose: the front page's revision note and this ledger's dated rows, which say what was true on their day, and the other companies' print pages, which are written as of their own print. One of them quotes a reading: nvidia.html gives Amazon's 107% and Oracle's 174%. Since 2026-10-05 its sentence opens "When Nvidia reported", so it stays true when the chart moves (row "nvidia.html analysis page"). A reading quoted on a print page must be dated in the same way. The other four companies are held in the aria-label only. No other sentence on the checked pages quotes them, and rule 3 cannot work for a two-digit reading: a search for "63%" stops on unrelated figures. Not held: the "about 244%" net of prepayments. Its arithmetic (row "Strain chart") needs the prepayments inside the twelve months, $11,363M + $4,592M, and the second is on no page and in no registry. One registry entry for that sum would let the checker hold it: 75.6 ÷ (47.0 − 16.0) = 244%. Until then it is recomputed by hand at each Oracle print. It sits in or beside every sentence rule 3 prints, so the failure leads its author to it. Two limits. A wrong figure that is neither the latest nor the previous reading is caught in the aria-label, in the two standing sentences and in any "from 174% to" sentence, but not in a newly written sentence of another form. And when a new quarter leaves the whole-percent reading unchanged, rule 3 has nothing to look for and is skipped. The design not taken: a list of the six sentences, each matched by its wording. It would hold every listed copy exactly. But three of the six are rewritten at every Oracle print, so the list would break each quarter, and a list cannot catch a sentence nobody listed, which is what happened here. Not a systemic change: one more check in an existing deploy gate, with no data and no page figure altered. Verified by breaking each of the six copies in turn, and by running the check against the page as it stood before the 5 Oct correction: every case failed and named the sentence. |
| meta.html after-hours move — corrected 2026-08-27, and again on its share card 2026-09-11 | The companion page reported Meta's Q2 print as "shares −5% after hours" in three places and "$556 after hours", and carried alphabet.html's Sources block verbatim — citing Alphabet's 22 Jul release as the source for Meta's 29 Jul figures. Both were copy-paste from alphabet.html, which the page's own line "the Alphabet reaction, repeated" makes legible. The watch-list row on index.html was right throughout at −9.6%. Corrected against source: close $585.61, after-hours $529.15, −9.64%, and the Sources block replaced with Meta's own (results release, call transcript, Fortune, TradingKey). The geoveu pipeline no longer holds the July Meta record — its capture rotates — so this was verified from the public record instead. Standing rule from this print (2026-08-27): every figure now needs two independent confirmations before it ships, geoveu's record being only one of them, with the primary filing settling any disagreement; and a new companion page starts from tools/companion-template.html with a blank sources block rather than from a sibling page. scripts/check-pages.py enforces the second mechanically — it fails the deploy on an unfilled placeholder, on a sources block that never names its own company, or on two pages sharing one. Verified against this very bug: restoring alphabet.html's block onto meta.html trips both the naming and the duplicate check. The page's data table was correct throughout: revenue $60.8B, EPS $6.18, FCF $0.8B (reported $784M), capex $31.1B gross against the $30.1B cash figure the strain chart uses net of finance leases. A fourth instance, found 2026-09-11. The correction reached the page's three occurrences but not its share card — the picture a chat app shows when the link is posted — which went on reading "shares −5% on a beat" for two more weeks. Nothing could see it: the card's words existed only inside a PNG, and no check in this repo reads pixels. The card now says −9.6%, matching the page's own clock, and the words of every card live as text in the manifest in tools/make-og-card.py, which scripts/check-pages.py re-renders and compares byte for byte on every deploy. The generator was verified against the seven cards already shipped before it was trusted: each regenerated identically from its recorded text, so the manifest records what those cards actually said rather than a later paraphrase. |
|---|---|
| spacex.html note of 8 Oct 2026: the $40B borrowing plan and the bond market's reading of 7 Oct | Press-reported, not filed; evidence beside the model. The plan: FT, 6 Oct 2026 ("SpaceX looks to raise $40bn to buy Nvidia chips"): $40B, about $10B of bank loans and $30B of investment-grade bonds, Apollo expected to lead, Pimco among lenders in talks; Reuters, 6 Oct, confirms the total from two of its own sources and the split and Pimco from one. The market, 7 Oct: five-year CDS 194bp (FT, "$19,400 annually to protect $1mn"), began trading in June "at around 110bp"; Bloomberg (Caleb Mutua, 7 Oct), from ICE Data Services, up as much as 14.5bp to around 195.4bp, "the highest intraday level since the swaps … started actively trading in June". The page says "about 195" and "roughly $19,500", the two readings rounded together; this row keeps both. The 2056 bond: 2.36 points over Treasuries, up 0.09 on the day, MarketAxess via the FT; 238bp, up 12bp as of 9:40 New York, Trace via Bloomberg; sold at 175bp in June (both). The page says "about 2.4" and "0.6 of a point" (2.36 − 1.75 = 0.61; 2.38 − 1.75 = 0.63). Comparisons, FT only: Oracle's five-year CDS 244bp, up 40bp since the start of September; Nvidia's about 81bp from 45bp in early July; Alphabet, Meta and Microsoft "close to record highs". Adam Jonas's note (Morgan Stanley) as the FT quotes it, an attributed claim. Share close 7 Oct $167.60 from $171.92, −2.51% (Yahoo Finance price history); "two and a half per cent" on the FT News Briefing of 8 Oct. Page arithmetic: $39.4B of debt (the June balance sheet, row above) + $40B ≈ double. Owner's decision of 2026-10-08 on confirmation: a figure reported by the FT, Bloomberg or Reuters stands on that one report; the project's rules carry the exception under "Confirming a figure", and the FAQ states it. Before it, Apollo's role, the June CDS level and the Oracle and Nvidia readings were single-sourced and off the page. Feeds nothing: the roster stays closed, and no registry entry moves. Not a systemic change: one dated section on a print page, one sentence in §2 and one paragraph in §5 of debt.html. The design not taken: refreshing the SpaceX row on the tracking page. The row reports prints and the date of the next; a market move between prints belongs on the page the row links to, which is where a company's story is kept between its prints. |
| Anthropic Q3 2026 row in the watch-list (added 2026-10-16; the FT's figures added the same day) | Press-reported, not filed; evidence beside the model. What Anthropic told prospective investors at an investor day in San Francisco on 14 Oct 2026. Both reports stand on one source each under the owner's FT, Bloomberg and Reuters exception (2026-10-08), and the row is labelled press-reported. Financial Times, 15 Oct (Amelia Pollard and George Hammond, "Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO", from multiple people with knowledge of the matter), read in full: third-quarter revenue $16.8B, up from $11.5B in the second quarter and about $4.8B in the first; profitable on an adjusted basis for the second quarter in a row; cash-flow positive; $1.70 of revenue for every $1 spent on computing power; adjusted operating margin 11%, excluding stock-based pay; no update to annualised revenue, which was $65B at the end of July; SemiAnalysis projects about $74B annualised from the new figures, below the summer, when some investors told the FT it would far surpass $100B by year-end; distillation 5–7% of revenue, cut back "materially" in the third quarter (four people with direct knowledge); Meta and Microsoft, two major customers, increasingly using their own AI tools; a price war with OpenAI and cheaper Chinese rivals. Reuters, 14 Oct (Milana Vinn and Krystal Hu, two people who attended), read at source: $16.8B, $11.6B and $4.7B for the three quarters, up 45% on the quarter, profitable on an adjusted basis; listing likely after the November US midterm elections. Reuters, 15 Oct (Milana Vinn and Anirban Sen, four people familiar): some investors discussing a valuation of about $1.5 trillion, against roughly $2 trillion expected before. The two reports differ on the earlier quarters, $11.5B against $11.6B and $4.8B against $4.7B. Neither is a filing, so neither settles it, and the row gives both. The FT's $11.5B agrees with Reuters' own report of 28 Sept (the row "Anthropic's draft prospectus" in §2, which carries the label "Not an official filing."). Page arithmetic: 16.8 ÷ 11.5 − 1 = 46% and 16.8 ÷ 11.6 − 1 = 45%, so the row says about 45% (the FT's headline "nearly 50%" rounds the same figures); 1 ÷ 1.70 = ≈59 cents of computing per dollar of revenue. The $65B and $74B are annualised rates, the latest month or weeks multiplied up, not revenue earned; the $74B is SemiAnalysis's estimate, not the company's. Outside the BIS's five: no slope sparkline, no gap-chart slice, nothing in the registry. The design not taken: not systemic, one watch-list row. |
| anthropic.html analysis page (added 2026-10-16, at the owner's request) | Press-reported, not filed; evidence beside the model. Every third-quarter figure is from the FT and Reuters reports in the row above, with the same two-outlet split on the earlier quarters; nothing on the page is filed, and the table note, the clocks and the share card say so. The page's own arithmetic: computing in the quarter = 16.8 ÷ 1.70 = ≈$9.9B (9.88); at that pace × 4 = ≈$40B a year (39.5); against the front page's floor at its defaults, 40 ÷ 1,300 = ≈3%. Computing per dollar of revenue = 1 ÷ 1.70 = ≈59 cents, against the bill's default s × (1 − m) = 0.75 × 0.90 = 67.5 cents. In the bill's formula floor × [d + (1 − d) ÷ (s × (1 − m))], the product s × (1 − m) is exactly computing ÷ revenue for a firm in the middle, so Anthropic's ratio replaces it without splitting it into a cost share and a margin: 1,300 × [0.17 + 0.83 × 1.70] = 1,300 × 1.581 = ≈$2,055B ≈ $2.06T, against 1,300 × [0.17 + 0.83 ÷ 0.675] = ≈$1,820B ≈ $1.82T at the defaults (written on 16 Oct at d = 16%: $2,064B and $1,826B; the rounded $2.06T did not move) (§2 rows on the bill). This is one firm in one quarter, on a ratio the company chose to disclose, of unknown scope (training may or may not be included); the page says so and changes no slider. Draft-prospectus figures quoted on the page, each in a block carrying the label "Not an official filing.": 2025 revenue $4.59B and operating loss $8.06B (8.06 ÷ 4.59 = ≈$1.76 lost per dollar of revenue); computing commitments $518B, of which $252.5B with Google, Amazon and Microsoft (§2 row "Anthropic's draft prospectus"). Not in the registry: the $1.83T is the bill's registered default restated, so a change to the bill's defaults leaves this page's sentence stale; it is written as a comparison with the defaults and is not held by the checker. The design not taken: waiting for the public prospectus and building the page from filed figures. It would have needed no labels and no rewrite, and it lost on the owner's decision to have the page now, while the investor-day figures are the news; the page will be rewritten from the prospectus when it is filed. Not otherwise systemic: one new page beside the model, feeding nothing. |
| Meta Q3 2026 row in the watch-list (reported 2026-10-21, folded in 2026-10-22) | Filed. Primary source: Meta's results release, SEC 8-K exhibit 99.1 filed 21 Oct 2026 (accession 0001628280-26-066700): revenue $64,601M against $51,242M (+26%); costs $51,613M including a $9.89B charge for the August 2026 agreement with 52 attorneys general (payments of up to about $18.59B); operating profit $12,988M (−37%), margin 20% from 40%, $22.88B and 35% without the charge; net income $9,860M; adverts shown +10%, price per advert +14%; capital spending including finance-lease principal $34.91B; operating cash flow $32,786M; free cash flow −$2,120M (against $10,625M a year earlier); cash and marketable securities $82.45B; staff 66,775, −15%; Q4 revenue guide $73–76B; 2026 costs $175–179B (from $165–169B); 2026 capital spending $135–145B, "narrowed from our prior outlook of $130–145 billion"; a non-voting Class C share class effective 19 Jan 2027, distributed one for one as a dividend (a two-for-one split in effect). Meta's 10-Q for the quarter (accession 0001628280-26-066773) confirms the income statement and adds: purchases of property and equipment $33,942M in the quarter; non-cancellable contractual commitments $454.52B at 30 Sept; $24.91B of net proceeds from notes issued in May 2026; "We expect to use shares of the Class C common stock to conduct equity offerings from time to time". Second confirmation: geoveu holds no capture (its record expected the print on 28 Oct), so an independent report stands in: StockStory on Yahoo Finance and Investing.com (21 Oct) both give $64.6B and +26%; Business Today and FXStreet (22 Oct) give the −$2.1B free cash flow and the $135–145B range. Press-reported, labelled on the page: analysts' revenue expectation $63.3B (StockStory $63.35B, Investing.com $63.34B) and Q4 expectation $73.6–73.7B; close $697.43 on 21 Oct and $723.93–725.88 after hours, +3.5–3.8% (the two reports); a fall of 6–7% during trading on 22 Oct (Business Today, citing CNBC: −7.22% to $647.16; FXStreet: about 6% in the morning). No closing price for 22 Oct was available when this was written. Not used: an "$825B" commitments figure in two press reports, which is not in the 10-Q (it gives $454.52B non-cancellable) and has no stated basis. Page arithmetic: $22.88B ÷ $64.60B = 35%; free cash flow $32.79B − $33.94B − $0.96B finance leases = −$2.12B, matching the release. Gap chart: TTM revenue to Q3 2026 = $228.2B (to Q2) − $51.24B (Q3 2025) + $64.60B (Q3 2026) = $241.56B, minus the CY2023 base of $134.9B = 106.7, appended to Meta's v. Strain chart: cash capex 33.9 and operating cash flow 32.8 appended; the TTM ratio moves from (18.8 + 21.4 + 19.0 + 30.1) ÷ (30.0 + 36.2 + 32.2 + 31.9) = 89.3 ÷ 130.3 = 69% to (21.4 + 19.0 + 30.1 + 33.9) ÷ (36.2 + 32.2 + 31.9 + 32.8) = 104.4 ÷ 133.1 = 78%, and the quarter's 33.9 above 32.8 is the first in Meta's series where cash capex exceeds operating cash flow. No slope sparkline: Meta reports no AI revenue line, so the row keeps the commitment gauge, as in July. scripts/check-prints.py's DUE entry advanced to the Q4 window (~27 Jan 2027, date not yet announced). |
| meta.html refreshed for the Q3 print (Earnings watch № 10, 2026-10-22) | Page construction on the figures in the row above; no new constant. The page is rewritten for the new print, as oracle.html was in September, and refers back to July's figures (Q2 capital spending $31.1B, operating cash flow $31.9B, free cash flow $0.8B) from the July version of the page and its ledger rows. The "$140B midpoint moves the five's plans to $842.5B" paragraph restates the registry change in the next row. Its share card (og-image-meta.png, number 10) restates figures on the page only. The design not taken: a second page for the Q3 print, keeping the July page beside it. It would preserve July's reading word for word, but the site has one page per company and the July figures are kept in this ledger and in git; two Meta pages would split the menu entry and the share card. Not systemic: one page rewritten. |
| Meta's narrowed plan carried through the site (2026-10-22) | Registry change, and every page that restates it. capexShareB.Meta 137.5 → 140, capexShareTotalB 840 → 842.5, aiCapexPoolT 0.84 → 0.8425. What follows, each held by check-derivations.py: the tracking page's counter ($842.5B ÷ 365 days = $26,715 a second, $2.31B a day), its split shares and the GAP comment; booms.html ($842.5B, 2.7% of US output, 5.5 times 2023 on plans); AI_CAPEX_T on the front page (its label still reads $0.84T). The buyers' bill's first slider is Meta's share of the five's plans: 140 ÷ 842.5 = 16.6%, which now rounds to 17%, so buyersBill.directPct.default 16 → 17 and the bill at its defaults becomes $1,300B × [0.17 + 0.83 ÷ 0.675] = $1,819.5B ≈ $1.82T (from $1.83T); measuresT.bill 1.83 → 1.82, and the bill page, the FAQ, the glossary, the economies page, the reading page's Bain entry (now "$842B" and "$0.52T") and the Anthropic page follow. The front page's §2 sentence says "about $1.8T" and does not move. The checker's patterns for the capex total read whole numbers only; they now read a decimal, because $842.5B is the first total that is not whole. The design not taken: keep 16% as a rounded convention and leave the bill at $1.83T. It touches nine fewer files, but the slider's note says it starts at Meta's share, and the check exists so that the note and the slider cannot part; a convention that ignores the share would need the note rewritten and the check removed. |
This section carries no page constants and feeds no model. Every figure in it is an attributed third-party claim, reproduced as the cited author's own and load-bearing for nothing on this page.
Moved 2026-08-27, and the section retired 2026-09-03. The entries live on their own page (reading.html), newest first. Between those dates the main page kept a §8 summary table of headlines with a one-line statement of what each tests (the annotation label changed from "where it presses" to "what it tests" on 2026-08-28 — plainer, per the owner); that section is now gone, and the reading page is reached from the site nav in the main page's header instead. This section number is therefore retired rather than reused: the essay runs §1–§7. The classification below is unchanged by either move: still attributed third-party claims, still feeding nothing.
- US Bureau of Labor Statistics, Productivity and Costs, Second Quarter 2026, Revised (USDL 26-1434, released 3 September 2026), flagged by Mike Konczal on Bluesky the same day. Quoted claims, all the BLS's: nonfarm-business labour share 52.8% in Q2 2026, "the lowest level in the series, which begins in the first quarter of 1947"; real hourly compensation −3.3% in the quarter and −0.1% over four quarters; unit profits of nonfinancial corporations +43.0% annualised in the quarter (fastest since Q2 2021) and +17.8% over four quarters (fastest since Q4 2021); labour productivity +1.4% in the quarter, +2.2% year on year, and a 2.1% annualised rate since Q4 2019 against 1.5% in the previous business cycle and 2.1% long-run since 1947. None enters the constants registry. Confirmed twice per the standing rule: the BLS release read at source, and the BLS public API series PRS85006173 (nonfarm business labour share index, 2017 = 100), which is an independent retrieval of the same revised data. Konczal's post is the flag, not a confirmation.
Page construction — the quarterly path. The release publishes the level as a percentage but the retrievable series is an index, so the two are joined by a single scale factor: 52.8 ÷ 93.446 (the Q2 2026 index) = 0.56502, applied to each quarter's index. Readings used on the reading page: Q4 2022 = 96.900 → 54.75%, Q4 2024 = 96.940 → 54.77%, Q2 2026 = 93.446 → 52.80%. Hence "still for the two years to the end of 2024" (+0.02 pt) and "about two points since" (−1.97 pt over six quarters, ≈1.3 pt/yr). The rescaled figures inherit the rounding of the published 52.8%, so they are stated to one decimal and never finer. This arithmetic feeds nothing above; it exists to size the claim that the fall is a post-2024 development.
The annotation is a reading of this page's own §6, not a BLS claim. The §6 dial takes its boundary condition from the BIS: in the no-substitution scenarios labour's share does not move at any growth rate, and on the dial the share falls only as far as the bet is paid for out of wages. A labour share falling while §4 still reads today's revenue as tools and circular financing is evidence against that premise, and is explicitly not filed as evidence the bet is working. Nothing in the release attributes any part of the fall to AI.
Corrected 2026-09-03, after publication (owner's objection). The entry first said the fall showed the share "moving for reasons that have nothing to do with an AI substitution shock". That is too strong: it treats "not substitution" as "not AI", and so erases the bargaining channel — AI lowering labour's share through the credible threat of replacement, which shifts the employer's outside option in wage-setting without anyone being displaced. The entry now names three non-exclusive channels (not AI at all; substitution already happening; the bargaining/anticipation channel) and states that this release separates none of them.
The bargaining channel is the one that matters for the dial, and it makes the objection stronger rather than weaker. The page had been treating displacement in the payroll data as the signal that starts the clock; on this channel the transfer begins with the threat and never shows up in payrolls, so §4's "no displacement visible" is not evidence that the share is safe. §6 already argues the reverse leg — that a shrinking share drains labour's leverage — so the two form a loop, not a sequence. Recorded here because it changes what the page is watching for, not because any figure moved: no constant changed.
No constant moved, and none should. The 60% labour share in §4 and §6 is the BIS's advanced-economy figure for whole economies. The 52.8% is the US nonfarm business sector, which excludes government, households, non-profits and farms and has always printed materially lower. Different scope, different geography; neither updates the other.
- Mike Konczal, two posts on Bluesky (a thread — the second is his own reply to the first), 3 September 2026. Reported claim: that the record-low labour share is "a big deal, and a new post-2024 development". The 52.8% itself is not an attributed claim — it is BLS's own published figure and is confirmed against the release (USDL 26-1434) and geoveu's capture, which is why it also appears as a sourced §7 row. What stays attributed to Konczal is the break reading: that the level dates from 2024 rather than continuing a decline tracked since the 1980s. The page states where that reading is harder to hold — on the comparable whole-economy measure the euro area's labour share has been rising since 2022 while the US falls, so on present evidence the move is a US development rather than a developed-world one. Neither post claims AI as the cause and the page does not supply one. None enters the constants registry.
- Adam Posen, interviewed on Bloomberg's Odd Lots, "Adam Posen Thinks Things Could Get Very 'Messy' for the Fed", 1 September 2026. Reported claims: that there is more indicative evidence of AI in productivity than in the labour market; that hiring continues to grow in long-haul trucking and junior coding; that Brynjolfsson's J-curve and Garicano's "messy jobs" explain the absence of displacement so far; that any AI productivity boost to growth registered only in the last year and a half; and that reduced hiring of younger workers cannot be attributed to AI alone given the post-Covid reshuffling. None enters the constants registry.
Sourcing deviation, stated on the page. Bloomberg's own transcript is behind bot protection, so Posen's remarks were read from a third-party automatic transcript (podscripts.co), which misspells several of the names it reports — Brynjolfsson, Garicano, Korinek, Kolko, Ozimek and Chorzempa all appear mangled. One automatic transcript does not meet the two-confirmation rule, so the entry paraphrases throughout and presents nothing as a verbatim quote, and both substantive claims are cited to their own primary sources instead:
- The messy-jobs argument: Luis Garicano, Jin Li and Yanhui Wu, Messy Jobs: The Work That AI Cannot Reach (June 2026), messyjobs.ai — confirmed twice, the publisher's site and the HKU Centre for AI, Management and Organization launch record, which agree on authors and framing. - The 0.3% GDP claim, traced before write-up and found to be misattributed in the interview itself: it is neither Posen's nor the Peterson Institute's (Posen name-checks PIIE colleagues as working the topic) but Isabel Juniewicz, Daniel Carey, Phil Trammell and Anson Ho, "The Nvidia-sized hole in US GDP statistics", Epoch AI, 24 August 2026, read at source. Its actual claim is narrower than the interview's shorthand: US GDP growth understated by about 0.3 percentage points over the past year (not the GDP level), from a gap averaging ~$30B/quarter attributable to Nvidia through 2025, projected toward ~2 percentage points of growth by end 2028 on current trends; Epoch states it confirmed the absence with the BEA. The mechanism is the fabless case: chips designed in the US but manufactured and sold abroad generate no recorded goods export and no recorded IP export.
The entry's annotation is a reading of this page's own sections, not a claim by Posen or by Epoch: the measurement gap is routed to the §6 growth dial rather than to the bill or the gap chart, because Nvidia's value-add already appears on this page as the five's capex seen from the supplier's side, and the gap chart is built from the five's reported revenue, which national- accounts convention does not alter.
- Heather Hennerich, "How Does Productivity Affect Inflation, Jobs and Pay?", Federal Reserve Bank of St. Louis, Open Vault, 26 August 2026, interviewing St. Louis Fed senior economic policy advisor Alex Bick. Quoted claims: that a 10 percentage point rise in AI adoption among an industry's workers went with 2.9 percentage points of additional cumulative labour productivity growth (US Q4 2022 – Q3 2025; Europe 2022–24), with no clear industry-level employment effect in either direction; that wage increases typically reflect productivity and inflation; Governor Christopher Waller's claim that sustained productivity growth above 2% supports rising real incomes without inflation pressure; and farming's fall from about 41% of US workers in 1900 to under 2% today. All are the interview's and its sources', not this page's, and none enters the constants registry. Confirmed twice per the standing rule: the Open Vault piece itself, and the underlying research post it cites — Bick, Blandin, Deming, Fuchs-Schündeln and Jessen, "Mind the Gap: AI Adoption in Europe and the U.S." (St. Louis Fed On the Economy, 30 March 2026, from a paper prepared for the Brookings Papers on Economic Activity, Spring 2026) — which states the 10pp/2.9pp figure identically and adds the authors' own caveat that the estimates are not causal. The entry's annotation is a reading of the page's own §3 and §4, not a claim by the authors: the Fed's mechanism has the productivity gain passing to buyers and workers, which is the opposite direction from the §4 capture rate.
Corrected 2026-09-03, same day as publication. The entry originally called the 10pp/2.9pp result "a small number next to a bill that needs about 40% revenue growth a year". That comparison was wrong on two counts: it set a cumulative figure against an annual one without converting (2.9pp over the eleven quarters to Q3 2025 is ≈1pp/yr), and it compared an industry-level productivity measure against five companies' required revenue growth, which share no denominator. The corrected entry states the conversion and drops the comparison, keeping the point that does hold: the study does not measure capture, which is what the bet turns on. No figure and no constant changed — the error was in the comparison, not the number — and nothing in the model could have caught it, since neither quantity is a page constant.
- Adam Shaw, "Bringing robotics to life", FT Free Lunch, 23 August 2026. Quoted claims: that investors have few incentives to automate business functions where labour costs are low; and a 2026 study led by Erik Brynjolfsson reporting a 10% higher statutory pay floor correlated with 8% higher robot adoption in factories. Both are the article's, not this page's, and neither enters the constants registry. The robot-adoption figure traces to a public primary source — Brynjolfsson, Li, Miranda, Seamans and Wang, "Minimum Wages and the Rise of the Robots", Stanford Digital Economy Lab — which the entry links directly, since the newsletter edition itself was published email-first and has no public article URL.
- Kristalina Georgieva, IMF, remarks of 25 August 2026. Quoted figures: global growth 3% for 2026 (from 3.1% in April), 3.4% for 2027, Middle East cut 1.2 percentage points to 0.7%. These are the Fund's projections as reported, not this page's, and none enters the constants registry.
- Michelle Chan, Martha Muir, Rafe Rosner-Uddin and Lee Harris, "The multiplying risks of financing data centres", FT The Big Read, 26 August 2026. Quoted claims: $7tn expected into data centres by 2030; hyperscalers tapping the $11.7tn US corporate debt market; $500bn of financing for Nvidia arranged by six firms in August; a QTS bond sold at 5.7% in April against over 7.2% on its new debt in August; Munich Re's largest natural-catastrophe exposure of €8.5bn against single-site construction costs above $10bn; only a fraction of Meta's $14bn Texas data centre insured; Oracle downgraded by S&P to triple B minus with a $7bn collateral requirement from Wisconsin's utility regulator. All are the article's figures (credited within it to Infralogic, DC Byte, Gallup and S&P among others), reproduced as attributed claims; none enters the constants registry. Oracle is in the model's roster, so the promotion rule below applies with particular force: none of these figures may reach Oracle's rows or charts without reclassification.
- Ruchir Sharma, The Economics Show (FT), 11 September 2026, interviewed by Chris Giles. Reading-page entry added 2026-09-11; the argument has its own page, debt.html, and every figure attributed to him is classified in the "US debt and the bubble page" section below, which the reading entry repeats in summary. Still attributed third-party claims, still feeding nothing.
- Robin Wigglesworth, "What is the AI capex breakeven rate?", FT Alphaville, 28 September 2026, read in full. Reading-page entry added 2026-09-30. The article collects three estimates of the revenue AI capital spending needs, plus one worked example of a single lease. All are attributed third-party claims; none enters the constants registry and no page constant changed. Each was traced to its own source before write-up: - Goldman Sachs, "More AI capex, more volatility", 25 September 2026, strategists led by Ryan Hammond. The note is not public. Headline figures — five US hyperscalers on track for $800B of capex in 2026, consensus $1.1T for 2027 with Goldman's own forecast above it, roughly $300B a year of AI revenue to break even, cloud revenue annualising about $70B above the pre-AI trend in Q2 2026, announced backlogs above $1.5T, about $1T a year of spending on AI applications for solid returns — are confirmed twice: the passages the FT quotes, and two same-day reports of the note (Investing.com and Yahoo Finance, both 25 Sept 2026). The note's title and the roster (Amazon, Alphabet, Microsoft, Oracle, Meta — the BIS's five) are as reported by TechTimes, 28 Sept 2026; the entry says "reported as the same five". Single-sourced, and labelled so on the page: the four required-revenue bars ($308B at 0% ROIC, $417B at 10%, $526B at 20%, $636B at 30%, "based on average annual capex in 2026 and 2027"), the realised-ROIC chart (about 40% falling to about 25%) and the application-layer table (compute at 75% of application-layer operating expenses; EBIT margin not applied to Meta) are read from Goldman's exhibits as the FT reproduces them. Cells used on the page: hyperscaler ROIC 10% and 20% at a 30% application margin = $925B and $1,300B ("about $0.9T to $1.3T"); ROIC 30% at a 40% margin = $1,925B ("$1.9T"). The "about four times" and "about ten times" are Wigglesworth's own division. - Stijn Van Nieuwerburgh, "Financing the AI Buildout", BPEA conference draft (draft dated 4 September 2026, conference 24–25 September), Brookings Papers on Economic Activity, Fall 2026, read at source. 182.7 GW of US capacity completed 2025–32; $8.2B per 200 MW campus; $10.3T of investment over 2025–32, averaging 3.63% of GDP; initial cost $8.54T, capital base $9.61T including the return on capital committed during construction; at a 10% unlevered return, a six-year life for the 68% IT share and twenty years for the rest, the annual capital charge is 19.37% of the capital base, and at a 50% operating cash-flow margin required revenue is 38.74% of the capital base = $3.725T a year by 2032 = 9.2% of his projected 2032 GDP of $40.481T (his Table 6). Three-year chip life: $6.0T. About $5.5 per installed GPU-hour at full utilisation, against current rental prices he gives as roughly $6 to more than $10. A second write-up of the same figures (Hanno Lustig, "Nine Percent of GDP", 27 Sept 2026) agrees with the paper. - Jared Bernstein and Ryan Cummings, "AI Companies Are In a Race Against Time", 25 September 2026, read at source; the FT quotes it verbatim. Six firms (the BIS's five plus SpaceX): about $1.2T spent on AI since 2024 against $277B of revenue; $2.4T to $3.8T of revenue needed over the next six years; $520–850B in the next year; $5.7T of spending over 2026–30 requiring $13.1–18.7T over ten years; footnote 1's slow-year case ($313B next year, then $612B–$1.2T in 2028). - Dan Davies, "if it floats, flies or megaFLOPs", Back of Mind, 14 August 2026, read at source: $1.25B a month ($15B a year) of rent on Colossus 1, at most $6B a year of running costs, at most $13B to build, five-year depreciation ($2.6B a year), profit $6.4B a year, about 50% on capital, with his own caveat that it lasts only while chips are scarce. The $1.25B a month is the same press-reported figure the §5 loop carries.
Page construction in the annotation — arithmetic on the authors' figures and on registered page values, feeding nothing: - §2 floor per dollar of capital at the slider defaults: $1,300B ÷ $3,500B = 37.1 cents a year, set beside Van Nieuwerburgh's published 38.74. "Nearly three times": $3,725B ÷ $1,300B = 2.87, and $9.61T ÷ $3.5T = 2.75. - Goldman per dollar, rough and labelled rough on the page, because the capital figure behind the chart is not reproduced: if the base is the average of 2026 and 2027 capex, it is ($800B + $1,100B) ÷ 2 = $950B on consensus or ($800B + $1,200B) ÷ 2 = $1,000B on Goldman's own 2027 forecast (reported by Bloomberg, 25 Sept 2026); $417B ÷ those = 44 or 42 cents, stated as "a little over 40 cents". - Bernstein and Cummings' ten-year requirement spread evenly: $13.1T ÷ 10 = $1.31T and $18.7T ÷ 10 = $1.87T a year. The entry states that this is six firms on $5.7T against the floor's five on $3.5T, so not the same sum. - Revenue arrived against revenue required, both at Q2 2026: Goldman's $70B ÷ $308B = 23% ("under a quarter"); the gap chart's trend-adjusted $149B ÷ $319B accrued = 47% ("under half"). The two gap-chart values are the chart's own readout at the last quarter all five have reported, and the entry dates them so a later print does not make the sentence false.
The annotation is a reading of this site's own sections, not a claim by any of the authors. Two points in it go beyond agreement with §2. First, §3 compares the floor (the builders' revenue) with the software pool (what final buyers pay) and carries no mark-up for the application firms in between; Goldman's table shows one, and the entry says §3's comparison is kinder to the bet than it should be, without scaling Goldman's dollar figures to the page's base. Whether the page should carry a mark-up was left open when the entry was written, as an owner's decision about the model. Decided the same day (2026-09-30): the page models it as a second measure beneath the §2 floor, the buyers' bill, and the floor stays the main yardstick. The §2 rows above carry the construction, the decision and the design not taken; the reading entry now points to the panel, which has been on its own page, bill.html, since 2026-10-03. Second, the FT's roughly $1.5T of global software spending is a wider measure than §3's $1.2T of enterprise software; the page keeps its own figure.
- Euan Healy, Emily Herbert and Michelle Chan, "AI hyperscalers are transforming debt", FT The Big Read, 28 September 2026, read in full. Reading-page entry added 2026-10-05, with four companions from the same fortnight. All are attributed third-party claims; none enters the constants registry and no model constant changed. Each was traced as far as it would go: - The FT article. About $500bn of financing to AI-linked groups so far this year (Goldman Sachs's count); the five (Amazon, Alphabet, Meta, Microsoft, Oracle, named) about $200bn of it, and expected to issue more than $1tn of new debt over the next few years; record bond sales in Canadian dollars (Amazon, C$14bn, June), Australian dollars (Alphabet, A$5.5bn, August) and Swiss francs (Alphabet, SFr3.1bn, February); Alphabet's 100-year £1bn sterling bond of February; highly rated bonds sometimes trading at a discount to lower-rated ones; PIMCO's strategist on the repricing, "in some cases" close to junk levels, and its cause; Alphabet, Amazon, Microsoft and Meta raising 2026 capex plans to a combined $745bn in the second-quarter results; some hyperscalers promising not to issue in two consecutive quarters; HSBC's and one investor's remarks on government borrowing; Meta's use of special-purpose vehicles. Goldman's three figures are single-sourced and labelled so on the page: the same reporter repeats them on the FT's Unhedged podcast of 1 October ("Hyperscale my debt!", read in the FT's transcript), where the $1tn is given as "to about 2030"; the page uses the article's "over the next few years". The podcast's "14 billion euros" for Amazon is the article's C$14bn and is not used. - KKR, report of 30 September 2026, read through Michelle Chan's FT account of the same day: about $600bn of AI-linked debt, 6.3% of the US investment-grade market; the largest sector share averaged 2.6% over 29 years; nearly $8tn of AI infrastructure spending by 2030; a fifth of the index could end up exposed; a 3% single-issuer limit would let the largest hyperscalers ("including" the five and SpaceX) raise up to $1.7tn, leaving more than $6tn unfunded; off-balance-sheet commitments make true exposure larger. A second report (the Globe and Mail) carries the $600bn, "roughly 6 per cent" and the fifth; the $1.7tn, 2.6% and $8tn rest on the FT alone, and the page says so. The Oracle-linked loans under strain are the FT's own reporting in that article, and the page attributes it so. - Bank of England, Record of the Financial Policy Committee meeting on 25 September 2026, published 30 September, read at source. Paragraph 11: Morgan Stanley's estimate of about $450bn of global AI-related debt issuance as of early September, more than double all of 2025; JP Morgan's of about $4.1tn of AI capex financed by debt between 2026 and 2030; AI hyperscalers at 47% of sterling corporate bond issuance so far this year. Paragraphs 12 and 13 and the headline judgements: opacity and "circular arrangements" could amplify losses, and a reassessment of AI's expected productivity gains could affect sovereign debt markets as well as AI asset valuations. - Kevin Warsh, the Federal Reserve's transcript of his press conference of 16 September 2026, read at source (the passage is quoted in the debt page's rows). The Treasury secretary's agreement is as the FT reports it. PIMCO's study of 28 September and Robert Armstrong's remarks on Unhedged are the two named dissents, with HSBC's in the FT article. - Torsten Sløk, "Hyperscaler Credit Rests on One Consensus Assumption", Apollo, 21 September 2026, read at source. The text gives $600bn rising to $2tn; the chart's title gives the years (2025 and 2030), its note the five companies (Google, Meta, Amazon, Microsoft, Oracle), and its source line FactSet. "Spending passes cash flow in 2026 and stays above it through 2027" is read off the chart's lines, where free cash flow is below zero in both years; no value is taken from the chart.
Page construction in the annotation — arithmetic on the authors' figures and on the strain chart's data, feeding nothing: - The five's operating cash flow from the STRAIN object, at the last quarter each had reported by 5 October 2026 (four to June 2026, Oracle to August 2026). Calendar 2025, the four quarters ending March to December (Oracle's ending February to November): 160.5 + 164.7 + 139.5 + 115.8 + 22.3 = $602.8B, set beside Sløk's $600bn. Latest four quarters: 183.0 + 185.7 + 161.4 + 130.3 + 47.0 = $707.4B; the four before: 136.1 + 133.7 + 121.1 + 102.3 + 21.5 = $514.7B; growth 37.4%. (Order: Microsoft, Alphabet, Amazon, Meta, Oracle.) - Capital spending, same source and quarters: 116.0 + 132.5 + 173.0 + 89.3 + 75.6 = $586.4B against 64.5 + 67.0 + 107.6 + 52.1 + 27.5 = $318.7B; growth 84.0%. Spending ÷ cash flow: 586.4 ÷ 707.4 = 82.9% ("83 cents") against 318.7 ÷ 514.7 = 61.9% ("62 cents"). - Growth needed: (2,000 ÷ 602.8)^(1/5) − 1 = 27.1% a year, 2025 to 2030. The rise: 2,000 − 602.8 = 1,397, "$1.4T", set beside the §2 floor's $1.3T with the page's statement that the two are different measures. - "Nearly three times": the five's roughly $200bn of bonds this year ÷ Goldman's $70bn a year of cloud revenue above trend (the §8 row above) = 2.9. - The spending slider: the front page's "about $1.3T already spent or planned for 2025–26" + Goldman's consensus $1.1T for 2027 = $2.4T by the end of 2027; a further $1.1T a year gives $3.5T at the end of 2028 and $4.6T in 2029, against the slider's default of $3.5T and top of $4.5T. Each extra $1T of spending adds 1,000 ÷ 5 + 1,000 × 10% = $300B a year to the floor at the registered defaults. - The FT's $745bn for four companies against the registry's capexShareB for the same four: 190 + 220 + 200 + 140 = $750B ($747.5B until Meta narrowed its plan on 21 Oct 2026). - The chart under the spending-slider paragraph (added 2026-10-05, owner's request): the same path as bars, total spending from 2025 to the end of each year, against the slider's range as a band and its default as a dashed line. Its note gives the floor at three totals with the front page's other defaults: 3,500 ÷ 5 + 250 + 350 = $1,300B; 4,500 ÷ 5 + 250 + 450 = $1,600B; 5,700 ÷ 5 + 250 + 570 = $1,960B. The last is outside the slider and so appears nowhere on the front page; it is this entry's arithmetic and feeds nothing. The chart is drawn by script from two constants, SLIDER and PATH, which the checker holds to cpxSlider and spendPath. The design not taken: a static drawing with the geometry written by hand. It lost because every number in it is a copy of the slider's range or of the path, and a hand-drawn copy is one the checker cannot read; if the range ever moves, a drawn chart would go on showing the old one. - The entry is dated so that a later print, which changes the STRAIN totals, does not make its sentences false. - The second chart, its table and the paragraph above them (added 2026-10-05, owner's request): what three totals for the five's spending would ask of wages or of growth. Page construction; no new assumption and no new constant. It is the Growth Saviour's growth-or-wages arithmetic (the rows for that page, below), run on the floor each total gives: floor ÷ the §4 capture rate of 25% = what customers must gain a year by 2030; found wholly from wages it is jobs at $80,000 each, and workers' share of income is 60% less that gain as a share of 2030 output at today's pace; found wholly from growth it is the rate that lifts the 43 advanced economies from $69.6T in 2025 to today's-pace output plus the gain by 2030. At $3.5T of spending (floor $1,300B): $5.2T, 65M jobs, 53%, 3.44% a year, $82.4T. At $4.5T ($1,600B): $6.4T, 80M, 52%, 3.74%, $83.6T. At $5.7T ($1,960B): $7.84T, 98M, 50%, 4.09%, $85.1T. Today's pace is 2.1% and gives $77.2T by 2030; the 2016–25 average is 2.0% (October 2026 round; on the April and July figures these were 3.16%, 3.46%, 3.82%, 1.8% and 1.9%). The first row is the front page's own; the other two appear nowhere else on the site. "Half as much again" is 7.84 ÷ 5.2 = 1.51. The checker recomputes every figure in the chart's CASES, the table and the paragraph. The design not taken: working the figures out in the page's own script. It lost because that needs a third copy of the nine constants the arithmetic runs on, beside the front page's and the Growth Saviour's. The page types the eighteen results and the checker recomputes them, so there is still one place where the arithmetic lives. What the reader should hold in mind is in the entry: the path is the cautious one for spending, and the three settings it is multiplied through (five-year life, 10% return, 25% capture) are the front page's defaults, each on a slider there.
Two questions the entry raises; both are the owner's. First, the slider's window and range are the BIS's (§2 rows above), and the entry says outside forecasts now sit above its top. Decided the same day (2026-10-05): the range stays the BIS's and the front page says so in a sentence under the slider; the §2 row carries the sentence and the reason. Second, the debt clock (§7, "Three clocks") is a page inference from typical tenors; the reporting describes long-dated borrowing but gives no breakdown by maturity, so the entry questions the clock's wording and leaves the clock as it is.
- Bain & Company, "New Innovation Is Required to Fund AI's $6 Trillion Buildout", Global Technology Report 2026 (the firm's seventh), published 29 September 2026, read at source: the chapter page and the press release of the same day, both fetched 13 October 2026 (bain.com refuses a browser that fails its bot check, so the text was read through a fetch of the page, not in a browser), with the report's data-centre chapter, "AI Data Center Boom: Can We Build It If They Come?", read the same way, and Bloomberg's report of 29 September ("AI Faces $6 Trillion Test to Justify Data Centers, Bain Says") as coverage. Reading-page entry added 2026-10-13, with a chart. All of Bain's figures are attributed third-party claims; none enters the constants registry and no page constant changed. - Bain's figures as the entry quotes them: AI infrastructure spending of $1.5T a year by 2031, new capacity plus the upgrade of the installed base of GPUs, memory and networking; capital spending taken as sustainable at about 25% of industry revenue ("ambitious but reasonable", on the cloud providers' record), hence an AI market "approaching $6 trillion annually"; existing uses $1.2T to $1.8T (consumer subscriptions and advertising $200B to $400B; enterprise productivity gains in software development, sales, marketing, customer service and IT operations $1T to $1.4T, as revenue to providers); a remaining gap of "about $4.2 trillion"; new sources: search and advertising $100B to $200B "or more", autonomous everything $400B, physical AI $900B (on a 10% reduction in R&D and manufacturing costs), new product development with no figure; "funding it sustainably will require adding approximately 1% to the annual global GDP growth rate"; hyperscaler capex (Microsoft, Google, Amazon, Meta, Oracle: the BIS's five) "could reach $780B in 2026", nearly five times three years earlier; "the infrastructure is being built ahead of the demand curve". The chapter gives no chip life, depreciation, interest or debt. From the data-centre chapter: $5T to $6.5T of spending for nearly 150 GW of new capacity by 2030 (excluding GPU refresh and retirements); hyperscaler data-centre capex "could top $1 trillion by 2028"; local opposition "blocked or delayed at least 75 projects worth $130 billion in the first quarter of this year alone", against $156B in all of 2025. - The 2025 edition: Bain's press release of 23 September 2025 ("$2 trillion in new revenue needed to fund AI's scaling trend", the sixth report), read at source: about $500B of capital expenditure and about $2T of new annual revenue by 2030, "the world is still $800 billion short to keep pace with demand"; Bloomberg's report of the same day ("An $800 Billion Revenue Shortfall Threatens AI Future, Bain Says") as coverage. The entry says the two editions imply the same quarter ($500B ÷ $2T and $1.5T ÷ $6T are both 25%) and that the scope differs (new data centres in 2025; all infrastructure with replacement in 2026).
Page construction in the entry and its chart — arithmetic on Bain's figures and on registered page values, feeding nothing. The checker recomputes every site-side figure (the Bain block in scripts/check-derivations.py, 2026-10-13); Bain's inputs sit in the page's BAIN constant and the site's in SITE, which the checker holds to the registry. - The unnamed part of Bain's gap: $6T − $1.8T = $4.2T (Bain's own figure, at the top of the existing-uses range); the named new sources at the top of their ranges, $0.2T + $0.4T + $0.9T = $1.5T; $4.2T − $1.5T = $2.7T with no source named. The chart draws the three blocks at these upper bounds. Bain's own text leaves the remainder unquantified; the subtraction is this page's. - The comparable site measure is the buyers' bill, not the floor, because Bain's $6T is customer revenue across the industry. The bill at the registered defaults is $1,825.8B (measuresT.bill, bill.html's arithmetic): $6,000B ÷ $1,825.8B = 3.3 times. - That distance split on the reading page's usual test. Spending counted: Bain $1,500B a year; the floor's $3,500B (cpxSlider.default) over the BIS window of six years (2025–30, the span of rampYears) = $583B a year; 1,500 ÷ 583.3 = 2.6 times. Revenue per dollar of yearly spending: Bain 6,000 ÷ 1,500 = 4.0; the bill 1,825.8 ÷ 583.3 = 3.1; the floor 1,300 ÷ 583.3 = 2.2; 4.0 ÷ 3.13 = 1.3 times. 2.57 × 1.28 = 3.29, the 3.3 above. The window's average is the only yearly figure the BIS model has. On the five's 2026 plan ($840B) the two factors would be 1.8 and 1.8, so the entry says only that the first "is the larger part" on the window's average, and quotes no fraction. - Bain's rule on the five's 2026 plan: capexShareTotalB $840B ÷ 0.25 = $3,360B, $3.4T. - Growth: the entry restates the Growth Saviour's registered results, the floor paid by growth of 3.2% a year to 2030 against today's 1.8% (1.4 points) and the bill by 3.7% (1.9 points), recomputed by the checker from the same _trade as the economies page's figures. Bain's "approximately 1%" is global and comes with no working; the comparison is of order only. - The chart's right column: the floor $1.3T (measuresT.floor) and the firms in between $1.82T − $1.3T = $0.52T ($0.53T until 22 Oct 2026); the dashed line is softwarePoolT, $1.2T. The legend and the note under the chart type these, and the checker holds them.
The design not taken (the chart): a bridge chart from the bill to Bain's $6T in two steps, "spending counted" then "revenue per dollar". It lost because the split between the two steps depends on which yearly figure stands for the five (the window's average or the 2026 plan), as the sums above show, so the picture would assert a proportion the arithmetic cannot settle. The stacked columns drawn instead compare only figures each account states, and put the one similarity on the page: both accounts set today's uses at about the software pool. Not a systemic change: one entry, one chart, no new constant.
- Gartner, "Gartner Says Over $3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by the End of 2030", press release, 14 October 2026, read at source on 24 October 2026 (gartner.com refuses a plain fetch; the text was read in a browser). Confirmed by CRN Asia and The Next Web, which report the same figures. The report behind it ("Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032") is for clients only and was not read. Reading-page entry added 2026-10-24, no chart. All of Gartner's figures are attributed third-party claims; none enters the constants registry and no page constant changed. - Gartner's figures as the entry quotes them: AI providers "expected to spend nearly $1 trillion on AI infrastructure" in 2026 and "an additional $6 trillion between 2027 and 2030"; AI data-centre owners "would need to generate a cumulative $8.5 trillion in revenue to cover announced capital spending, financing costs, and a 12% return on invested capital"; AI revenue "that can be attributed to AI data centers during this period is forecast to reach no more than $3.6 trillion"; "less than half of AI revenue will be earned by providers that own significant AI infrastructure"; more than $3.5T exposed to impairment by the end of 2030 and more than $4T through 2032; the 12% not reached "until the end of 2032 at earliest"; over 80% of neocloud providers to "cease independent operations" by 2030; AI-optimised IaaS spending from $66B in 2027 to $174B in 2030; hyperscalers' "massive and highly diversified revenue sources ... could plausibly sustain AI investments for years". Quoted analyst: Will Sommer. The 80% is not a bankruptcy figure, and the entry says so, because some coverage reported it as one.
Page construction in the entry — arithmetic on Gartner's figures and on registered values, feeding nothing. The checker recomputes each (the Gartner block in scripts/check-derivations.py, 2026-10-24); Gartner's inputs sit in that block's GARTNER constant. - Gartner's spending: $1T + $6T = about $7T for 2026–30; $8.5T ÷ $7T = 1.2 dollars of revenue per dollar spent; $6T ÷ 4 = $1.5T a year for 2027–30. - The floor summed along the required ramp: rampFrac gives the share of the full floor due at the start of each year, so a year's average share is the mean of its two ends: 2026 (0.13 + 0.36) ÷ 2 = 0.245, 2027 0.44, 2028 0.60, 2029 0.76, 2030 0.92; total 2.965 years of the full floor. × $1.30T (measuresT.floor) = $3.85T; ÷ $3.5T (cpxSlider.default) = 1.1. × $1.82T (measuresT.bill) = $5.40T, "about $5.4T". - The five's 2026 plans as a share of Gartner's 2026 figure: capexShareTotalB $842.5B ÷ $1,000B = 84%, "about 85%" (rounded to the nearest 5, because Gartner's figure is itself "nearly $1 trillion"). - Share of the bill that reaches the builders: 1.30 ÷ 1.82 = 71%. If less than half does, as Gartner says, the bill exceeds 1.30 ÷ 0.5 = $2.6T a year. The bill at the computing-share slider's minimum (buyersBill.computeSharePct.min, 50%), the other two at their defaults: 1.30 × (0.17 + 0.83 ÷ (0.50 × 0.90)) = $2.62T. - Against Gartner's revenue: $3.6T ÷ $5.40T = 67%, "about two-thirds"; $3.6T × 0.5 = $1.8T reaching owners, against the summed floor of $3.85T, 47%, "less than half". The checker fails if either comparison stops being true. - Scope, stated in the entry: Gartner counts every AI provider worldwide; the floor and the bill count the BIS's five. Gartner's share is of all AI revenue; the bill's is of what customers pay for the five's computing.
The design not taken: a chart setting Gartner's $8.5T and $3.6T beside the summed floor and bill. It lost because Gartner's spending base is twice the BIS's and its method is not published, so bars on one scale would invite a comparison of totals the entry says are not comparable; the two ratios (revenue per dollar, and the share reaching the builders) are the comparisons that hold, and prose carries them. Not a systemic change: one entry, no new constant.
Entries here may be added without a constants-registry change. If a claim from this section is ever promoted into a model on the page, it stops being a reading entry and must be classified under the section it feeds.
| Page claim | Source & status |
|---|---|
| China is not an advanced economy | BIS classification, stated in the chapter itself (Ch I p. 1 discusses China under the EME heading; every cross-country graph note lists CN among EMEs, e.g. Graph 8.A's "16 AEs including EA and 22 EMEs"). Follows the IMF grouping. |
| US weight in the advanced bloc: slider 30–50%, default 37% (2026-10-03) | Page derivation. Measured shares (IMF, WEO October 2026, 2025): US GDP $30.9T of $69.6T at market exchange rates = 44.3%; on purchasing-power weights (series PPPSH), which the IMF and the BIS average growth with, 37.1% (April 2026: $30.8T of $69.4T, the same shares). The default is the purchasing-power share, for the reason in the §6 reality-check row; the range covers both measures. Until 2026-10-03 the slider ran 40–55% from 47%, on wording ("≈$30T of ≈$62T, ≈48%"; PPP "≈40–45%") that mixed years and overstated both shares. The widget's formula — bloc growth = weight × US growth + (1−weight) × rest growth — is an identity; the weight is the only assumption, hence the slider. Worked default: rest at 2.25%, target 2.5%, weight 37% ⇒ required US growth 2.25 + 0.25 ÷ 0.37 = 2.93%, US contribution 0.37 × 2.93 ÷ 2.5 = 43% of bloc growth; the 3.4% at which growth pays for the bet (§6, since 2026-10-05, on the October 2026 round; 3.2% before) at those settings ⇒ 2.25 + 1.15 ÷ 0.37 = 5.36% US growth; the 2.9% at which total pay grows as it used to ⇒ 2.25 + 0.65 ÷ 0.37 = 4.0%. With the rest near today's 2.0%, the 3.4% ⇒ ≈6%. The FAQ's history sentence is the IMF's US series (NGDP_RPCH): outside 2021 the US last grew faster than 5.4% in 1984 (7.2%), and grew 4.0–4.8% in 1997–2000. (At the old 47%: 2.78%, 52%, 4.27%.) |
| Three items added 2026-10-04: growth as a way to pay for the bet; the capture rate and its evidence; and the growth dial (the third replaced 2026-10-05: it explained the doubling, and now says what the dial measures and why it stops in 2030) | Restatements. Figures and sources are in the rows for §3 (the third source), §4 (the capture rate's evidence), §6 (dial machinery, the three rates) and the Growth Saviour (growth or wages). The checker holds the restated figures to the same arithmetic. The Tooze attribution in the first item is as in the §3 row. |
| The revenue floor and the buyers' bill, restated (added 2026-10-03) | Restated, not new. The answer "What is the revenue floor, and why do customers pay more than it?" gives the floor at $1.30T and the bill at $1.82T ($1.83T until 22 Oct 2026), both at their starting values; their derivations are the §2 rows above and the bill.html section below. The checker holds both figures to the arithmetic at the registered defaults, so a moved default stops the deploy until the answer is changed too. |
| Two confirmations for every figure, and the draft-prospectus exception (added 2026-10-03) | The answer on where the numbers come from now states the site's standing rule: two independent sources that agree, the filing winning a disagreement, press-reported figures labelled. It names the one exception, Anthropic's draft prospectus, with the label its figures carry everywhere: "From Anthropic's draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing." (see the §2 row on Anthropic). No figure. |
| US and bloc weights in the page's footer (corrected 2026-10-03) | The footer still read "≈ $30T of ≈ $62T" and "toward 40%" after the answer above it was corrected to $30.8T of $69.4T, 44% and 37%. It now carries the same IMF figures as the row above. |
debt.html, added 2026-09-11)The page reads Ruchir Sharma's argument, made on the FT's Economics Show of 11 September 2026, against the §2 floor. It carries one interactive construction (the floor with the required return split into the ten-year Treasury yield and a premium) and one measured input (the yield). Everything attributed to Sharma is an attributed third-party claim and feeds nothing.
| Page claim | Source & status |
|---|---|
| Sharma's figures: about $1T a year of AI infrastructure spending worldwide against about $200B of AI revenue, a gap of at least $800B a year, part paid from the builders' other businesses and the rest raised as bonds or new shares; a US deficit of about 6% of GDP, or about $2T of borrowing a year; federal debt of about $40T gross and about $32T net of intra-government holdings, roughly 100% of GDP against about 37% at the start of the century; an interest bill above 3% of GDP, near $1T a year, against $800B on defence; the ten-year yield about 4% in mid-February and about 4.8% at recording; the four signs (overvaluation, over-ownership, over-investment, over-leverage) with the first three ticked and the fourth left unticked; AI investment approaching 5% of GDP; about 40% of the US market as direct AI plays; every bubble in 300 years ending as a monetary event; 5% as the line, for two reasons (a twenty-year cap, and interest above nominal growth of about 5%) | Attributed third-party claims, all Sharma's as given on the show, read from the FT's published transcript (ft.com/content/2980df94-8968-48b5-89f8-05ffaed20a7b), with his column of the same week ("Why America's debt binge is starting to matter", ft.com/content/8827baa7-163d-4122-9ea8-a2c9ff6e8a4d) as the second confirmation of the argument in print. None enters the constants registry; none feeds a model. The one direct quotation on the page, that AI "has now increasingly become a capital market story", is his sentence from the transcript. The $1T is a global figure and the site's $840B is the five's 2026 guides, and the page says so where it sets them side by side. Where the page says the yield was "about 4%" in mid-February, that is the host's "just under 4 per cent" checked against FRED DGS10 (4.04% on 13 Feb 2026). |
| Ten-year Treasury yield 4.95% on 10 September 2026 (the stat strip, §2's "where the yield stands", and the slider's default) | Reported. Primary source: the US Treasury's daily par yield curve, 10-year column, 10 Sep 2026 = 4.95 (9 Sep = 4.83, 8 Sep = 4.80). Geoveu holds no record of the series, so per the standing rule the primary plus independent reports stands in: two market reports of the 10 Sep close at 4.95% (up 0.12 points on the day, the highest since October 2023) were read and agree. FRED series DGS10 and the Federal Reserve's H.15 release, which carry the same series a day behind, confirm 9 Sep at 4.83 and will carry 10 Sep after their next update. Registered as tenYearYield (value and date) and checked against the slider default, the JS constant and the page's prose. Refresh policy: the reading is dated on the page and moves only when the page is revised for a reason; it is not a live feed, and the slider is the reader's way to move it. |
| "Highest since October 2023, when it peaked at 4.98%"; "has not closed above 5% since July 2007" | Reported, from FRED DGS10 (the Fed's H.15 10-year constant-maturity series): maximum since 2008 = 4.98 on 19 Oct 2023; last observation ≥ 5.00 = 5.04 on 19 Jul 2007. Both are lookups, not constructions; neither enters the registry. Consistent with Sharma's "capped at 5 per cent for the last 20 years". |
| The split: profit investors require = ten-year yield + premium; floor = $3.5T ÷ 5 + $250B + $3.5T × (yield + premium) | Page construction. The §2 floor at its defaults, with the return written as two addends. FLOOR_DEFAULTS on the page is held equal to index.html's slider defaults and the registry by the checker, as tracking.html's copy is. The premium slider's default is returnPct − tenYearYield.pct = 10 − 4.95 = 5.05 points, so at rest the two sliders sum to the front page's 10% and the floor is the front page's $1,300B; the checker enforces that identity. Slider ranges: yield 3–7% (from below the 2026 low of about 4% to two points above Sharma's line), premium 2–8 points (from roughly the spread on the safest corporate borrowers to a distressed one), both step 0.05. Sensitivity: each point on either slider adds $3,500B × 1% = $35B a year; the page prints it from CAPEX_B / 100. Worked points at the default premium: yield 5% → 10.05% → $1,301.75B; 6% → $1,336.75B; 7% → $1,371.75B. Named limitation, stated on the page: the premium is held fixed as the yield moves, the gentle case; in a scare spreads widen with yields (the FT's QTS bond, 5.7% → over 7.2% in four months, a 1.5-point widening that on the second slider is ≈ $52B a year), and the reader can move both sliders to see it. Relation to the BIS: the BIS's own sustainability test subtracts debt service (Graph 11.B notes: net economic surplus = revenue − capex − debt service, from Rungcharoenkitkul (2026b)'s contest model with cross-firm financing and debt); this floor charges a required return on all committed capital instead, which is stricter and larger at any yield, and this page changes none of that. The point-for-point transmission of the yield into the required return (premium constant) is the standard first approximation and is stated as such. |
| The front page's slider priced at 10 / 12 / 14%: floor $1.30T / $1.37T / $1.44T, multiples 9.6× / 10.1× / 10.6× of Microsoft's office business; "about $35B a year per point, about 11% across the slider" | Page construction, the §2 arithmetic at three settings of its own slider: 700 + 250 + 3,500 × 0.10 = 1,300; × 0.12 = 1,370; × 0.14 = 1,440 ($B). Multiples ÷ officeTtmB 135.3 = 9.61, 10.13, 10.64. Across the slider's 10 → 14%: +140 ÷ 1,300 = 10.8%. |
| "Six and a half times what AI earns today" | Page construction on Sharma's figure: the §2 floor at defaults ÷ his $200B revenue estimate = 1,300 ÷ 200 = 6.5. His figure, this site's floor; labelled as such on the page. |
| Oracle as the fourth sign already ticking: $28.5B of capital spending in the quarter, operating cash flow $23.1B, free cash flow −$5.4B, TTM capex/OCF 161% and about 244% net of $11.4B of customer prepayments, a completed $20B share sale | Reported, all from the §7 Oracle FY27 Q1 row (8-K of 10 Sep 2026) and oracle.html; nothing new is constructed here. The reading of it as Sharma's "over-leverage" is the page's interpretation, stated as such. |
| §5's circuit figures repeated: Amazon's $53.4B Anthropic gain; Anthropic and Alphabet paying SpaceX "about $2.2B a month"; SpaceX's $15.8B of AI capex in Q2 | Repeated from the §3 and §5 rows above: $1.25B + $0.92B = $2.17B a month, rounded on the page to "about $2.2B". Contract terms press-reported, as the §3 row says. |
| Builders' borrowing costs cited: SpaceX bonds at 5.855% average; QTS data-centre bond 5.7% (April) to over 7.2% (August); Oracle downgraded to triple B minus | Repeated from the tracking page (SpaceX prospectus figure) and the §8 FT Big Read entry (attributed claims). None enters the registry. |
| Ten-year yield after 10 September (added 2026-10-05): first close above 5% at 5.01% on 16 Sept 2026, again on 18 Sept; above 5% on all eight trading days from 23 Sept to 2 Oct; 5.29% on 30 Sept, "the highest close since May 2002"; 5.28% on 2 Oct. Refreshed 2026-10-24: above 5% on 20 trading days in a row, 23 Sept to 21 Oct; 5.31% on 5 Oct, the highest close (also "the highest since May 2002"); 4.96% on 22 Oct, back under the line, now the stat strip's and the chart's latest reading; §2 and §5 tell the run as dated history | Reported. Primary source: the US Treasury's daily par yield curve, 10-year column: 11 Sep 4.96, 14 Sep 4.97, 15 Sep 5.00, 16 Sep 5.01, 17 Sep 4.94, 18 Sep 5.01, 21 and 22 Sep 4.96, 23 Sep 5.11, 24 Sep 5.18, 25 Sep 5.17, 28 Sep 5.24, 29 Sep 5.26, 30 Sep 5.29, 1 Oct 5.24, 2 Oct 5.28. Geoveu holds no record of the series, so the primary plus an independent report stands in: CNBC's bond report of 2 Oct 2026 has the ten-year "almost 5 basis points" higher at 5.281% and says it reached its highest since 2002 earlier that week. FRED DGS10 (the Fed's H.15 series) carries the same closes through 1 Oct and gives the history: the last close at or above 5.29 before 2026 was 5.32 on 14 May 2002, and the 2007 peak was 5.26 on 12 June 2007. Registered as tenYearYieldLatest; the checker holds the JS constant, the stat strip and the chart note to it. Refresh of 2026-10-24: Treasury closes 5 Oct 5.31, 6 Oct 5.27, 7 Oct 5.28, 8 Oct 5.22, 9 Oct 5.24, 13 Oct 5.21, 14 Oct 5.19, 15 Oct 5.10, 16 Oct 5.09, 19 Oct 5.15, 20 Oct 5.12, 21 Oct 5.13, 22 Oct 4.96, 23 Oct 5.00 (not above the line; not on the page, as no second report of it was found). FRED DGS10 carries the same closes from 23 Sept to 22 Oct. Independent report of the 22 Oct close: Seoul Economic Daily, 23 Oct 2026, 4.953% from Koscom market data, "the first time the yield had dipped below 5% since Sept. 23". The cause the page gives (a sell-off in technology shares after Meta's results, money moving into government bonds) is the same report's, and CNBC's bond report of 23 Oct is headlined the same way ("on AI worries"). 5.31 is under the 5.32 of 14 May 2002, so it too is the highest close since May 2002. |
| The stat strip shows the latest reading; the sliders stay at 10 September (decision record, 2026-10-05) | Page construction, unchanged; one reading added beside it. The premium slider's default is defined as the front page's 10% less the yield of 10 September, which is what lets the panel start at the front page's floor and claim it adds no assumption of its own. The design not taken: moving the sliders to the latest yield. It lost because it has two forms and both are worse. Keep the premium at 5.05 and the panel rests at 10.33% and a floor of $1.31T, so the page no longer starts where the front page leaves off. Or cut the premium to 4.72 to keep the 10%, and the page lowers the premium in the weeks when the reporting says lenders are raising it. The cost of the design taken: the stat strip's reading is refreshed by hand. No scheduled run reads the yield, so the reading is dated on the page and in the registry, and a stale one is visible as stale. |
| "At that yield, with the premium left where it is, the floor is $1.30T a year, within $1B of the front page's" (since 2026-10-24; until then "$1.31T … about $12B more", at 5.28%) and "At the highest close, 5.31% on 5 October, the floor was $1.31T, about $13B more" | Page construction, the panel's own arithmetic at one setting: 700 + 250 + 3,500 × (4.96 + 5.05) ÷ 100 = 1,300.35 ($B), so $1.30T, 0.35 above the front page's 1,300, "within $1B". At 5.31: 3,500 × 10.36 ÷ 100 = 362.6, floor 1,312.6, "about $13B more". (The old reading: at 5.28, 1,311.55, "about $12B".) The checker recomputes the sentence at the latest reading from the registry, and words the gap "within $1B" when it rounds to zero; the 5.31% sentence is dated history and is not checked. |
| §4's "Where the cause may run both ways": Kevin Warsh, the Bank of England's Financial Policy Committee, PIMCO (added 2026-10-05) | Attributed third-party claims, each read at source; none enters the registry. Warsh: the Federal Reserve's transcript of his press conference of 16 Sept 2026, where he gives three reasons for higher long-term yields (a stronger economy, competition for capital, geopolitics) and says of the second that the hyperscalers are raising funding, "the competition for capital is real", and it partly explains the rise in yields. Bank of England: Record of the FPC meeting of 25 Sept 2026, published 30 Sept, headline judgements and paragraph 13 (growth prospects and fiscal outlooks depend in part on expected AI productivity gains, so a reassessment could affect sovereign debt markets as well as AI asset valuations). PIMCO: "The Credit Market Lens: Hyperscalers Are Repricing, Not Displacing (So Far)", 28 Sept 2026, which finds the evidence for investors rebalancing from Treasuries into AI bonds weak. "The trigger is not wholly outside the industry" is the page's reading, stated as conditional on Warsh being right. |
| §5's second reading: the builders pay more above the government's rate than similar companies, "in some cases nearly what a junk-rated company pays", because the amount of debt to come cannot be predicted | Attributed claim, from the FT's Big Read of 28 Sept 2026 (the §8 entry of 2026-10-05): PIMCO's credit strategist on the repricing and its cause, and the FT's own statement that highly rated bonds sometimes trade at a discount to lower-rated ones. The reporter repeats the reason on the FT's Unhedged podcast of 1 Oct. No figure is given, and the page says so: the FT's chart of the extra rate was not legible in the copy read, and the only published numbers found (a Reuters count of about a tenth of a point of widening, to August) predate the move the article describes. |
| §2 and §5: SpaceX's premium and insurance cost of 7 Oct 2026, and Oracle's (added 2026-10-08) | Reported, press not filing, with the sources in the §7 row "spacex.html note of 8 Oct 2026". Page construction in §5: the widening on SpaceX's 2056 bond, about 2.4 − 1.75 ≈ 0.6 of a point, × $35B a year per point (the panel's own slope: 1% of $3.5T) = ≈$21B a year, what the floor would rise by if every dollar of the $3.5T paid that wider premium. The same arithmetic as the QTS sentence in §3's note (1.5 points → $52B). A comparison, not a forecast, and the page says "if every builder paid it". Not held by the checker, as the QTS figure is not: both are arithmetic on the registered $3.5T default and a press figure. §2 gains one sentence pointing at the same figures. Nothing on the panel moves. Not a systemic change. The design not taken: raising the premium slider's default by 0.6. The decision record above keeps the sliders at 10 September so that the panel at rest is the front page's floor, and one company's premium, the sixth builder's at that, is not the sector's. |
bill.html, added 2026-10-03)The page carries the buyers' bill (its construction, defaults, Anthropic check and decision record are the §2 rows above) and one construction of its own: the jobs arithmetic run on the bill beside the floor. Nothing on any other page reads its figures.
| Page claim | Source & status |
|---|---|
| Jobs table: the floor $1.30T → $5.2T of wages, 21% of the pool, 65M jobs-equivalent; the bill $1.82T → $7.3T, 29%, 91M | Page construction, the §4 identity (wages = revenue ÷ capture rate; jobs = wages ÷ $80k) run twice at the essay's §4 defaults, which the checker holds equal on both pages (capture 25%, pool $25T since the output base moved to $69.4T on 3 Oct; it was $22T, giving 24% and 33%). Bill: $1,819.5B ÷ 0.25 = $7,278B → $7.3T; 7.278 ÷ 25 = 29.1% → 29%; $7.278T ÷ $80k = 91.0M → 91M (from $1,825.8B until 22 Oct 2026; the rounded figures did not move). Floor: 1,300 ÷ 0.25 = $5.2T; 5.2 ÷ 25 = 20.8% → 21%; 65M. The bill row uses the live bill, unrounded, where the front page's §4 on the morning of 3 Oct rounded it to its slider's $0.1T step first (1.8 → $7.2T, 90M); the difference is that rounding, not a changed input. The checker recomputes every figure in the table and the stat strip from the registry. |
| "Raises the jobs count by about two-fifths"; "about $1.4 for every $1 the builders need" | Page construction: 91.3 ÷ 65 = 1.40 and 1,825.8 ÷ 1,300 = 1.40; the jobs figure scales with the bill because both rows share the capture rate and salary. |
| Why the front page keeps to the floor: one bank's note, not public, read through one newspaper; one lab's draft accounts seen by one news agency | A reading of the sources in the §2 rows above (Goldman via FT Alphaville; Anthropic's draft prospectus via Reuters, not an official filing). No new figure. |
economies.html, added 2026-10-03 as "The economies in the bet")The page sets out the IMF's advanced economies, the group every economy-wide figure on the site stands for, country by country. Its 2025 total is the front page's output base (aeGdpT, since 2026-10-03), which the checker holds equal to this page's data.
| Page claim | Source & status |
|---|---|
| The group is 43 economies | IMF-sourced, confirmed twice. IMF, World Economic Outlook, April 2026, Statistical Appendix, Table B1, lists 43 advanced economies (the October 2026 round's data sum to the same group's total, so the list stands): the United States, 21 euro-area members (Bulgaria now among them) and 21 others (Liechtenstein and Andorra now among them). The second confirmation is the IMF's own arithmetic: the 43 members' GDP in current dollars (series NGDPD) sums to the IMF's group total (ADVEC) in every year 2016–2026, to within rounding. scripts/sync-gdp.py refuses to write if that stops holding, which is how a change of membership in a new round is caught. |
| Each country's output, 2016–2026, in 2025 dollars | Page construction on IMF data. For each economy, level(2025) = NGDPD 2025 (current US dollars at 2025 market exchange rates); earlier years divide back by the IMF's real growth rate (NGDP_RPCH) one year at a time, and 2026 multiplies forward by the 2026 forecast. So every year is in 2025 prices and exchange rates, and the 2025 column equals current dollars. The total is the sum of the 43. The group's growth in this construction weights countries by 2025 market-rate output, which is not the IMF's own method for averaging growth (it uses purchasing-power weights), so the total's growth rate here can differ slightly from the IMF's headline rate for the group. 2026 is the October 2026 forecast and is labelled so. Second source: the World Bank's GDP in current US dollars (NY.GDP.MKTP.CD) for the latest year it publishes for most members; each country within 5% of the IMF and the covered total within 2%, or the script writes nothing. Cross-checked 3 Oct 2026 (after the World Bank's API had been down most of the day): for 2025 the World Bank covers 40 of the 43, every one within 5% of the IMF, and their totals are $68,455.4B (IMF) against $68,435.6B (World Bank), 0.03% apart. It has no 2025 figure for Taiwan (which it never covers), Liechtenstein or San Marino; their only check is the IMF's total. Cross-checked 22 Oct 2026, October 2026 round: 39 members against the World Bank, every one within 5%, totals $67,993.0B (IMF) against $67,713.9B (World Bank), 0.41% apart; the same three uncovered. Ireland is checked against its statistics office instead (named exception, 2026-10-22). The IMF's October figure for Ireland's 2025 output is $680.4B (April: $718.1B); the World Bank's is $721.7B, more than 5% away, which stopped scripts/sync-gdp.py. The CSO's Annual National Accounts 2025 put GDP at €602,441M (PxStat table NA005, "Gross domestic product at current market prices"); at the ECB's 2025 average of $1.12998 per euro that is $680.7B, 0.05% from the IMF. The World Bank holds €638.7B in euros for the same year (NY.GDP.MKTP.CN, updated 8 Oct 2026), matching neither, so the statistics office settles it, as the figures rule says the filing does. The exception lives in NATIONAL_CHECKS in scripts/sync-gdp.py, keyed to Ireland and 2025, and the data block records it under crossCheck.national; it goes quiet when the base year moves on. The design not taken: widening the 5% tolerance, which would also let through the next real error in any of the 43; and fetching the CSO and ECB figures live, which adds two services that can fail, for one country. |
| 2025: $69.6T for the group; the US 44%, the 21 euro countries 26% | IMF NGDPD 2025, October 2026 round: ADVEC $69,605.3B; United States $30,861.3B (44.3%) (April: $69,386.7B and $30,767.1B); euro-area members summed from the same series. The markup's figures are held equal to the data block by scripts/check-pages.py. |
| 2022: $59.4T; output after inflation up 6% from 2022 to 2025 | IMF NGDPD 2022: ADVEC $59,396.4B (October 2026 round; April: $59,358.6B), the year the front page's former round $60T matched. The 6% is this page's 2025-dollar total, 2025 against 2022. The rest of the rise in current dollars (+16.9%) is prices and exchange rates. |
| The US share at purchasing-power parity: 37% | IMF series PPPSH (share of world GDP at purchasing-power parity), 2025: United States ÷ advanced economies = 37.1%, against 44% at market rates. Since 2026-10-03 it is the US weight in §6's growth chart and the FAQ calculator's default (aeUsGrowthWeight = 0.37), which the checker holds to this page's usPppSharePct; before that they used 45% and 47%. |
| Growth or wages: the second chart (rebuilt 2026-10-04, owner's decision) | Page construction on the bet's own numbers; no new assumption. Required revenue by 2030 (the §2 floor, $1,300B, or the bill at bill.html's defaults, $1,825.8B) ÷ the §4 capture rate (25%) = what customers must gain a year: $5.2T (floor) or $7.3T (bill). The gain is either wages no longer paid or output that would not otherwise exist. "Today's pace" is the IMF's 2026 growth forecast for the group, 2.1% in the October 2026 round (the last point of §6's reality chart; 1.8% before), carried five years from 2025 to 2030, the end of the BIS's capex window when the floor falls due: 69.605 × 1.021⁵ = $77.23T. Growth that supplies the whole gain: ((77.23 + V) ÷ 69.605)^(1/5) − 1 = 3.44% a year for the floor, "3.4%", and 3.96% for the bill, "4.0%": 1.3 and 1.9 points above today's pace, 1.7× and 2.0× the 2016–2025 average of 1.97%. At a growth rate g the extra output is 69.605 × ((1 + g)⁵ − 1.021⁵), all of it credited to the bet (generous: growth has other causes); the remainder comes from wages. Wages alone: 5.2 ÷ $80k = 65M jobs' worth (bill: 7.303 ÷ $80k = 91M); workers' share in 2030 = 60% − wages displaced ÷ 2030 output = 60 − 5.2 ÷ 77.23 = 53% (bill: 50.5%, "51%"). Each extra 0.1 point of growth a year: 69.605 × (1.022⁵ − 1.021⁵) = $0.38T by 2030 = 4.7M jobs' worth. The US column: 1.98 + (need − 1.98) ÷ 0.37 = 5.9% and 7.3%, with everyone else at §6's 2.0%. Dollar figures are treated as today's prices throughout; the floor is in the dollars of the spending and output in 2025 dollars, an approximation. The checker holds every constant (BET) to the front page, bill.html and the registry, and every typed figure here and where restated (§3, §6, the FAQ, tracking.html, debt.html, bill.html, the share card). |
| Growth or wages: the BIS check and the likelihood note | BIS check (Ch I p. 19; Box C pp. 20–21): the BIS reports task-level time savings of 20–50% but aggregate productivity growth estimates of "less than 1% over a long horizon" (read here as an annual rate; if it is a cumulative gain the case for growth is weaker still), against the 1.3 points the floor needs. Box C's bounded-productivity-boost scenario is the growth case: trend growth permanently higher, labour share flat, which the BIS notes "has not occurred since the Industrial Revolution". The page needs only a rise in the level of output by 2030, not a permanent rise in the rate, and says so. Workers' share is held at 60% for the part growth supplies, following that scenario. The design not taken: the chart this replaced, the §6 dial solved for growth (wages keep pace at 3.2%, share holds at 4.7%, built 2026-10-03). It answered a question about a stylised fifteen-year claim by capital that is not derived from the spending, and its "share holds" rate exists only because the dial fixes capital's income; in the BIS's model no growth rate restores the share once work is replaced. The rebuilt chart asks the question the site's argument turns on: can growth pay for the bet in place of wages. |
| Growth or wages: the target stated as a level (wording, 2026-10-05) | The claim now gives the level the economy must reach, because only the level matters, not the path: output of 77.23 + 5.2 = $82.43T, "about $82.4T", a year by the end of 2030, where today's pace reaches $77.2T (October 2026 round; $81.1T and $75.9T before). The 3.4% is the average rate over the five years that gets there; growth could start slowly and build. After 2030 the faster rate is not needed: growth at today's 2.1% keeps the extra output (5.2 × 1.021⁵ = $5.8T five years on). The earlier wording, "3.2% a year until 2030", left the path and the afterwards unstated. |
| If the growth comes late (added 2026-10-05) | Page construction on the same inputs (floor $1.30T, capture 25%, pace 2.1%, $80k a job, 2025 output $69.605T; October 2026 round). Output runs at today's pace until the faster growth starts, at a steady rate r until it is $5.2T above today's path, and at today's pace after: r = ((trend(end) + 5.2) ÷ trend(start))^(1/(end − start)) − 1, with trend(y) = 69.605 × 1.021^(y − 2025). "Supplied by the end of 2030" = output(2030) − trend(2030); "still missing" = 5.2 − supplied, shown as jobs' worth (÷ $80k) and as the floor left unpaid (× 25%). One year late (2027–2031): r = 3.41%; supplied $4.05T, 78%; missing $1.15T = 14M jobs' worth, or $0.29T of the floor. Two years late (2028–2032): 3.39%; $2.96T, 57%; $2.25T = 28M, or $0.56T. Over ten years (2026–2035): 2.70% (0.60 points above today's pace); $2.31T, 44%; $2.89T = 36M, or $0.72T, which is twice the required return of $3.5T × 10% = $0.35T a year; unpaid floor summed over the year-ends 2030–2034: 0.72 + 0.59 + 0.45 + 0.31 + 0.16 = $2.2T. A later start needs a slightly lower rate because it compounds from a larger economy. BIS check: the ten-year rate's 0.6 points sit inside the BIS's surveyed "less than 1%" for aggregate productivity growth (Ch I p. 19), the on-time 1.3 points do not; the comparison with canals, railways and the dot-com boom, and the phrase "boom-bust cycles are a regular feature of past major innovations", are Graph 11.C and its text (Ch I pp. 22–23). The reading that late growth spares workers but not the first investors is this page's, from the five-year chip life in §2 and the end-2030 due date; it treats the unpaid floor as falling on investors or wages, and does not model which. The checker recomputes every figure in the table and notes. |
| The country chart's two paths to 2035 (2026-10-05; replaces an extension of fifteen years at 3% a year) | Page construction on the growth-or-wages arithmetic, labelled on the chart as not a forecast. A button adds two paths after the IMF's 2026 forecast. Today's pace: trend(y) = 69.605 × 1.021^(y − 2025), which is $77.2T in 2030 and $85.7T in 2035 (October 2026 round); each country is drawn at its 2026 share of that path, because projecting each separately would need country forecasts nobody publishes. Growth that pays for the bet: 69.605 × 1.0344^(y − 2025) to 2030, which is $82.43T, then 2.1% a year: $91.5T in 2035. The band between them is the extra output: $5.2T a year in 2030 (the growth-or-wages row), 5.2 ÷ 77.23 = 6.7%, "about 7%", of 2030 output at today's pace, and 5.2 × 1.021⁵ = $5.8T in 2035. The note's other figures: the 43 averaged 2.0% a year 2016–2025 (this page's total, 1.97%), so 3.44 ÷ 1.97 = "1.7 times"; and because the IMF's forecast for 2026 is today's pace (this page's total: $70.98T), reaching $82.43T by 2030 from there takes (82.43 ÷ 70.98)^(1/4) − 1 = 3.8% a year from 2027. The checker recomputes every figure in the note. Why the old extension went: fifteen years at 3% a year was the front page's old dial drawn as a chart, not the bet's arithmetic. It reached $110.0T in 2041 and suggested that the growth must last fifteen years, where the bet needs a level of output by 2030. The decision and the design not taken are in the §6 record. |
booms.html, added 2026-10-05)The page sets the five's spending beside earlier investment booms on two measures. One is the BIS's own: how many times investment multiplied from its low point before the boom. The other, the share of the economy, is this site's addition. Nothing on the page feeds the floor or any other model. Its inputs are the booms block in the registry below, and the checker recomputes every figure the page states from that block.
| Page claim | Source & status |
|---|---|
| The BIS's comparison (§1) | BIS-sourced. Ch I pp. 22–23 names four earlier episodes: the canal mania of the 1830s, the British railway mania of the 1840s, the electrification exuberance of the late 1920s and the dot-com boom of the late 1990s. Its common trait: a genuine breakthrough that attracted more capital than commercial returns could justify, ending in a reversal of investment and a recession. Graph 11.C plots each as multiples of its pre-boom trough against years since the trough. Its notes (p. 35) give the series and the trough years: US canal construction spending, 1835; GB real investment, 1843; US private fixed asset investment, 1921; US private fixed investment in information processing equipment and software, 1995; AI hyperscalers' capital expenditure, 2023. The graph's vertical scale runs from 0 to 4. |
| The share of the economy as a measure (§2) | Named addition, stated on the page in §1 and under the chart. The BIS does not compare the booms by share of GDP. Each bar is one year's investment at the boom's peak ÷ that country's GDP in the same year. |
| US computers and software: 3.3% of GDP in 1995, 4.4% in 2000, 4.5% in 2025 | Reported, confirmed twice. US Bureau of Economic Analysis, national income and product accounts, annual, read from the BEA's own data file on 5 Oct 2026. Series A679RC, private fixed investment in information processing equipment and software, which is the series the BIS's note names: $253.8B (1995), $450.6B (2000), $1,404.0B (2025). Series A191RC, GDP: $7,639.7B, $10,251.0B, $30,861.3B. Second source: the St Louis Fed's copy of both series (identical), and for GDP the World Bank (NY.GDP.MKTP.CD), identical for 1995 and 2000. For 2025 the World Bank gives $30,769.7B and the IMF $30,767.1B; both predate the BEA's latest revision, and on either the five's share below still rounds to 2.7%. Arithmetic: 253.8 ÷ 7,639.7 = 3.32%; 450.6 ÷ 10,251.0 = 4.40%; 1,404.0 ÷ 30,861.3 = 4.55%, which prints as 4.5%. |
| US telecom build-out, 2000: 1.4% | Page construction on reported figures. BEA series C275RC, private fixed investment in communication equipment, $123.3B, plus C314RC, communication structures, $18.8B: $142.1B ÷ $10,251.0B = 1.39%. The equipment alone is 1.20%, the "1.2%" usually quoted for the telecom boom. Second confirmation that these are the right series and the right size: the Federal Reserve Bank of Richmond's study of the boom (Couper, Hejkal and Wolman, Economic Quarterly 89(4), 2003) describes it with the same two, in 1996 dollars: communications equipment from about $62B a year in early 1996 to over $135B in late 2000, and telecommunications structures from about $12B through most of the 1990s to over $21B in late 1999. |
| British railways, 1847: 6.7–7.3% | Other people's estimates, two that agree to within a point; both read at source on 5 Oct 2026. Andrew Odlyzko, "Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s" (2010): railway investment of £44M in 1847 against GDP of £600M, 7.3%. B. R. Mitchell (1964), as cited by Gareth Campbell, "Deriving the Railway Mania", Financial History Review 20(1), 2013: 5.7% of GDP in 1846, 6.7% in 1847, 4.7% in 1848. The bar is solid to 6.7 and shaded to 7.3. |
| US railroads, 1870s and 1880s: 2–6% | Other people's estimates, which disagree. Drawn as a range and labelled as one; not confirmed to the two-source standard, because the sources do not agree. Low end: Goldman Sachs, peaks of "roughly 2% to 3% of GDP" for railroads, electrification and automobiles, as reported by Yahoo Finance on 11 June 2026. High end: 6.0% for 1870, from Melville Ulmer, Capital in Transportation, Communications, and Public Utilities (1960), as quoted by Tomasz Tunguz on 6 Nov 2025; Smead Capital quotes 6% for the 1880s. Wells Fargo's estimate, about 3% in the 1850s (Fortune, 13 May 2026), falls inside the range. A figure of 7.7% also circulates without a source and is left out. The page says "no agreed figure" in the table and "not a measurement" in the text, and the footer labels both ends as reported estimates. |
| The five's 2026 plans: $842.5B, 2.7% of US output | Page construction. capexShareTotalB (§7, the five's own 2026 plans) ÷ BEA GDP for 2025: 842.5 ÷ 30,861.3 = 2.73% (840 ÷ 30,861.3 = 2.72% until Meta narrowed its plan on 21 Oct 2026). Two things make it a little too high as a share of the United States, and the page states both: the five spend worldwide, and this year's plans are divided by last year's output. |
| The five's capital spending: $154.2B (2023), $239.1B (2024), $411.5B (2025) | Reported. Cash paid for property and equipment, from each company's cash-flow statements as filed with the SEC, read through the SEC's company-facts data on 5 Oct 2026, by calendar year. Microsoft (year-end June) is its fiscal year less that year's first half plus the next year's first half. Oracle (year-end May) is taken on the quarters ending February, May, August and November, as the strain chart takes it. 2023: Alphabet 32.25, Amazon 52.73, Meta 27.05, Microsoft 35.20, Oracle 6.93, total 154.2. 2022 was 157.8, so 2023 is the low point, as the BIS's note has it. 2024: 52.53, 83.00, 37.26, 55.55, 10.74, total 239.1. 2025: 91.45, 131.82, 69.69, 83.09, 35.48, total 411.5. Second confirmation: the strain chart's quarterly figures on tracking.html, each confirmed when its results were folded in, add up to $239.2B for 2024 and $411.5B for 2025; the 0.1 is quarterly rounding, and the checker holds the two pages to each other. 2023 is before the strain chart begins. Its cross-check is Epoch AI's independent compilation for the same five companies ("Hyperscaler capex has quadrupled since GPT-4's release", 26 Feb 2026), which counts cash spending plus assets taken on finance leases: $161.3B (2023), $260.4B (2024), $448.2B (2025). Those are higher than the cash figures here by the leases, in every year, and give a 2023-to-2025 multiple of 2.78 against 2.67 here. The page uses cash only, to match the strain chart. |
| Speed: 1.8 times in five years; 2.7 times in two; 5.5 times in three on plans (§3) | Page construction, on the BIS's measure and the BIS's two series and trough years. Dot-com: 450.6 ÷ 253.8 = 1.78 in 2000, the peak; the series then falls in 2001 and 2002. AI: 411.5 ÷ 154.2 = 2.67 in 2025; 842.5 ÷ 154.2 = 5.46 on the 2026 plans (840 ÷ 154.2 = 5.45, shown as 5.4, until 21 Oct 2026), above the top of the BIS's scale. The plan is on the companies' own definitions and the earlier points are cash spent; the chart draws the plan as an open circle and the note says so. Neither series is adjusted for inflation. The page also says the comparison is kind to the older boom, since five firms can grow faster than a national total. |
| Five-year equipment | The front page's §2 life slider at its default, 5 years (see §2 above). The checker fails if that default moves and this page still says five. |
| What repeats and what does not (§4–§6) | Attributed readings; none is a figure of this site's. BIS Ch I p. 22: the race may be driven by the belief that few firms will win. Couper, Hejkal and Wolman (2003): the four established carriers held 72% of US long-haul fibre in 1996 and 30% in 1999, new firms having built much of the rest; executives said throughout the boom that internet traffic was doubling every 100 days, which Odlyzko worked out to mean 17 million-fold growth from 1994 to 2000; significant overcapacity in long-haul fibre was clear by late 2000. Odlyzko on the 1840s: capital came from private sources; shareholders lost heavily, hardly any lines went bankrupt, almost all stayed in service, and it is "quite likely" the nation gained. Odlyzko, "This time is different: An example of a giant, wildly speculative, and successful investment mania" (2010), on the 1830s: prospects extrapolated from a couple of early lines; investment of about 2% of GDP a year in 1838–39; costs twice the plan; investors earning above-market returns within a few years. Each is one named source, reproduced as its author's finding, as Sharma's figures are on the debt page. Two figures are restated from the reading page and held to it by the checker: Apollo's $600B rising to $2T, and Van Nieuwerburgh's 3.6% of US GDP a year. |
| The design not taken (2026-10-05) | Three designs lost. An entry on the reading page with one chart, which was the first proposal: the owner chose a page. The reading page holds other people's arguments, and this is the site's own comparison, with two charts and a table. Redrawing all five lines of the BIS's Graph 11.C: the canal series and the 1920s series could not be read at source and confirmed twice, so the page draws the two that could and says which. A section on the front page: the front page keeps one measure and one argument (owner's decision, 2026-10-03), and links here from §5's sentence on the parallels. The change is not large but it does add a dependency: the page's 2.7% and 5.4 times follow from capexShareTotalB, so results that change a company's 2026 plan now stop the deploy until booms.html, its share card and the registry follow. |
glossary.html, written 2026-10-03, merged 2026-10-05)| Page claim | Source & status |
|---|---|
| Definitions of the site's own terms (the floor, the buyers' bill, the firms in between, the direct share, the two pools, the capture rate, the jobs equivalent, the loop, the growth dial, the required ramp, upper limit and trend-adjusted, the strain ratio, the clocks, beside the model, the premium) | Restated, not new. Each definition describes a construction already recorded in this ledger, in the row for its section, and adds no arithmetic. Figures are kept out where an illustration will do: the capture rate is explained with "a quarter", which holds whatever the slider's default. |
| The floor's worked sum: $3.5T ÷ 5 years = $0.70T, plus $0.25T of running costs, plus 10% of $3.5T = $0.35T: $1.30T a year; the buyers' bill, $1.82T a year | Page construction, restated from §2 and the bill.html section. The checker rebuilds the sum from the registered capex default, the essay's life and return sliders and opexFixedB, and holds the bill to the same arithmetic as bill.html. |
| Software pool "about $1.2T a year worldwide"; the US at "44% in dollars, 37% on purchasing power" | Restated from §3 (softwarePoolT) and the economies page (aeUsGrowthWeight); the checker holds the software pool and the 37% to the registry. |
| BIS terms: hyperscaler, circular financing, Box C and its four scenarios, r-star | From the BIS. Circular financing: Ch I p. 25, in the text and in note 2 under Graph 13: chip makers and hyperscalers take stakes in AI labs, which commit in turn to multi-year purchases of chips or computing, "rechannelling capital back to investors as revenue". Box C (pp. 20–21): business as usual (labour stays the binding input; output grows at its 2% trend), a bounded productivity boost (trend growth shifted by a constant margin), transformative AI (a self-reinforcing growth engine; the labour share falls towards zero) and the demand bottleneck (every displaced worker is a lost consumer; growth rises, then falls below trend). R-star is the report's "natural rate of interest". "Hyperscaler" is the report's word for the five (Graph 11.A's note names them); it gives no formal definition, so the glossary uses the site's. |
| Economic and financial terms (labour share, purchasing power, free cash flow, operating cash flow and the like) | Standard definitions in plain words; no figures beyond those restated above, the 60% labour share (§6) and Box C's 20% by 2060 (§5). |
| Entries brought up to date at the merge (2026-10-05) | Restated, not new. The glossary and the FAQ changes were written on 3 Oct and joined main on 5 Oct; the three FAQ rows dated 2026-10-03 above arrived with them. Between those dates §3 named growth as the other way the bet could be paid for (4 Oct) and §6's dial was rebuilt to run to 2030 (5 Oct). Four entries described the pages as they had been. "The bet" and "The two pools" said the revenue could only come from software budgets or wages. "The growth dial" described the fifteen-year dial and the doubling of capital's income. "Transformative AI" called it the world the bet needs to win. They now restate §3 and §6 as they stand, and "Transformative AI" uses the wording of §5's chart, "replacing workers pays off". The link to the economies page uses the name it has had since 3 Oct, The Growth Saviour. The dial's entry carries no figure, so it adds nothing for the checker to hold. The design not taken: merging the entries as written and correcting them afterwards. A deploy ships the whole of main, so any deploy in between would have published definitions the pages contradict. |
scripts/check-derivations.py verifies these against index.html, tracking.html, debt.html, bill.html and booms.html on every deploy, and the figures faq.html and glossary.html restate in prose. Change them here and on the page in the same commit, with a row above explaining why.
{
"opexFixedB": 250,
"officeTtmB": 135.3,
"cpxSlider": { "min": 3, "max": 4.5, "default": 3.5 },
"spendPath": { "to2026T": 1.3, "perYearT": 1.1 },
"wcpoolSlider": { "min": 20, "max": 35, "default": 25 },
"aeGdpT": 69.6,
"growthDial": { "minPct": 2.1, "maxPct": 4, "defaultPct": 2.1, "oldPayGrowthPct": 2 },
"captureEvidence": { "agentforceUsdPerConversation": 2, "finUsdPerOutcome": 0.99, "liveContactUsd": 8.01 },
"softwarePoolT": 1.2,
"buyersBill": {
"directPct": { "min": 0, "max": 50, "default": 17 },
"computeSharePct": { "min": 50, "max": 100, "default": 75 },
"marginPct": { "min": 0, "max": 40, "default": 10 }
},
"measuresT": { "floor": 1.3, "bill": 1.82 },
"aiCapexPoolT": 0.8425,
"aeUsGrowthWeight": 0.37,
"tenYearYield": { "pct": 4.95, "date": "2026-09-10" },
"yieldSlider": { "min": 3, "max": 7, "default": 4.95 },
"premiumSlider": { "min": 2, "max": 8, "default": 5.05 },
"tenYearYieldLatest": { "pct": 4.96, "date": "2026-10-22" },
"capexShareB": { "Microsoft": 190, "Amazon": 220, "Alphabet": 200, "Meta": 140, "Oracle": 92.5 },
"capexShareTotalB": 842.5,
"rampYears": [2025, 2026, 2027, 2028, 2029, 2030, 2031],
"rampFrac": [0, 0.13, 0.36, 0.52, 0.68, 0.84, 1.0],
"cy2023BaselineB": { "Microsoft": 227.6, "Alphabet": 307.4, "Amazon": 574.8, "Meta": 134.9, "Oracle": 51.6 },
"trendGrowth": { "Microsoft": 0.115, "Alphabet": 0.087, "Amazon": 0.118, "Meta": 0.157, "Oracle": 0.121 },
"booms": {
"usGdpB": { "1995": 7639.7, "2000": 10251.0, "2025": 30861.3 },
"itInvestB": { "1995": 253.8, "1996": 279.2, "1997": 316.5, "1998": 349.3, "1999": 396.4, "2000": 450.6, "2001": 423.6, "2002": 389.3, "2003": 397.7, "2004": 422.1, "2005": 445.6, "2025": 1404.0 },
"telecom2000B": { "equipment": 123.3, "structures": 18.8 },
"fiveCapexB": { "2023": 154.2, "2024": 239.1, "2025": 411.5 },
"britishRailways1847Pct": { "low": 6.7, "high": 7.3 },
"usRailroadsPct": { "low": 2, "high": 6 }
}
}This page is generated from DERIVATIONS.md, the canonical machine-checked ledger.