Five companies are spending more than a trillion dollars on AI. This page works out what that spending must earn back, where the money could come from, and what each answer means for wages. Every assumption is on a slider, so you can change it.
11 Sept 2026, night — will US debt burst the AI bubble? A page beside the model. Ruchir Sharma’s argument on the FT’s Economics Show of the same day (the government and the AI builders borrowing from the same savings; the ten-year Treasury yield as the number to watch, with 5% as the line) gets a page of its own, reached from the nav in the header. It sets his $800B funding gap beside the §2 floor (the same fact counted two ways), lays out his five-step route from the Treasury market to the data centre from the FT’s published transcript, puts his mechanism on a slider (the §2 return split into the Treasury yield and the premium above it, so the bond market moves the floor by about $35B a year per point at the default $3.5T), and reads the 10 Sept Oracle print as his fourth sign, over-leverage, ticking. The sliders rest at the front page’s defaults, so nothing here changed: the floor is still $1.30T at 10%. The only constants added are the yield reading (4.95% on 10 Sept, from the Treasury’s daily curve and two independent reports) and the two new sliders’ ranges, all in the ledger and held by the checker. The floor bar’s styling moved into the shared stylesheet, since two pages now draw it. Sharma’s own figures are his claims and feed nothing.
11 Sept 2026, later still — every page gets its own share card, and a wrong figure on one is corrected. A share card is the picture a chat app or social network shows when someone posts a link. Six pages had none of their own and borrowed the front page’s, so a link to the FAQ, the reading list, the tracking page, the ledger, Nvidia or Oracle all previewed as the front page. Each now has its own. Two existing cards were wrong. Meta’s carried −5% as its after-hours share move — Alphabet’s figure, not Meta’s, the same copy that was corrected on the page itself on 27 August and missed here, because the words on a card exist only inside an image and nothing on this site could read them. It now says −9.6%, matching the page. Microsoft’s capex note was too long for the box it sat in and ran over the one beside it, so the live card read “non-raiseshares +8% after hours”; it is shortened, with the same figures. Every card’s words now live as text alongside the site’s code and are re-drawn and compared before each publish, so a card that drifts from what it should say stops the publish rather than shipping. Text too wide to fit does the same instead of overlapping. No figure on this page or any other changed.
11 Sept 2026, later — the scoreboard moves to its own page. Everything that was §7 — the three clocks, the gap chart, the strain chart and the watch list — now lives at Tracking the bet, reached from the nav in the header, and this essay ends at §6 with a pointer. The owner’s reason: the argument and the record are different things, read at different times. Nothing in the scoreboard changed in the move. The one construction that could have drifted is the gap chart’s required ramp, which on this page followed the §2 sliders; on the tracking page it is drawn at the sliders’ defaults ($3.5T, five years, 10%, $250B of running costs, a floor of $1.3T a year), and the constants checker now holds those defaults equal to the registry and to this page’s sliders. The print-sync’s detector reads both pages. §7 is retired rather than reused, as §8 was, since the entries below refer to it by number.
11 Sept 2026 — plain language, every page. The prose of this page, the eight company pages, the FAQ and the reading page was rewritten so that a reader without a finance background can follow every sentence. Terms of trade are now explained where they first appear: capital spending, operating cash flow, free cash flow, the labour share, a company’s “guide” (its own guidance), “backlog” (orders signed but not yet delivered) and the like. Labels were renamed to say what they hold: the gap chart’s “Ceiling” view is now “Upper limit”, the clock chart’s “Depreciation wall” is “Chips need replacing”, the scoreboard’s third column is “What they mean for the bet”, and the nav’s “Reading beside the model” is “Further reading”. One factual slip was fixed on the way: the jobs row pointed “above” to the labour-share row, which sits below it. No figure, hedge, source, model or constant changed; the constants checker passed unaltered before and after. The 10 Sept Oracle print, folded in the same day from the trunk, got the same treatment in its row, in the §7 notes and on its page. Earlier entries below are left in their original wording, since they are the record.
10 Sept 2026 — Oracle FY27 Q1: the loudest demand reading yet, and the first equity sale. OCI +121% against +93% a quarter earlier, RPO to $664B, revenue $19.3B +30%, non-GAAP EPS $1.92 past a ~$1.74 street; FY27 revenue guide lifted to at least $90B, the FY27 capex guide held at $90–95B gross. Against it: $28.5B of capital expenditure in one quarter, more than half of all FY26 spending, free cash flow −$5.4B, and a completed $20B at-the-market share sale — the first of the BIS five to fund the build with equity. The watch-list row is rewritten and moved to the top of the table; the 10 June row it replaces survives in the companion page, refreshed for this print. Two charts moved. The gap chart takes Oracle to +$20.2B/yr above its 2023 baseline, and the strain chart takes its capex-to-operating-cash-flow ratio from 174% to 161% — a fall that is an artefact, and is now annotated as one. $11.4B of Oracle’s record $23.1B operating cash flow is customer prepayments for capacity not yet built; strip them, as Oracle itself does in its “net cash outlay” table, and the same quarter reads ≈244%. That prepayment is also the cleanest instance of §5’s circuit on the record so far: the customer financing the builder, counted as the builder’s operating cash. Both charts now carry ragged series by construction — Oracle’s August fiscal quarter lands a month before the other four report calendar Q3 — so the strain chart’s x-axis was extended a quarter, its end labels now follow each company’s own last point rather than a shared right edge, and the gap chart’s aggregate deliberately stays at the last quarter all five have reported. The aggregate actual line therefore did not move. The §1 and SpaceX capex-to-revenue comparisons are dated rather than overwritten: Oracle’s $0.86 per dollar of revenue was the June quarter, and the August quarter reached $1.47. Oracle sits outside geoveu’s capture, so every figure here is taken from the 8-K earnings exhibit filed 10 Sept, with the share move press-reported and stated as a range. Next reading ~mid-Dec.
4 Sept 2026 — August payrolls, and July’s contraction revised away. August payrolls +162,000, unemployment unchanged at 4.1%, average hourly earnings +3.1% year on year. The same release revised July from −23,000 to +21,000 and June from +20,000 to +31,000, so the “first contraction on this row’s clock” that the 7 Aug entry reported is no longer in the record. The jobs row and the watch-list preamble say so rather than silently carrying the new figure, and the two reading-page passages that quoted July’s −23,000 carry the revision inline. The divergence test still points at the same fork, more mildly: earnings beating against payrolls flat, which is the “tool” world. A jobs print changes no gap-chart or strain data, so the row’s +$553B/yr, ≈$2.2T and ≈28M figures stand. Figures confirmed against the BLS release (USDL-26-1435) and geoveu’s capture of the same morning. Next reading 2 Oct.
3 Sept 2026, later still — the labour share charted, and a channel the entry had missed. The BLS entry gains the series it argues from: all 318 quarters from 1947, with the 2013–2026 window enlarged beneath so the recent fall is legible at all. Both panels start at 52% rather than zero, and say so — the whole eighty-year range is thirteen points wide. The entry’s reasoning is also corrected. It had said the fall showed the share moving “for reasons that have nothing to do with an AI substitution shock”, which quietly treats not substitution as not AI. It leaves out the bargaining channel: what sets pay is the employer’s alternative to paying you, and a credible replacement moves that the moment it is credible — no-one need be displaced. The entry now names three channels it cannot separate, and the correction cuts against this page rather than for it. §4 has been treating displacement in the payroll data as the signal that starts the clock; on the bargaining channel the transfer starts with the threat and never reaches payrolls, so “no displacement visible” stops being reassurance. §6 already argues the other leg — a shrinking share drains labour’s leverage — so the two make a loop, not a sequence. No figure moved; what changed is what the page watches for.
3 Sept 2026, after the BLS revision — the labour share, measured. §7 gains two rows carrying the one variable this page has always modelled and never observed, and the reading page gains the entry behind them. BLS’s revised Productivity and Costs print put the US nonfarm-business labour share at 52.8%, the lowest in a series that starts in 1947, with real hourly compensation down 3.3% against productivity up 1.4%. Checked against the underlying index, the fall is recent: the share sat still through the two years to the end of 2024 and has dropped about two points since, roughly 1.3 points a year. The second row exists because the first invites a wrong reading: 52.8% is one sector of one country, while §4 and §6 run on a bloc-wide whole-economy figure, so the two cannot be set against each other. AMECO’s adjusted wage share is the measure that can, and it says two things worth having. The 60% intercept is high — the euro area reads ~57%, the EU ~56%, the US ~54% — and the blocs are moving in opposite directions, the euro area up since 2022 while the US falls, which makes this morning’s record low a US divergence rather than the developed world’s. Both rows are evidence beside the model in the standing sense: no slope, no gap-chart slice, no contribution to any total. The print is filed as evidence against the §6 dial’s boundary condition rather than for the bet: the dial is geared to the BIS’s premise that labour’s share does not move until substitution actually happens, and by §4’s own reading substitution has not started, so a share falling anyway says the premise is not safe, not that the bet is paying. Nothing in the release attributes any of the fall to AI, and no page constant changed. The intercept question the rows raise is left open on purpose — changing a load-bearing constant is the owner’s call, and the ledger now carries the row that states it.
3 Sept 2026, late — the floor, in units of a business that exists. §2 now states its floor as a multiple of Microsoft’s productivity franchise: Microsoft 365, Office, LinkedIn and Dynamics together turned over $135.3B in the twelve months to March 2026, so the floor at the default sliders is ≈9.6× that. The multiple is computed live and moves with the sliders, from 8.5× at the bottom of the capex range to 11.8× at the top. The comparator was chosen to answer the objection that the capture argument implies these companies do badly — it does not. Microsoft did extraordinarily well and still took four decades to build one franchise of that size; the floor asks for roughly ten, by 2030. The denominator is Microsoft’s own reported segment revenue (nine months FY26 plus Q4 FY25), confirmed against the 29 April press release and an independent write-up of the same print, and it is registered as a load-bearing constant with a checker test, so a stale figure now fails the deploy rather than sitting quietly in the prose.
3 Sept 2026, night — a correction to this morning’s entry. The St. Louis Fed entry called the measured productivity effect “small next to a bill that needs about 40% revenue growth a year”, and the note below said the same. That comparison was wrong twice over. It set a cumulative figure against an annual one without converting, and more importantly it compared quantities that do not share a denominator: an industry-level productivity gain measured in output per hour, against the sales growth five companies need. 2.9 percentage points cumulative over eleven quarters is roughly one point a year, which is not small — against the 2% threshold Waller uses it is substantial, if it is causal, which its authors do not claim. The entry now says what is actually true: the productivity finding is not too small to matter, it simply does not measure capture, which is the thing the bet turns on. No figure changed and no model above was affected; the error was in the comparison, not the number.
3 Sept 2026, evening — a nav, and §8 retired. The essay now ends at §7. The summary table that stood between the scoreboard and the footer was a second index of the reading page, kept in step by hand and read after the argument had already finished, so it is gone; the reading page is reached from the nav in the header instead, alongside the FAQ, the derivations ledger and RAAK. Section numbers are not reused — §8 is retired rather than refilled, and the ledger says so, since the revision entries below refer to it by number. The reading page itself is unchanged apart from stating what its order already was: newest first.
3 Sept 2026, later — Posen on why the displacement is not in the data. The reading page gains the Odd Lots interview with Adam Posen of 1 September. He splits the question the page keeps separate too: more indicative evidence on productivity than on jobs. Hiring still grows in long-haul trucking and junior coding, the two occupations most often named as first to go, which is where §4 already stands; but his endorsement of Luis Garicano’s messy-jobs argument goes further than the page does, holding that the “most exposed occupations” lists any wage-pool estimate leans on are the wrong shape. The interview’s other claim was traced to its source before it was written up: the 0.3% is not Posen’s and not the Peterson Institute’s but Epoch AI’s report of 24 August, and it concerns US GDP growth rather than the level. It moves the §6 growth dial, not the bill — Nvidia’s value-add is already on this page as capex seen from the supplier’s side, and the gap chart runs on the five’s own reported revenue, which no accounting convention changes. Bloomberg’s transcript sits behind bot protection, so Posen is paraphrased from a third-party automatic transcript and nothing is presented as a verbatim quote; the entry says so and cites both substantive claims to their own sources. As always for §8, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.
3 Sept 2026 — the productivity case, read against the capture rate. The reading page gains the St. Louis Fed’s Open Vault interview of 26 August with Alex Bick, on how productivity affects inflation, jobs and pay. It is the standard benign case, and it tests §3 and §4 rather than supporting them: in the Fed’s mechanism a productivity gain shows up as lower prices and higher wages, meaning the gain leaves the seller, while the bet needs a quarter of the displaced wage bill arriving as vendor revenue. The measured effect so far — 2.9 percentage points of extra cumulative labour productivity per 10 percentage points of AI adoption, US Q4 2022 to Q3 2025 — is about one point a year, and the same study found no industry-level employment effect either way, which matches the jobs row’s reading of the July payroll contraction. Both figures were confirmed against the interview and against the underlying research post the interview cites. As always for §8, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.
28 Aug 2026 — the financing of the bill, read beside the model. The reading page gains the FT's Big Read of 26 August on how the data-centre build-out is financed: debt from even the cash-rich, special-purpose vehicles that keep projects off tech balance sheets, synthetic risk transfers moving bank risk elsewhere, and insurance that cannot stretch to a $10bn site. It tests §7's strain chart (capex measured against a company's own operating cash flow cannot see spending financed off the books) and it adds to the patience clock: the same issuer's data-centre debt sold at 5.7% in April and over 7.2% in August. Oracle is where the credit market is already pushing back on one of the five, with a downgrade to triple B minus. The annotation label on §8 and the reading page also changed with this entry, from "where it presses" to "what it tests", because the old label did not say what the reader would get. A plain-language pass followed the same day, over the reading page and this section: em-dashes replaced where a comma, colon or full stop does the job, and two sentences reworded. No figure or hedge changed. As always for §8, every figure is the cited author's claim, feeds no model above, and the ledger records it that way.
27 Aug 2026, late — Oracle, the last of the five. A companion page for Oracle's 10 June print, written from the new blank-sourced template — the fifth BIS hyperscaler was the only one without one. Oracle sits outside geoveu's capture, so every figure was taken from its results release directly. Two things came out of that. The release's own trailing-four-quarter table independently confirms this page's strain series for Oracle, peak and all: capital spending ran to 205% of operating cash flow in the February quarter before easing to 174%. And the watch-list row's "shares −10%" did not survive checking — reports range from −5% to "as much as 10%" after hours, with the following morning at −12.6%, so the row now states the range. Free cash flow for the year was −$23.7B, which the row had described only as capex having "ate" it.
27 Aug 2026, evening — a reading page, and a correction. §8 now carries only each piece's headline and a one-line statement of where it presses; the full entries moved to a page of their own, which is also where new ones will land. A copy-edit pass went through the essay for sentence length — no figure, model or hedge changed, and the constants checker passed unaltered before and after. One real error found and fixed: meta.html reported the 29 July print as "shares −5% after hours" and carried alphabet.html's source list verbatim, both copied from that page when it was written. The move was −9.64% ($585.61 to $529.15) — as the watch-list row above had said all along — and the sources are now Meta's own. The ledger records the correction and how it was verified.
27 Aug 2026 — the supplier gets a page, and the scoreboard turns round. Nvidia was the only reported entity on the watch-list without a companion analysis, so it now has one. The material the row had no room for is the distance between the two financial statements: $59.7B of net income against $24.1B of operating cash flow, receivables up $22.3B in the quarter, and $42.4B spent on equity stakes over the half-year against $4.4B on its own plant. That is §5’s circuit with the supplier’s own capital inside it. The standing rule is now written down: sitting outside the BIS’s five governs the charts and the totals, never whether a print gets analysed, which is why Apple, SpaceX and Nvidia each have a page. Oracle is the one member of the five still missing one. The scoreboard table is also reversed to newest-first, with the two not-yet-reported rows kept at its foot, and the paragraph above it now says where the first wave actually landed rather than anticipating it.
27 Aug 2026 — reading beside the model. The page gained §8, a place for arguments that bear on the bet without feeding it. Two entries open it. The IMF's 25 August growth revision counts AI capital spending on the growth side of the ledger, which lengthens §7's patience clock without moving the revenue floor by a dollar. An FT Free Lunch piece argues that capital keeps flowing to knowledge-work augmentation because that is where the returns are, which narrows the wage pool §3 and §4 measure while leaving the bill unchanged. Everything quoted in the section is the cited author's claim rather than a page constant. The ledger's new §8 records them that way, and none enters the constants registry or any model above.
26 Aug 2026 — Nvidia Q2, and the supplier's view of the bill. The 26 August print beat its own guide and the street on every line — revenue $96.2B against a $91.0B guide, data-centre revenue $89.0B +117% yoy — and guided Q3 to $108.0B. The watch-list row is updated accordingly, and its slope is drawn as none: Nvidia sits outside the BIS's five, so under the standing roster policy it gets no sparkline against the required ramp and feeds no gap-chart slice or capex total. Two corrections travel with the print. The pre-print row described the $91B figure as a data-centre guide; it was Nvidia's guide for total revenue, and the row now says so. And the placeholder had promised to draw a slope after the print, which the roster policy adopted on 30 July does not allow for a non-BIS row — the promise, not the policy, was the error. The strain and gap charts are unchanged: Nvidia is the supplier, and its revenue is the five's spending seen from the other side of the invoice, so adding it anywhere in the model would double-count the same dollars. 27 Aug: the row now also carries the sequential base it was measured from — Q1 FY27 revenue $81.6B and data-centre revenue $75.2B — so the +18% q/q step reads as figures rather than only a percentage.
7 Aug 2026, afternoon — the other ledger reports. July payrolls −23,000, the first contraction in the watch-list's payroll column, with unemployment down to 4.1% — a pairing that points at a shrinking labour force rather than displacement at the bet's scale (the §4 arithmetic still says wage-drawn revenue at the gap chart's pace would mean ≈28M jobs-equivalent, unmissable in this column). The divergence test widens: earnings beating, payrolls now falling. Next reading 4 Sept.
7 Aug 2026 — SpaceX, and the third pool. The 4 August print — the first since June's listing — put the BIS's capex-to-revenue gauge somewhere none of the five have taken it: $2.35 of capital per dollar of revenue, $6.18 inside the AI segment. It also made the report's "circular financing" concrete enough to draw, so §5 now carries the circuit, with a reported figure on every arrow and the loop closing through asset prices rather than cash. §3 gained a third strip: the builders' own $0.84T of 2026 capex budgets, drawn to the same scale as the software and wage pools, because that budget is where a growing share of today's AI revenue actually comes from — the one source that cannot last. The strain chart gained an off-scale marker (≈820%, half-year) and §7 a watch-list row, with a full companion analysis — the ratio, the buyer, and why the $14.1B booked backlog is softer than the contracts' headline value. The model stays the BIS's five; SpaceX is tracked beside them, not inside them.
31 Jul 2026 — the growth reality check. §6 gained a decade of actual advanced-economy growth (IMF data, 2016–2026), split US versus the rest of the bloc against the dial's own reference lines. Outside the post-Covid rebound, neither half has touched the dial's happy zone in ten years; since 2022 the US carries a rest-of-bloc running ≈0.7–1.7%. The FAQ's growth answer now carries the current readings (US Q2: +1.5% annualized; euro area +0.4%).
31 Jul 2026 — the FAQ. A questions-and-answers page grown from real reader questions: advanced-economy definitions (China is not one), the US share of the bloc's growth (with its own sliders — the §6 dial's optimistic zones decompose into ≈4.3% sustained US growth), sources and inspiration, what is BIS-sourced versus page construction, and what would falsify the bet.
31 Jul 2026 — the strain chart. §7 gained the financed system's dashboard: each company's capital spending as a share of its operating cash flow, trailing twelve months. Two of the five now build past 100% — Amazon at ≈107%, Oracle at ≈174% after peaking above 200 — meaning the build-out there runs on the balance sheet, not the business. The revision-history note also became this expandable block.
31 Jul 2026 — audit pass. The gap chart gained a crediting toggle (ceiling vs trend-adjusted — the strict reading shows ≈$149B/yr of above-trend growth against $319B accrued); the white-collar wage pool moved onto a slider ($18–30T, default $22T ≈ 60% of labour income — the launch value $30T was the range's most generous end); and the §3 pools graphic was re-drawn to true scale after the audit found the software pool drawn ≈2.5× too small, overstating the floor's dominance.
31 Jul 2026 — the wave's last two analyses. Full pages for Apple, the bet's control group, and Amazon, where the bill rises with the proof. Amazon's capex-guide raise ($200B→$220B) folded into the gap chart's split, moving the five's 2026 total to ≈$840B.
30 Jul 2026, post-close — Amazon Q2 captured. Revenue $200.6B beat its guide, AWS +37% was the fastest in 18 quarters, and $53.4B of the quarter's net income is the Anthropic stake marking up. The July wave complete; all five of the BIS's hyperscalers on the board; the gap chart's aggregate moved to +$553B/yr against $319B accrued.
30 Jul 2026 — BIS-fidelity pass. Every model on the page audited against the BIS report's own constructions, deviations fixed (the trillion is two years of spend, not one; the capex window is the BIS's 2025–30; Box C's horizon is 2060, not 2065), every derivation documented in a public ledger the deploy now checks mechanically. Oracle's June print joined the scoreboard — the fifth of the BIS's five had been missing. Same day: the gap chart added (§7), and Microsoft and Meta Q2 actuals captured — Meta answered the conviction question this page poses of it, capex floor lifted to $130–145B while quarterly free cash flow fell to $0.8B.
26 Jul 2026. Street consensus and company guides added ahead of the earnings wave.
23 Jul 2026. Alphabet's capex raise — the first live test of this framework.
Launch assumptions stay visible as each earnings print revises them.
Five technology companies, Alphabet, Amazon, Meta, Microsoft and Oracle, will spend more than a trillion dollars in 2025 and 2026 on the chips, buildings and power that AI runs on. Accountants call spending on things that last for years capital spending, or capex, and the word appears throughout this page. The five’s own plans for 2026 alone add up to about $840B. That is more than they earn, so the difference is borrowed. The Bank for International Settlements (the BIS, a bank owned by the world’s central banks) picks these five out in its 2026 annual report and calls them the hyperscalers: the firms that run the world’s largest computing fleets. The report compares this moment with the canal boom, the railway boom and the dot-com boom. Each was a real breakthrough. Each drew in more money than it could ever pay back.
This page does one piece of arithmetic in three steps. How much revenue must the spending earn? Where could that revenue come from? And what does each answer mean for wages, jobs and growth? None of the figures is a forecast. They are rough estimates, meant to be right in their order of magnitude rather than to the dollar, and every assumption sits on a slider so you can change it and see what follows.
The five are the BIS’s choice, not a complete list. On 12 June a sixth builder, SpaceX (Nasdaq: SPCX), listed on the stock market. Its first quarterly results as a public company, on 4 August, showed $18.4B of capital spending against $7.8B of revenue: $2.35 spent for every dollar earned. Oracle, the heaviest spender of the five by this measure, spent $0.86 in the same quarter. By its next quarter, reported on 10 Sept, Oracle had reached $1.47, so the gap has narrowed sharply. The model on this page keeps to the BIS’s five. SpaceX is tracked alongside them for two reasons: it shows how extreme the spending can get, and where its revenue comes from matters more than how much it spends.
Chips wear out fast, so a one-off pile of spending becomes a bill that comes due every year.
A railway line earns for a century. An AI chip earns for a few years, then a better one replaces it. So the spending cannot be treated as a one-off. Spread it over the years the chips will last, add the running costs (mostly electricity), and add the profit investors expect in return for taking the risk. The total is the revenue the sector must earn every year. This page calls it the revenue floor.
For scale: that is 9.6× what Microsoft’s office-software business earns. Microsoft 365, Office, LinkedIn and Dynamics together turned over $135.3B in the twelve months to March 2026, and Microsoft has been building that business since the 1980s. The question is not whether a business that size can be built. It is whether about ten of them can be built by 2030.
Set every slider to its most forgiving end and the floor is still close to a trillion dollars a year. The capex slider matters because the spending keeps rising. On 23 July Alphabet raised its 2026 plan to $195–205B and said 2027 would be “significantly” higher, which pushes the five’s total towards the top of the slider’s range.
The third slider is the one the bond market moves. Will US debt burst the AI bubble? splits it into the ten-year Treasury yield and the premium investors ask above it, and works through Ruchir Sharma’s argument that a yield held above 5% is what ends the boom.
Where this comes from: the slider’s window and range are the BIS’s own (Graph 11.B: total spending 2025–30 of $2–4.5T, with the Nvidia chief executive’s $3–4T projection marked). The slider starts at $3T because about $1.3T is already spent or planned for 2025–26 alone. The floor arithmetic is this page’s construction, not the BIS’s. The BIS asks whether revenue covers the spending and the interest on the debt. This floor asks for more: it spreads the spending over the chips’ life and adds the profit investors expect. Both are ways of asking whether the revenue can justify the spending. One thing the floor cannot show: it is a still picture. A bill is either covered or missed, by exactly the amount revenue falls short. The BIS models the build-out as a system that runs on borrowed money, and there a shortfall feeds on itself. Lenders charge more, new money dries up, spending is cut, and the cut deepens the shortfall that caused it. A small miss becomes a bust. That is the fork §5 describes. The same loop runs upward in good times, when the firms’ stakes in each other inflate their reported profits; the clearest case is the paper gain on Anthropic in Amazon’s results. Full sources and arithmetic: the derivations ledger.
There are only two pools of money that revenue can come from, and one is nearly twenty times the size of the other.
If AI is a tool that helps people work faster, companies will pay for it out of the budget that pays for all their other software. If AI does the work instead of people, companies will pay for it out of what they used to pay in wages. The blocks below show the two pools at their true relative size.
The dashed outline is the revenue floor from §2. In “tool” mode it is larger than the whole software pool: AI would have to replace every piece of business software on earth just to break even.
The wage pool is worked out, not measured. The advanced economies produce about $60T a year, and about 60% of that goes to workers as pay, in line with the BIS’s figures. That is $36T of wages. How much of it is white-collar depends on how you count: about 50–55% if you count jobs, 60–70% if you weight by pay, because office jobs pay more. The slider in §4 runs from $18T to $30T, and the default of $22T is about 60%. The BIS publishes the workers’ share of income but not a white-collar split. The split is this page’s assumption, which is why it sits on a slider.
That is the whole argument in one picture. If the revenue needed is more than the software budget can possibly supply, the rest has to come from wages. The size of the bet tells you what the bettors believe. They are not paying for a better spreadsheet. They are paying for a share of the payroll.
The amber strip is a third source, and it is where much of today’s AI revenue actually comes from. It is the five’s own planned 2026 AI capital spending, about $0.84T, drawn to the same scale: seven-tenths of the whole world software market, in one year, from five buyers. The builders rent computing power to each other, so one builder’s spending is another builder’s revenue. SpaceX’s AI division is the clearest public example. It took in $2.6B in the June quarter, up 247% on a year earlier, and almost none of that came from a software budget or a payroll. Anthropic, an AI lab, rents at least 300 megawatts of computing at SpaceX’s Memphis site for a reported $1.25B a month. Alphabet, one of the five, which is spending about $200B on its own data centres, starts paying SpaceX a reported $920M a month in October. That is about $26B a year of one builder’s revenue paid out of other builders’ capital budgets.
This third source is real and large, and it is the only one of the three that cannot last. Capital spending is a cost, not a market. Revenue paid out of it is the bet paying itself, and it stops the moment the spending stops. So it does not change the arithmetic above. It only postpones it. §5 draws the loop, with the reported figure on every arrow.
Every dollar of revenue drawn from the wage pool means several dollars of wages no longer paid to people.
An AI company cannot charge a customer the full wage of the worker it replaces. The customer keeps most of the saving, or there would be no reason to buy. The share the AI company gets is what this page calls the capture rate. At a capture rate of 25%, every dollar of AI revenue means four dollars of wages no longer paid. Divide the required revenue by the capture rate and you get the wages that must go to fund it.
“Jobs equivalent” is a way to picture the scale, not a forecast of layoffs. The same loss of wages can arrive as slower hiring, lower-grade jobs or smaller raises. The grey bar is the number of US factory jobs usually attributed to competition from Chinese imports between 1999 and 2011, a shock whose political effects are still with us.
The BIS describes two ways this can go wrong, and its own model shows they are the same road.
The BIS’s 2026 annual report describes the two ways the boom can end more plainly than central banks usually do.
Revenue never reaches the floor. Share prices fall. The web of money between chipmakers, AI labs and their investors, in which each one’s spending is another’s income, unwinds. The capital spending that has been propping up economic growth is cut, and the cut is what causes the recession. The report’s parallels: canals, railways, dot-com.
The revenue comes out of wages, on a large scale. In the report’s own scenario (Box C, the “transformative” branch), the share of national income that goes to workers falls from about 60% today to about 20% by 2060, and the text says it heads “towards zero” after that. The economist Adam Tooze, borrowing a phrase from Keynes, calls that a “euthanasia of the working class.” Past machines replaced muscle, the report notes. AI competes with thinking itself.
The catch in the report’s own model: a worker who loses a wage is also a customer who stops buying. In the BIS’s “demand bottleneck” scenario, replacing workers shrinks the wages that buy the output. Firms then find that further automation does not pay, and growth falls below its long-run trend. Workers lose and investors are disappointed at the same time. Replacing workers can destroy the very demand that was supposed to prove the bet right. The two endings are one road.
Simplified paths based on the BIS’s Box C scenarios. The amber line follows the red one for as long as replacing workers pays. Then the shrinking pool of wages stops buying the output, further automation stops paying, and growth sinks below the trend the investment was meant to beat. The end state is both failures at once: workers displaced and investors disappointed.
“Circular financing” is the report’s name for the way a bust would spread. The firms in the boom are each other’s investors, customers and suppliers, so one firm’s trouble quickly becomes everyone’s. Until this summer that was an abstract idea on this page. SpaceX’s first public accounts, and the contracts behind them, now let the loop be drawn with a reported figure on every arrow. What stands out is not any one number but the shape. The cash moves in one direction, and what comes back to the start is not cash.
How this was built: it illustrates the mechanism the BIS describes (Chapter I, pages 22–23, and the notes to Graph 11.B, which define financing between firms and circular financing), using figures the companies have disclosed. It is not a BIS diagram. SpaceX’s June-quarter AI capital spending ($15.8B) and Amazon’s $53.4B gain on Anthropic come from their results releases. The Anthropic and Alphabet contracts are terms reported in the press, not filed accounts. SpaceX’s own release counts only $14.1B of “contracted sales” across all its cloud deals, and defines that as the part the customer cannot cancel. That is why the figure is a fraction of the contracts’ headline value, and the gap is the point: the order book that supports these valuations is softer than the headlines suggest. Sources and arithmetic: the derivations ledger.
Read the loop both ways. Forwards, it is the “tool” world working: a lab rents computing, a builder buys chips, everyone books revenue, and nobody’s payroll is touched. That is exactly what the jobs row on the tracking page keeps finding. Backwards, it is the bust ready to happen. Amazon’s profit at the start of the loop is not cash. It is an accounting gain, the rise in the value of its stake in a lab whose income depends on the middle of the loop. If demand fails at the far end, there is no chain of separate failures to wait for. The same dollar is being counted as income at four points, so all four fall together. The BIS’s warning is not that these firms are lending to each other. It is that the boom’s headline figures count this loop as growth.
Whether workers win or lose comes down to about 1.5 percentage points of economic growth.
Suppose the bet pays off in full. Not just the revenue floor for 2030, but the thing the floor is the entry ticket to: AI doing a large share of the work across the whole economy. In that world the income that goes to owners of capital (profits, rents and interest) roughly doubles over fifteen years, from $24T to about $48T a year across the advanced economies. Owners take that fixed amount first. Workers get whatever is left. Because workers get the remainder, how fast the economy grows decides everything. Drag the dial to see.
One condition keeps the dial honest: it only applies once the bet has paid off. In the BIS’s scenarios where AI does not replace workers, business as usual or a one-off boost to productivity, workers stay essential and their share of income does not move, whatever the growth rate. The dial starts turning only when AI does replace workers and capital’s larger claim is locked in. From then on, growth decides how much of the loss workers feel.
Move the dial and watch which lines move. Capital income never changes, because the bet fixes it. Everything else is what is left over.
At the historical growth rate of 2%, doubling capital’s income is only possible by cutting wages in dollar terms. At 3.2%, wages grow at exactly their historical rate. Above 3.5%, a rate the advanced economies last sustained in the decades after the Second World War, everyone gains even as workers’ share shrinks. Each extra percentage point of growth is worth about 7–8 points of workers’ share.
Where this comes from: the dial is this page’s own arithmetic. Capital’s income is fixed at double today’s $24T, and workers keep whatever growth leaves over. It is a deliberate simplification of the BIS’s Box C model, which works task by task; the BIS’s no-replacement scenarios set the condition described above. Every readout can be recomputed from three inputs: output of $60T, a workers’ share of 60%, and the growth rate. The arithmetic is checked line by line in the derivations ledger. One join to keep in view: the dial runs fifteen years while the capex model in §2 runs to 2030. That is deliberate and follows the report’s own two horizons (Chapter I tests the boom, Box C runs the long transformation). The doubling of capital’s income does not follow from the §2 floor. $1.3T a year of revenue is the down payment on replacing workers, not the transfer itself. The floor asks whether the bet survives. The dial asks what it wins.
The BIS scenario runs longer than the dial above. In its Box C simulation, workers’ share of income falls to about 20% by 2060, thirty-five years from now, and the text says it heads “towards zero” after that. Year by year the change is small, about a point a year, never a sudden drop, and that is what keeps it politically quiet. Added up, it is nothing like the past: about ten times the decline the advanced economies have absorbed since 1980.
It also rests on a stronger assumption than the dial. A one-off doubling of capital’s income needs only for AI to take over a large share of today’s tasks. A share that keeps falling for forty years needs AI to keep taking over new tasks, year after year. In the BIS’s model that happens only in its transformative branch, where AI improves AI. The dial is a bet that AI can replace workers once. This chart is a bet that it can keep doing so.
The historical line is a simplified average for the advanced economies. The distance between the two endings, 56% against 20%, is how much the scenario claims will change.
A share of income is also a share of power. Workers’ bargaining power rests on three things: a strike must hurt, replacing a workforce must be costly, and the pay packet must carry political weight. All three shrink as workers’ share of production shrinks. In an economy where workers get 20%, most of that power is gone. Firms can produce without workers, so a strike is no threat. Households live on government payments rather than wages. Workers bargain with the state for benefits instead of with employers for pay. That is a different kind of society, arrived at one unnoticeable year at a time.
The dial above asks what growth rate the advanced economies reach. Here is what they have actually delivered, split between the United States and the rest, because just under half of the group’s output is American and the two halves have not moved together.
Only the rebound after Covid has ever reached the dial’s good zone, and the dial needs that rate for fifteen years running, starting from a group growing about 1.8% a year with its non-US half at about 1.4%. Because the US is just under half the group, any target for the whole quietly demands much more of the US. At today’s weights, the 3.2% at which wages keep pace means the US growing at about 5% a year for fifteen years. The FAQ works through that arithmetic, with sliders. Source: IMF World Economic Outlook, real GDP growth, July 2026 edition; 2026 is a projection. The rest-of-group line is worked out from the total using a 45% US weight, as set out in the ledger.
Set the dial to 3.5% and everything a worker can feel improves: wages grow 2.5% a year, above the historical 2%, the best sustained pay rise since the postwar decades. As people would live it, this is prosperity.
Over the same fifteen years, workers’ share of income falls eight points (60% to 52%) while capital income doubles, growing nearly 5% a year. Capital’s gains do not stay as income. They pile up as assets, so ownership concentrates even as pay packets grow. Workers are better off in dollars and worse off in position. Nobody notices for as long as growth holds. If growth ever falters, the lost share is already gone, and the bargaining power that might have won it back drained away during the good years.
The argument ends here. Whether it is right is a matter of record, and the record is kept on its own page.
Tracking the bet carries the scoreboard: the three clocks that time the boom, the revenue arriving against the revenue required, which companies can still pay for their own building, and every set of results as it lands, newest first. It is updated after each set of results, and the revision history above says when.