The front page sets out what the AI spending must earn back and where the money could come from. This page keeps the score: 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.
This boom has a deadline, because the chips it runs on wear out on a known schedule.
Three clocks are running. First, investors can lose patience with the missing revenue at any time, a risk made sharper by the 2026 inflation shock and the interest-rate rises markets now expect (for the geopolitical background, see Geo-V). Second, the chips bought in 2024–25 must be replaced around 2027–29, and that second round of spending will have to be paid for out of revenue rather than confidence. Third, the debt raised in 2025–26 comes due around 2028–30 and will have to be borrowed again, at whatever rate lenders then demand. The three clocks strike close together. What that rate might be, and why the US government’s own borrowing is what sets it, is the subject of Will US debt burst the AI bubble?
June gave the first clock a reading. Within two weeks SpaceX raised about $85.7B by selling shares on Nasdaq and $25B by selling bonds, the largest sum the build-out has yet drawn from public markets. The bonds pay an average interest rate of 5.855% and the longest run to 2056. Six weeks later its first quarterly results as a public company sent the shares down 8%, below the price at which they had been sold, and the complaint was the spending, not the sales. That is the patience clock at work: the money was there, and the doubt arrived straight after.
The same test, drawn as a race. The dashed ramp below is the revenue floor from the front page’s §2 arriving on schedule: the revenue the build-out must be earning each year by 2030, rising as the spending piles up. It is split across the five companies in proportion to their announced spending. The solid lines are the revenue actually arriving: each company’s reported revenue above what it earned in 2023, before the boom. That is deliberately generous. It counts every new dollar since 2023 as AI revenue, including online shopping and advertising, so the lines are upper limits, not measurements. The gap between the lines and the ramp is the bet.
How this was built: the actual lines are each company’s revenue over the latest twelve months minus its revenue in calendar 2023 (Oracle on its own financial quarters, which end in February, May, August and November). Only audited figures are used, and no share of them is guessed to be AI. 2023 is the year the BIS dates the boom’s spending from. The toggle changes how revenue is credited. Upper limit counts every dollar of growth since 2023 as AI revenue, which by construction overstates it. Trend-adjusted subtracts the growth each company was already managing in 2022–23, carried forward, and keeps only the growth that pace cannot explain. That understates it, because the trend years are themselves inflated by Meta’s recovery and Oracle’s purchase of Cerner. The true AI figure, which nobody can observe, lies between the two. The required ramp scales the front page’s §2 floor by how much of the BIS’s 2025–30 spending window is in place at each date: 2025 spending, then the 2026 plans, then a straight line to 2030. Spending before 2025 is left out, because the window starts in 2025. The five-way split uses announced 2026 spending: Microsoft $190B, Amazon $220B (raised from $200B on the 30 July call), Alphabet $195–205B, Meta $130–145B, Oracle’s plan for its 2027 financial year. The ramp is drawn at the front page’s default settings: $3.5T of spending, chips lasting five years and a 10% return, which give a floor of $1.3T a year. Move the sliders there and the floor changes; this page shows the defaults. Every figure’s source and arithmetic: the derivations ledger.
The gap chart shows whether the revenue is arriving. The chart below shows who can afford to wait for it. Each line is a company’s capital spending as a share of the cash its business brings in (its operating cash flow), over the latest twelve months. Below 100%, the business pays for the build-out itself. Above 100%, something else does: borrowing, selling assets, or other people’s money. This is the borrowed-money system from the front page’s §2 made visible. The further above the line a company builds, the more a revenue shortfall becomes a problem with its lenders rather than a disappointment for its shareholders.
How this was built: capital spending over the latest twelve months divided by operating cash flow over the same twelve months, from the companies’ quarterly cash-flow statements (Oracle on its own financial quarters). Oracle’s last point sits a quarter ahead of the other four, and it falls, from 174% to 161%, for a reason the ratio hides: its August quarter booked $11.4B of customer prepayments inside operating cash flow, which is the bottom half of this fraction. Strip those out and the same twelve months read about 244%. Oracle publishes that adjustment itself, as “net cash outlay for capital expenditures”; the chart plots the unadjusted figures from the accounts, so read this fall as the funding arriving early, not the strain easing. Only cash spending is counted. Equipment taken on long leases is left out, which understates the strain: Meta’s June quarter was 94% on cash spending but about 97% counting leases, and Microsoft’s $41B quarter includes leases against $35.8B in cash. The BIS’s own gauge (Graph 11.A) divides capital spending by revenue. This chart divides by operating cash flow, because the question here is whether the business can pay for the spending, not how heavy the spending is. That is a deliberate departure from the BIS, recorded in the ledger with its sources.
The amber flag is the sixth builder, SpaceX, shown as a marker rather than a line for two reasons. It is not one of the BIS’s five, so it belongs beside the chart and not in it. And it has only two quarters of public accounts, so a twelve-month figure is not possible: $28.5B of capital spending in the first half of 2026 against $3.5B of operating cash flow is about 820%, over six months rather than a year. Read it as the extreme case, not a sixth line. What pays for the difference is in plain sight on its balance sheet: about $85.7B raised from the June share sale and $25B from bonds, with $100B of cash and securities at the end of the quarter. Oracle, at 174%, is spending its reserves and borrowing. SpaceX raised the money first and is now spending it.
The first test is already under way. The results below are the first readings on which road the economy is on. The starting points matter. Growth rates this high slow down every year as the base gets bigger, so to average the 40% a year that the revenue floor requires over several years, the early quarters must grow much faster than 40%. Even an impressive-sounding 60–70% could mean the slowdown has started early. Newest readings first.
As of mid-September the first round of results is complete and the second has opened, and it points one way. Demand mostly cleared the bar. The cloud businesses that sell AI computing grew fast: Alphabet’s by 82%, Microsoft’s Azure by 43%, and Amazon’s AWS by 37%, the only one below the required 40%. Oracle, first into the new quarter because its financial year is out of step with the calendar, came back on 10 Sept at 121%, up from 93%. Nobody cut capital spending, several raised it, and Nvidia’s August guidance says the spending continues well into the autumn. The jobs numbers have not moved with it. July’s payrolls, first reported as a fall, were revised on 4 Sept to a gain of 21,000, and August added 162,000 against an average of 31,000 a month over the past year. That is the split this table exists to catch, in its mild form: the companies’ results beating expectations while jobs and wages stay flat. Oracle’s result also moved the question along: its demand reading is the strongest on the board, and it still had to sell $20B of new shares to pay for the capacity. The rest of the third-quarter results land from late October.
| Date | The figures | What they mean for the bet |
|---|---|---|
| 10 Sept — reported | Oracle FY27 Q1 (the quarter to 31 August 2026): cloud infrastructure revenue (OCI) up 121% to $7.4B, against 93% last quarter and the 40% this page requires. Total cloud revenue $11.6B, up 62%. Revenue $19.3B, up 30%. Adjusted earnings per share $1.92 against analysts’ $1.74. Orders signed but not yet delivered $664B, up $26B in the quarter and $209B on a year earlier, after more than $30B of new AI cloud contracts. Capital spending $28.5B in a single quarter, more than half the $55.7B spent in the whole of the previous financial year, against $23.1B of operating cash flow, leaving free cash flow at −$5.4B. Paid for in part by a completed $20B sale of new shares on the open market. 850 megawatts of data-centre capacity and more than 300,000 AI chips delivered since the May quarter. Guidance for the year to May 2027 raised to revenue of at least $90B and earnings per share of $8.10; the capital spending plan held at $90–95B gross, or about $70B after customers’ prepayments. Shares fell 5.4% during the day, then recovered about 4% in after-hours trading · press-reported share moves, and reports range up to +7%. Next reading around mid-December. Full analysis → Cloud infrastructure +121% vs +40% required | Demand answered more loudly than any result on this board: the growth rate did not slow, it doubled. The funding question grew louder with it. One quarter now costs more than half of last year’s entire build, free cash flow is negative again, and Oracle became the first of the five to sell new shares to pay for capacity, $20B of them. The record operating cash flow is the figure to read carefully: $11.4B of the $23.1B is customers paying in advance for capacity that is not yet built. That is the demand and the funding arriving as the same dollar, the most literal form yet of §5’s loop: the buyer financing the builder. Strip out those prepayments and the strain reading above goes the other way, from 174% to about 244% rather than down to 161%. What the result cannot settle is whether the $664B of orders turns into revenue. It is contracted, not delivered, and the 300,000 chips are the first instalment of what delivering it costs. |
| 4 Sept — reported | US jobs report, August: employers added 162,000 jobs, against an average of 31,000 a month over the previous twelve months. Unemployment 4.1%, unchanged. The same release revised July, which this row had carried as a loss of 23,000 and the first fall on its watch, to a gain of 21,000, and June from 20,000 to 31,000. The fall is gone from the record. Average hourly pay up 3.1% on a year earlier. The information sector lost 23,000 jobs, 8,000 of them in computing infrastructure, data processing and web hosting. Source: BLS Employment Situation; next reading 2 Oct. · This is one country, and the model is not: the BIS frames the bet across all the advanced economies, and the $24T to $48T in the front page’s §6 is a figure for the whole group. The US is still the place to watch first. The spending, the industries most exposed to AI, and the BIS’s own early evidence of change (US industries exposed to AI gaining productivity while adding fewer jobs, its Graph 10.C) are all American, so job losses on the bet’s scale should show up here before anywhere else. An early reading, not the whole test. | This is the other side of the ledger. The split now reads “results beating, jobs flat” rather than “results beating, jobs falling”: the companies cleared their bars while this column added jobs at a third of last year’s pace. The revision is stated rather than quietly overwritten because it changes what this row said last month. The July figure that pointed to a shrinking workforce was a first estimate, and this column is revised with every release. Run the front page’s §4 backwards to see the scale that would count. If the gap chart’s $553B a year of new revenue were coming out of wages, a 25% capture rate would mean about $2.2T of wages no longer paid, or about 28 million jobs’ worth. That could not be missed here. A month of +162,000 with pay up 3.1% is still the “tool” world. Two cautions from 3 Sept still apply. The industry group that includes computing infrastructure and data processing lost jobs, and the release does not say why. And “no job losses visible” is no longer reassurance. Pay can fall behind without anyone being sacked, because a credible replacement weakens every worker’s bargaining position, and the labour-share row below shows exactly that: prices running ahead of pay. That kind of transfer never reaches this column at all. Around 2027–28 the required revenue passes what software budgets can supply. From then on, revenue on schedule and a flat jobs column cannot both continue. One of them gives way, and which one decides the fork. |
| 3 Sept — reported | US labour share, Q2 2026 (revised): workers received 52.8% of what US private businesses produced, as pay and benefits. That is the lowest in a series that starts in 1947; the first estimate on 6 Aug said 52.9%. In the same release: output per hour worked up 1.4%, labour cost per unit of output up 1.2%, hourly pay up 2.6%, and hourly pay after inflation down 3.3%. Source: BLS Productivity and Costs; next reading 5 Nov (Q3 first estimate), revised 8 Dec. · This covers the “nonfarm business sector”, which leaves out government, non-profits and households. It is not a whole-economy figure, and not the measure the AMECO row at the foot of the table carries. | The dial in the front page’s §6 has workers’ share of income falling from 60% to 52% over fifteen years. This is the first row on the page that measures that share rather than modelling it, and it reads 52.8% today. Resist the obvious conclusion. This is not the dial’s ending arriving early. The series covers one sector of one country, while the dial covers whole economies across the advanced world, and the two are built differently. That is why the AMECO row at the foot of the table exists. What this release does show is the pairing, and the pairing is the mechanism §6 describes: output per hour rose 1.4% while pay after inflation fell 3.3%. That gap is a transfer from workers to owners, arriving as prices running ahead of wages rather than as anyone losing a job. What it cannot settle is the cause. Nothing in the release attributes the move to AI, and workers’ share can fall for reasons as old as the series. Read it as a measurement §6 never had before, not as proof that §6 is right. |
| 26 Aug — reported | Nvidia Q2 FY27 (the quarter to July 2026): revenue $96.2B, up 106% on a year earlier and 18% on the previous quarter, above both its own guidance of $91.0B ±2% and analysts’ average of about $91.9B. Data-centre revenue $89.0B, up 117% on a year earlier, from a previous-quarter base of $81.6B total revenue and $75.2B data centre. Earnings per share $2.46 against analysts’ $2.08; gross margin 75.0%. Guidance for the next quarter $108.0B ±2%, still assuming no sales of data-centre chips to China. $26.0B paid out to shareholders in the quarter. Announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to raise more than $500B of outside money for AI infrastructure, subject to final agreements. Full analysis → · Not one of the BIS’s five. Nvidia is the supplier their spending flows to. No slope plotted: Nvidia is outside the model’s five, so it adds nothing to the required ramp. Data-centre revenue +117% on the year, for the record. | The spending side is running hot, not cooling. Guidance of $108B for the next quarter is another 12% step up in three months. Nvidia sells to the five, so its results show their spending about two months before their own results do, and the reading is that nobody is easing off. Note what the beat does not settle. This is the bill being run up, not the revenue that must one day pay it. Every dollar of Nvidia’s revenue is somebody else’s capital spending. The financing partnerships make the same point more sharply: the build-out is reaching past the five’s own cash to outside investors, which widens §5’s loop rather than closing it. Next reading around late November. |
| 4 Aug — reported | SpaceX Q2 2026, its first results since listing on 12 June: revenue $7.81B, up 92%. AI division revenue $2.56B, up 247%, against $15.83B of AI capital spending in the same quarter. Total capital spending $18.37B. First-half capital spending $28.5B against $3.5B of operating cash flow. Computing capacity 1.4 gigawatts, up from 0.4 a year earlier. $14.1B of cloud contracts counted as firm sales. Net loss $541M. Shares down 8%, below the $135 at which they were sold in June. · Not one of the BIS’s five. A sixth builder, tracked as the extreme case on the BIS’s own spending gauge. Full analysis → No slope plotted: SpaceX is outside the model’s five, so it adds nothing to the required ramp. AI revenue +247% on the year, for the record. | The BIS’s spending gauge (Graph 11.A: capital spending divided by revenue) taken further than any of the five: $2.35 of spending per dollar of revenue, against Oracle’s $0.86 in the same quarter (Oracle’s next quarter, reported 10 Sept, reached $1.47), and $6.18 per dollar inside the AI division alone. Apply this page’s floor arithmetic to that one quarter (chips lasting 5 years, 10% required return) and $15.8B of spending adds about $4.7B a year to the permanent bill, for an AI business whose whole revenue, at the current quarterly rate, is about $10.2B a year. Four quarters at that pace and one year’s building costs more than the AI business’s entire revenue today. What is arriving to pay it is other builders’ capital budgets: Anthropic’s rent now, Alphabet’s from October. The clearest public example of §5’s loop. |
| 30 Jul — reported | Apple FY26 Q3 (the quarter to June 2026): revenue $109.4B, up 16%, a record for a June quarter. Capital spending $3.4B in the quarter, about $11B a year. · Not one of the BIS’s five. Apple is the control group: it sells to the same customers but has not made the bet. Full analysis → No AI spending to plot. No floor to clear and no clocks running. | The comparison case: a record quarter with no build-out, and still the worst reception of the round, down 8% on cautious guidance. Investors are pricing confidence in demand, not the spending. |
| 30 Jul — reported | Amazon Q2 2026: revenue $200.6B, up 20%, above its own guidance of $194–199B. Operating profit $27.5B against guidance of $20–24B. AWS, its cloud division, grew 37% on a year earlier, the fastest in 18 quarters, to a yearly rate of $169B, with $496B of orders signed but not yet delivered. 2026 capital spending plan raised to $220B. Net profit $62.6B, of which $53.4B was the rise in the value of its stake in Anthropic. Full analysis → AWS +37% vs +40% required: closing in, still below | Demand held up: the fastest AWS growth in 18 quarters, closing in on the required slope from below. But the quarter’s headline profit is itself part of the loop: a $53B paper gain on a stake in an AI lab. The last result of the round; all five are now on the board. |
| 29 Jul — reported | Meta Q2 2026: revenue $60.8B, up 28%, at the top of its own $58–61B guidance. Earnings per share $6.18 against analysts’ $7.23, as costs rose 55% (including a $2.4B legal charge). 2026 capital spending plan raised to $130–145B. Shares down 9.6% in after-hours trading. Full analysis → No AI revenue line to plot. Commitment gauge instead: capital spending $31.1B out of $31.9B of operating cash flow, leaving $0.8B of free cash. | Meta’s commitment is not in doubt. It raised its spending plan again, and the quarter’s building used up almost all the cash the business generated: operating margin down from 43% to 31%, free cash flow close to zero. Its advertising business makes money from AI without a separate AI revenue line, and beat the top of its guidance, but nothing is left over. After Oracle, Meta is the second of the five to meet the revenue-floor arithmetic inside its own accounts. The loudest tick of the patience clock in the July round. |
| 29 Jul — reported | Microsoft Q2 2026 (the quarter to June): revenue $90.0B against analysts’ $87.6B. Earnings per share $4.81 against $4.24. Azure, its cloud division, grew 43% on a year earlier; Microsoft Cloud as a whole $59.3B, up 27%. Full analysis → AI revenue rate not restated this quarter. Last known +123% vs +40% required. | Beat expectations on every line, but left out the one figure this row exists to track: its AI revenue rate. Azure’s 43% is the closest stand-in and clears the required slope. The AI line stays dark until Microsoft reports it again. |
| 22 Jul — reported | Alphabet Q2 2026: Google Cloud grew 82% (analysts expected 64%), with $514B of orders signed but not yet delivered. 2026 capital spending plan raised to $195–205B. Shares down 5% in after-hours trading. Full analysis → Cloud +82% vs +40% required | Demand cleared the bar easily and spending was not cut, but the market punished the spending. The patience clock ticked. |
| Autumn 2026 | Anthropic listing documents (the S-1 filing) expected before its share sale; a stock-market listing is planned for late 2026. · An AI lab, outside the BIS’s five. Tracked for what its revenue shows, not for its spending. | The first audited revenue record of a company that sells nothing but AI. |
| 2026–27 | OpenAI listing: filed in June 2026; a delay to 2027 is under discussion. · An AI lab, outside the BIS’s five. Tracked for what its revenue shows, not for its spending. | Starting point: revenue running at $20B a year at the end of 2025, while it spends several billion dollars more than it earns each quarter. |
| 2026 vintage | Workers’ share of income, whole economies: the European Commission’s AMECO database measures the share of national income paid to workers across the whole economy, counting the self-employed as if they earned an employee’s wage. For 2024: euro area 56.3%, EU27 55.7%, US 54.5%, Germany 59.1%. The Commission’s forecasts for 2026: 57.0 / 56.2 / 53.6 / 60.3. Source: AMECO, spring 2026 edition. · Annual figures, updated when the Commission publishes a new forecast rather than with each company’s results. 2025–27 are forecasts, not measurements. | This is the series that can be compared with the 60% that the front page’s §4 and §6 both start from, and it does two jobs. First, it shows that 60% is on the high side. On the whole-economy measure no developed economy or group of them reaches it: the euro area is about 57%, the EU about 56%, the US about 54%, and only Germany comes close. The 60% is the BIS’s own setting for its Box C model, which the report says was chosen to match data and established estimates rather than measured directly. Whether this page’s starting point should change is recorded as an open question in the ledger, not quietly altered here. Second, and more telling, it splits the trend in two. The euro area’s share has been rising since 2022 (55.4 → 57.0) while the US’s falls (54.9 → 53.6). So the record low in the US row above is a US story, not a developed-world one. That matters, because if AI were replacing workers across the advanced economies, the European half should be falling too. It is doing the opposite. |
The signal is in the gaps between the columns, not in any single result. If the companies keep beating expectations while the jobs column stays flat, the “tool” world is quietly winning, and the arithmetic above says that world runs out of software budget around 2028. If both columns move together, AI has started replacing workers. Either way, by about 2028 the chips need replacing and the answer is forced: the sector shows it can take revenue from wages, or it cannot.