A RAAK.work project
Capex WatchTwo lists. The first holds the terms this site made up, or uses in a narrower sense than usual, such as the revenue floor. The second holds the economic and financial terms the pages rely on. Each entry can be linked to on its own.
In the order the front page meets them. These are this site’s constructions, not the BIS’s, unless the entry says otherwise. The derivations ledger shows the arithmetic behind each.
The wager the five biggest AI builders are making: that the money they spend now on chips, data centres and power will earn enough revenue later to pay for itself. The site’s argument is that today revenue on that scale can come from only two places, business software budgets or wages, and that the software budget is too small. The rest has to come from wages, or from growth the economy has not yet produced. See the two pools and the growth dial.
Alphabet, Amazon, Meta, Microsoft and Oracle. These are the five companies the BIS picks out as the hyperscalers, and the only companies inside the site’s model. Every total, chart line and required figure is about these five.
The revenue the five must take in every year for their AI spending to pay for itself. It adds three things: the chips wearing out (the spending divided by the years the chips last), the running costs, mostly electricity, and the profit investors require for putting up the money. At the front page’s starting values that is $3.5T ÷ 5 years = $0.70T, plus $0.25T of running costs, plus 10% of $3.5T = $0.35T: $1.30T a year.
It is a floor because it is what the builders must receive. Their customers must pay more; see the buyers’ bill. It is also a still picture. On the floor, a shortfall is simply a shortfall. In the BIS’s own model it feeds on itself: lenders charge more, spending is cut, and the shortfall grows. The front page’s §2 has it on sliders.
The companies that rent computing power from the builders, turn it into products and sell those to final customers: AI labs such as Anthropic and OpenAI, and software companies. They have costs of their own, such as staff, and want a profit, so they charge their customers more than they pay the builders.
What final customers, the businesses and households that buy AI products, must pay in all for the builders to receive the floor, once the firms in between have added their other costs and profit. At its page’s starting values it is $1.82T a year, about 1.4 times the floor. The method comes from a Goldman Sachs note. It has its own page; the front page keeps to the floor, because the floor is what the BIS model tests and what the builders’ own accounts can check.
The part of the builders’ computing that reaches the final customer with no firm in between. Meta is the example: it uses its own computing to sell advertising. On the bill’s page it is a slider, set at Meta’s share of the five’s planned spending.
The two places the revenue can come from today. If AI is a tool that helps people work faster, firms pay for it from what they spend on business software (the software pool, about $1.2T a year worldwide). If AI does the work instead of people, firms pay for it from what they used to pay in wages (the wage pool). The pages call the first case “the tool world”. The only other source is growth that has not happened yet: if AI helps people produce more, its customers can pay out of new output, and no wages have to go. The front page’s §3 draws both pools to scale.
The pay earned each year by office workers across the advanced economies. It is worked out, not measured: total output, times the share of it paid as wages, times the share of wages that goes to white-collar jobs. The last step is this site’s assumption, which is why it sits on a slider in §4.
The share of a customer’s saving that the AI seller keeps. If AI replaces a worker, the customer will not pay the whole wage, or there would be no point in buying; it keeps most of the saving. If the seller keeps a quarter, each dollar of its revenue stands for four dollars of wages no longer paid. Dividing required revenue by the capture rate gives the wages that must go to fund it.
Wages displaced divided by the pay of one white-collar job ($80,000 a year). It shows the scale, not a forecast of layoffs: the same loss can arrive as slower hiring, lower-paid jobs or smaller raises.
The circle of money between firms in the boom: a builder invests in an AI lab, the lab rents computing from builders, the builders buy chips. Each firm’s spending is another’s revenue, and what returns to the start is often a higher valuation rather than cash. §5’s diagram draws it with a reported figure on each arrow. The BIS calls the core of it circular financing.
§6’s slider for economic growth between now and 2030, when the spending falls due. What AI must be worth to its customers by then can come out of wages, or out of growth above today’s pace, or any mix of the two. The dial shows the mix. At today’s pace growth pays for nothing and all of it comes out of wages. The faster the economy grows, the less wages have to give. The Growth Saviour sets out the same trade in more detail.
On the tracking page’s gap chart, the dashed line: the floor arriving on schedule, rising year by year to 2030, and split across the five in proportion to their planned spending. The required slope is the yearly revenue growth a company must sustain to stay on it.
The two ways the gap chart counts AI revenue, since no company reports it in figures that can be audited. Upper limit counts all of a company’s revenue growth since 2023 as AI revenue, which overstates it. Trend-adjusted counts only the growth above the pace the company was already managing before the boom, which understates it. The true figure lies between.
A company’s capital spending divided by its operating cash flow, both over the latest twelve months. Below 100%, the business pays for its own building. Above 100%, borrowing, asset sales or new shares pay for the rest. The strain chart draws it for each of the five.
Three deadlines that time the boom, on the tracking page. Market patience: investors can lose faith in the missing revenue at any time. Chips need replacing: the chips bought in 2024–25 wear out around 2027–29, and the replacements must be paid from revenue. Debt comes due: money borrowed in 2025–26 must be borrowed again around 2028–30, at whatever rate lenders then ask. A fourth counter is a rate, not a deadline: the five’s announced spending for the year, shown per second.
The dated table on the tracking page of every set of results and statistics release the site follows, newest first: what was reported, against what was expected, and what it means for the bet.
A company or figure the site tracks as evidence but keeps out of the model’s totals and charts. Nvidia is the supplier the five’s spending flows to. Apple is the control group: it sells into the same economy but has made no bet. SpaceX is a sixth builder, outside the BIS’s five. The AI labs show where revenue may come from. Every such row on the watch list says so in the row.
On the debt page, the part of investors’ required return that pays them for the risk of a data centre over the risk of lending to the US government. The required return is split there into the ten-year Treasury yield plus this premium, each on its own slider. Economists usually call it a risk premium.
The page that records where every number carrying weight comes from. Each entry is one of three kinds: reported (from a filing or an official statistics release), from the BIS (cited to its page or graph), or worked out here (with the sum shown). Where the site departs from the BIS, the entry says so and says why. A script checks the site against the ledger before every publish.
The scale of a number, rather than its exact value: whether something is in the hundreds of billions or the trillions. The site’s models aim to get this right, not the last dollar, and none of them is a forecast.
In alphabetical order. Where a term comes from the BIS report, the entry gives the BIS’s meaning.
The IMF’s group of 43 rich economies, which the BIS also uses: the United States, the euro area, Japan, the United Kingdom, Canada, Korea, Australia and others. China is not one; it is the largest of the emerging market economies. The Growth Saviour lists all 43.
The European Commission’s database of yearly economic figures. The site uses it for the workers’ share of income across whole economies, which counts the self-employed as if they earned an employee’s wage.
A figure for one quarter or one month restated as if it held for a whole year. US growth of 1.5% at an annual rate means the economy grew about 0.37% in the quarter.
Orders a company has signed but not yet delivered, and so not yet counted as revenue. Some of it can be cancelled, which is why a company’s “contracted” figure is often far below the headline value of its deals.
The Bank for International Settlements, a bank owned by the world’s central banks. Its Annual Economic Report 2026 is the reference model for this site: Chapter I on the five’s spending and how it is paid for, and Box C on what AI could do to growth and wages.
A loan sold to investors as a tradable piece of paper. The borrower pays interest and returns the money on a set date. A bond’s yield is the interest rate the market charges its issuer at today’s price.
The section of the BIS report titled “Transformative AI, long-term growth and r-star”. It models four ways AI could play out over the long run: business as usual, where workers remain essential and growth stays at its trend of 2% a year; a bounded productivity boost, which lifts growth by a fixed margin; transformative AI; and the demand bottleneck.
The part of national income that goes to owners rather than workers: profits, rents and interest. With the labour share, it makes up the whole.
Money spent on things that last for years, such as chips, buildings and power lines, rather than on the costs of running the business this year. Accountants spread it over the asset’s life as depreciation.
The BIS’s term for deals in which chip makers and hyperscalers take stakes in AI labs, and the labs in return commit to years of buying chips or computing power. The money invested comes back to the investor as revenue. The BIS warns that the terms are poorly disclosed. The site draws an example as the loop.
An agency’s grade for how likely a borrower is to repay. Investment grade is the safer band; one notch below it is junk, also called high yield, where borrowers pay much higher interest.
What a government spends beyond what it collects in a year, which it must borrow. It is usually stated as a share of GDP.
One of Box C’s scenarios. Every worker AI replaces is also a customer who stops buying, so the market for what AI produces shrinks. Firms find that automating further no longer pays, and growth rises at first, then falls below its old trend. Workers lose and investors are disappointed at once.
The cost of an asset wearing out, spread over the years it is useful. AI chips are useful for three to five years, so their cost returns as a bill every year, which is what makes the floor so large.
A company’s profit for the period divided by the number of its shares. Results are usually judged by whether it beats what analysts expected.
Ownership of a company. A firm can raise money by selling new shares instead of borrowing; it then owes nothing back, but each existing share owns a smaller part of the company.
The twelve months a company reports on, which need not match the calendar. Oracle’s ends in May and Nvidia’s in January, so their quarters are named differently from the others’, and the pages say which months each covers.
Operating cash flow minus capital spending: the cash left after the business has paid for its own building. When it is negative, the difference is borrowed or raised.
Gross domestic product: the value of everything an economy produces in a year. The site also calls it output.
A company’s own forecast of its next quarter or year, such as its expected revenue or how much it plans to spend. The watch list judges results against it.
A firm that runs computing fleets on the largest scale. The BIS uses the word for the five big US builders it studies, and so does this site.
The part of national income paid to workers as wages and benefits, the rest being capital income. The site starts from about 60% for the advanced economies. Box C’s transformative scenario has it falling to about 20% by 2060.
Two ways to add up countries’ output. At market exchange rates, each country’s output is converted to dollars at the going rate. At purchasing power (PPP), it is adjusted for what money buys at home, which raises poorer countries’ weight. The IMF averages growth rates on purchasing-power weights, which is why the site gives the US two shares of the advanced economies: 44% in dollars, 37% on purchasing power.
Data centres are sized by the electrical power they draw, not by how much computing they do. A gigawatt is a thousand megawatts, roughly the output of a large power station.
A company whose debts exceed the cash and securities it holds.
The part of the US economy covered by the government’s productivity and labour-share figures. It leaves out government, non-profits, households and farms, so its labour share is not a whole-economy figure.
The cash a company’s business brings in during a period, before it spends anything on new buildings and equipment.
Operating profit as a share of sales: how many cents of each dollar of revenue are left once the costs of running the business are paid.
A rise in the value of something a company owns, booked as profit although no cash has changed hands. Most of Amazon’s profit in the June quarter of 2026 was the rise in the value of its stake in Anthropic.
The number of people on employers’ pay lists, counted monthly by the US Bureau of Labor Statistics (BLS). The monthly change is the headline jobs figure.
Output per hour worked. It can rise because people produce more, or because fewer hours produce the same output, which is why the site always states hours worked beside it.
The document a US company files with the securities regulator before it first sells shares to the public, setting out its accounts and risks. A company may file a confidential draft first, and Anthropic has so far filed only that. Every figure the site takes from it carries the same label: “From Anthropic’s draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing.” The figures will be replaced when the public version appears.
A figure with inflation taken out, so that a dollar in any year buys what a dollar buys today. Growth rates on the site are real unless they say otherwise.
The profit a year investors expect in return for putting money into a risky venture, as a share of what they put in. It is the third part of the floor, set at 10% on the front page.
The “natural” rate of interest, the rate at which money is neither cheap enough to speed the economy up nor dear enough to slow it down. Box C asks how each AI scenario would move it.
Figures with the regular pattern of the year taken out, such as extra shop staff at Christmas, so that one month can be compared with the next.
The interest rate investors demand to lend the US government money for ten years. The government is the safest borrower in dollars, so this rate sets the floor under what every other borrower pays. The debt page puts it on a slider.
The total of a company’s latest four quarters. It smooths out a single strong or weak quarter, and is how the gap and strain charts are drawn.
One of Box C’s scenarios, in which AI improves itself and so drives growth on its own. Output grows ever faster and the labour share falls towards zero. It is the world in which replacing workers pays off, and the one in which workers lose most.
What employers pay in wages and benefits for each unit of output. It rises when pay grows faster than productivity.