A RAAK.work project
Capex WatchThe front page works out what the five AI builders must receive every year for their spending to pay off. Their customers must pay more than that, because most computing reaches them through firms like Anthropic, which rent it, add their own costs and profit, and sell it on. This page works out how much more, and what the larger figure means for jobs.
The builders rent out computing. Most of it is bought by firms that turn it into products, and those firms need paying too.
The front page’s revenue floor is what the five builders, Alphabet, Amazon, Meta, Microsoft and Oracle, must take in every year: the chips wearing out, the running costs, and the profit investors require. It is $1.30T a year at the front page’s starting values. It is not what their customers must pay out, and the two are not the same.
Most of the builders’ computing power is rented by firms that turn it into AI products: the AI labs, such as Anthropic and OpenAI, and the software companies. Anthropic, for example, rents computing from Amazon and Google and sells Claude. Those firms have other costs, such as staff, and they want a profit of their own, so they must charge their customers more than they pay the builders in rent. A smaller share of the computing never passes through such a firm. Meta uses its own to sell advertising, so its customers pay it directly.
The panel below works out what final customers must pay in all, the businesses and people who buy AI products, for the builders to receive the floor. This page calls that the buyers’ bill.
At cautious starting values, customers must pay about $1.4 for every $1 the builders need.
For scale: that is 1.5× the $1.2T the world spends on business software each year, the smaller of the two pools in the front page’s §3. Set computing to 100% of costs and profit to zero and the bill is the floor again. The BIS model is this measure with the firms in the middle taken out.
The floor is held at the front page’s starting values here: $3.5T of spending, chips lasting five years and a 10% return, which give $1.30T a year. Move the sliders on the front page to see the floor change; the bill moves in proportion, because every term in it is a multiple of the floor.
Why the starting values are cautious. The bill gets smaller when computing is a larger share of a firm’s costs, when the firms in the middle keep less profit, and when more is sold straight to customers. Each slider starts towards that end of the evidence, so the page is more likely to understate what customers must pay than to overstate it.
Where this comes from: this is the site’s own arithmetic, like the floor. The method and the computing share are Goldman Sachs’s. In a note of 25 September 2026 its analysts worked out what final customers must spend on AI products for the builders to earn a return. They assumed computing is 75% of a product firm’s running costs and tried profit margins from 10% to 40%. FT Alphaville published the table on 28 September. The reading page sets it out. The third slider starts at 10%, the bottom of Goldman’s range, because the one AI lab whose accounts have been reported, Anthropic, has only just begun to make a profit (see below). Goldman added no mark-up for Meta, which uses its own computing power to sell advertising and has no product firm in between. The first slider follows that: it starts at 17%, Meta’s share of the five’s planned capital spending for 2026. Goldman’s table is built on two years of spending and the floor on six, so its dollar figures are not used here, only its method. The BIS’s model has no such layer. It tests the builders’ revenue against their spending and their debt, and AI labs appear in it as partners in financing, not as firms that add a mark-up. Sources and arithmetic: the derivations ledger.
A check against one AI lab’s accounts. Anthropic’s figures give a second view of the firms in the middle. In 2025 computing was 58% of its costs ($7.33B of $12.65B), below Goldman’s 75%, and it lost $8.06B on $4.59B of revenue. It has made an operating profit in 2026. From Anthropic’s draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing. The size of that profit has not been reported from the prospectus. A company forecast reported by the Wall Street Journal in May put it at about 5% of sales in the second quarter. The two figures pull in opposite directions. The computing share points to a larger bill than the page’s: at 58% it would be $2.29T. The profit points to a slightly smaller one: at 5% it would be $1.74T. The page keeps 10%, because one forecast quarter, which leaves out what the company pays its staff in shares, is too thin a basis for going lower.
What the bill leaves out. This leaves out computing bought from builders outside the five, such as SpaceX, or built by the labs themselves. Counting it would raise the bill. Less than half of the $518B Anthropic has committed to spend on computing goes to the five; the tracking page has the split. From Anthropic’s draft prospectus, as reported by Reuters on 28 Sept 2026. Not an official filing. The site still models only the BIS’s five builders, so this spending is noted here but not added.
The larger figure is the fairer one to set against wages, and it raises the jobs count by about two-fifths.
The front page’s §4 turns required revenue into wages. 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 the capture rate. Divide the required revenue by the capture rate and you get the wages that must go to fund it.
The capture rate is the share of the saving kept by whoever sells to the customer. So it divides what customers pay, not what reaches the builders. The front page starts from the floor, which is kind to the bet. The table below runs the same arithmetic on both: the floor, what the builders must receive, as the front page’s §4 does, and the bill, what final customers must pay. The bill row follows the sliders above, and every row follows the two sliders beside it. “Jobs at $80k” is the wages divided by $80,000 a year, the front page’s figure for one white-collar job.
| Measure, per year | Wages displaced | Share of pool | Jobs at $80k |
|---|---|---|---|
| The floor, $1.30T | $5.2T | 21% | 65M |
| The bill, $1.82T | $7.3T | 29% | 91M |
The jobs column 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. For comparison, the number of US factory jobs usually attributed to competition from Chinese imports between 1999 and 2011 is about 1–2 million. The capture rate and the wage pool start where the front page’s do; the front page explains both. Wages are not the only place the money could come from: The Growth Saviour works out how much growth would pay the bill instead: an average of about 4.0% a year to the end of 2030.
The floor can be checked against the builders’ own accounts. The bill rests on assumptions nobody has yet been able to check.
The floor is the BIS’s question. The BIS tests the five builders’ own revenue against their spending and their debt. The tracking page measures the same thing, quarter by quarter, from the five’s published results. Every reading there is held against the floor, and would stop being comparable if the requirement changed.
The bill rests on thinner evidence. Its method and its computing share come from one bank’s note that is not public, read through one newspaper’s account of it. Its check against a real company rests on one lab’s draft accounts, seen by one news agency. Both may turn out right. Neither is yet the kind of source the floor is built on.
So the two are kept apart, and both are shown. The front page says, where it matters, that its figures are kind to the bet because they stop at what the builders receive. This page says how kind. When Anthropic files its prospectus in public, its figures will replace the draft ones here, and this page will say whether they move the bill.