Further reading

What Gartner counts. The release gives totals, not the method behind them. Its spending is about $7T over 2026–30, for every AI provider in the world. Its test is a 12% return on the money invested, which Gartner calls the rate investors require, and it does not expect that return to be reached before the end of 2032. It asks for $8.5T of revenue on about $7T of spending, about 1.2 dollars of revenue for every dollar spent over the period. It also forecasts what customers will spend renting AI computing as a service: $66B in 2027, rising to $174B in 2030, which it says is too slow for most of the firms that rent it out to earn back their chips before the next generation replaces them. “Over 80% of neocloud providers” ceasing “independent operations” includes being bought; some reports turned it into 80% going bankrupt, which the release does not say.

It tests the revenue floor in §2, and agrees with it. The floor is a yearly figure, $1.3T once the spending is complete. Added up year by year along the required ramp on the tracking page, it comes to $3.85T for 2026–30, on the BIS’s $3.5T of spending: 1.1 dollars of revenue for every dollar spent. Gartner’s 1.2 is close. The two methods differ, so the closeness is a check, not a match. Gartner asks for the money back within the window; the floor writes the chips off over five years and charges a 10% return. The spending differs far more. Gartner’s $7T is twice the BIS’s $3.5T, because it counts every provider worldwide, and the five are most of this year’s figure: their plans for 2026, $842.5B, are about 85% of Gartner’s “nearly $1T”. Gartner’s $6T for 2027–30 is $1.5T a year, above the $1.1T analysts expect the five to spend in 2027. It is one more forecast above the range of the front page’s spending slider, which keeps the BIS’s $3T to $4.5T.

It tests the buyers’ bill, and finds it cautious. The bill asks how much of each dollar customers spend on AI reaches the builders. At its starting values, about 71% of what customers pay reaches the builders: the floor of $1.30T out of a bill of $1.82T. Gartner says less than half reaches the firms that own the infrastructure. If so, the bill is more than twice the floor, more than $2.6T a year. On the bill page that is the computing-share slider at its lowest setting, 50%, where the bill comes to $2.62T. Gartner sits at the edge of the page’s range, not inside it, which supports the page’s claim that its starting values understate what customers must pay. The two measures are not identical. Gartner’s share is of all AI revenue in the world; the bill’s is of what customers pay for the five’s computing. They ask the same question, counted two ways.

Set against Gartner’s own forecast, the bet falls short by more than half. At the bill’s starting values, customers would need to pay about $5.4T over the same five years for the five builders alone. Gartner’s $3.6T for the whole industry is about two-thirds of that. Of Gartner’s $3.6T, less than half, under $1.8T, reaches the firms that own the data centres, against the five’s floor of $3.85T over the same years. On Gartner’s forecast the builders receive less than half of what the floor needs, before any builder outside the five is paid.

What it cannot settle. The report behind the release, “Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032”, is not public, so the $3.6T cannot be checked: the release does not say which revenue it counts as traced to data centres, or how it splits between the five and everyone else. And Gartner’s point about the five is about patience, not payback. Their other businesses can carry the spending for years; that is what the tracking page’s patience clock measures, and it is not the same as the spending paying for itself.

Gartner, “Gartner Says Over $3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by the End of 2030”, press release, 14 October 2026, read at source; the quotations are from Will Sommer, a senior director analyst at Gartner. Confirmed by CRN Asia and The Next Web. The $7T, the 1.2 and 1.1, the $3.85T, the 85%, the 71%, the $2.6T and $2.62T, the $5.4T, the two-thirds and the $1.8T are this page’s arithmetic on Gartner’s figures and on registered page values, and the ledger shows each. As always here, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.

How Bain gets to $6T. The figure is not a count of demand. It starts from the spending and works back. Bain puts AI infrastructure spending at $1.5T a year by 2031: new capacity, plus the ongoing upgrade of the installed base of chips, memory and networking equipment. It then assumes that capital spending at about 25% of revenue is sustainable, a ratio it calls ambitious but reasonable on the record of the cloud providers, and divides: $1.5T is a quarter of $6T. The chapter gives no life for the chips, no depreciation, no interest and no debt. Its spending figure for the companies this site models is close to the front page’s. The five hyperscalers, and Bain’s five are the BIS’s five, could spend $780B in 2026, nearly five times the level of three years earlier; the front page carries $842.5B, from the plans the five raised in July and Meta narrowed in October, after Bain’s figure was set. The report’s data-centre chapter adds the global base: $5T to $6.5T of spending to build about 150 gigawatts of new capacity by 2030, not counting the replacement of chips, and hyperscaler capital spending above $1T a year by 2028.

Where Bain says the money comes from. Existing uses: $200B to $400B a year from consumer subscriptions and advertising, and $1T to $1.4T from enterprise productivity gains in software development, sales, marketing, customer service and IT operations, counted as revenue to the providers. New sources: $100B to $200B or more from advertising in chatbots and from AI replacing part of internet search; $400B from autonomous vehicles, trucks, drones and industrial automation, through higher equipment uptime and lower operating costs; $900B from physical AI, meaning simulation, digital twins and robotics in research and manufacturing, on the assumption of a 10% cut in R&D and manufacturing costs; and new products, among them drug discovery, mental-health support, materials science and automated research, with no figure. The named new sources add up to $1.4T to $1.5T, so about $2.7T of the $4.2T gap has no source named. That subtraction is this page’s, not Bain’s. Bain’s own summary of what it takes: funding it sustainably will require adding approximately 1% to the annual global GDP growth rate.

Last year’s edition. In September 2025 Bain put the requirement at $2T of new revenue a year by 2030, for about $500B a year of data-centre investment, and found the world $800B short. The same quarter sits under both editions: $500B is a quarter of $2T, and $1.5T is a quarter of $6T. The spending base tripled in a year, and the requirement tripled with it. The scope widened too. Last year counted new data centres; this year counts all AI infrastructure and the replacement of what is already installed, so the two figures are not one measure taken three times over. The 2026 edition names no shortfall of its own. Its gap is the $4.2T above.

How to read it: the left column is Bain’s $6T, drawn at the top of each range it gives: $1.8T from today’s uses, $1.5T from the new markets it names with a figure, and the $2.7T left over, for which the report names no source. The right column is this site’s buyers’ bill, $1.82T a year at its starting values: the $1.3T that must reach the builders, and the $0.52T the firms in between add for their own costs and profit. The dashed line is the $1.2T the world spends on business software each year, the smaller of the two pools in the front page’s §3. Both accounts put today’s uses at about that line, and that is where they agree. Bain’s figures are its own and feed nothing here; the right column and the line are the site’s registered values.

What it tests: the buyers’ bill, and through it the revenue floor in §2. Bain’s $6T is what customers pay for AI products across the whole industry, so the figure on this site to set beside it is the bill, $1.82T a year at its starting values, not the floor, which is what reaches the builders. Bain asks 3.3 times the bill. This page’s usual test splits such a distance in two: how much spending an estimate counts, and how much revenue it asks of each dollar of spending. Both differ here. Bain counts $1.5T a year for every builder in the world, with replacement; the floor’s $3.5T over six years is $583B a year for five companies, so Bain counts about 2.6 times the spending. And Bain asks four dollars of revenue for each dollar of yearly spending, where the bill asks 3.1 and the floor 2.2 on the window’s average, about 1.3 times as much. Multiplied together the two give the 3.3, and on that average the first is the larger part. The site’s other check, cents of revenue a year for each dollar of capital, cannot be run on Bain: it gives a yearly flow and never a stock of capital. The two rules are of different kinds. The floor is a break-even: the chips wearing out, the electricity and a 10% return, and nothing more. A quarter of revenue is what a healthy cloud business looked like before the boom. Read as a test it is the strictest on this page: on the five’s own 2026 plan of $842B it asks $3.4T of customer revenue, against the bill’s $1.82T. None of the three estimates below found the floor too high. Bain finds even the bill low.

It tests the two pools in §3, and agrees with them. Bain’s enterprise figure, $1T to $1.4T of revenue to the providers from productivity gains, is the size of the $1.2T the world spends on business software, the smaller pool in §3. That is §3’s own finding restated: AI would have to replace all of business software to pay for itself out of software budgets. Bain does not ask whose budget the money leaves. Its new sources are §4’s capture arithmetic run on pools the front page does not draw: a share of the cost savings in manufacturing and research, in transport, and in advertising. That is the one thing this entry adds to the site. §3 says the only other source is growth that has not happened; Bain says the physical economy. Whether its figures hold cannot be checked from the chapter, which states the 10% saving but not what it is 10% of, nor the share the AI seller keeps. The category is real, and the front page’s drawing has no place for it.

It tests the growth dial in §6, and lands close. Bain’s closing line is about one extra point of growth for the world economy. The dial has the floor paid for by growth of 3.4% a year to 2030 against 2.1% today, 1.3 points, for the advanced economies; the Growth Saviour puts the bill at 4.0%, 1.9 points. Different countries counted, by different routes, and the same order of answer. What Bain does not give is the record. §6 does: outside the rebound after Covid, the advanced economies have not grown at either pace in any year since 2016.

What each side lacks. Bain has no chip life, no depreciation, no financing and no measure of strain, no payroll test and no way to be proved wrong. Its gap is mostly unnamed. And its answer is a function of whatever spending forecast is current, as the tripling since last year shows, which is the front page’s own point about the capex slider. What the report has and the site does not, beyond the third pool: the data-centre chapter counts at least 75 projects worth $130B blocked or delayed by local opposition in the first quarter of 2026 alone, nearly as much as in all of 2025. That is a constraint none of the three clocks on the tracking page carries, and it works the other way from the rest of this entry: capacity that is never built never has to be paid for.

Bain & Company, “New Innovation Is Required to Fund AI’s $6 Trillion Buildout”, Global Technology Report 2026, published 29 September 2026, read at source, with the press release of the same day, which carries the $6T, the $1.2T to $1.8T, the $4.2T and the line on global growth, and “AI Data Center Boom: Can We Build It If They Come?” from the same report, for the $5T to $6.5T, the 150 gigawatts, the $1T by 2028 and the 75 blocked projects. Press coverage: Bloomberg, “AI Faces $6 Trillion Test to Justify Data Centers, Bain Says”, 29 September 2026. The 2025 edition’s figures: Bain’s press release of 23 September 2025, read at source, and Bloomberg’s report of the same day. The $2.7T, the $583B, the multiples 3.3, 2.6, 1.3, 3.1 and 2.2, and the $3.4T are this page’s arithmetic on Bain’s figures and on registered page values, and the ledger shows each. As always here, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.

Why the five pay more. When more bonds are on offer than lenders want, the borrower has to pay a higher rate to sell them. PIMCO’s credit strategist tells the FT that the five’s bonds have been marked down for that reason, in some cases nearly to the level of junk-rated companies, because the amount still to come cannot be predicted. In this summer’s results Alphabet, Amazon, Microsoft and Meta raised their spending plans for 2026 to $745B between them, and he says the size and pace of such revisions leave lenders unable to judge how much debt is still to come. The higher cost has not slowed the borrowing so far. Some of the five have promised lenders that they will not sell bonds in two quarters running.

How much the bond market can hold. KKR, an investment firm that manages $796B, published its own count on 30 September. About $600B of AI debt is outstanding, which is 6.3% of the US market for high-grade company bonds. Over the past 29 years the largest share held by any one industry averaged 2.6%. A bond fund usually caps what it lends to a single company at 3% of the fund. KKR works out that if the largest builders, the five and SpaceX among them, each reached that cap, they could raise $1.7T between them. It puts spending on AI infrastructure by 2030 at nearly $8T, so more than $6T would have to come from somewhere else. It adds that the true exposure is larger than the bonds show, because leases, guarantees and other promises sit off the companies’ accounts. The FT notes that loans tied to Oracle’s data centres have come under strain from building delays.

What the Bank of England adds. The Bank’s Financial Policy Committee, which watches for risks to the financial system, published the record of its September meeting on 30 September. It cites Morgan Stanley’s estimate that about $450B of AI debt had been sold worldwide by early September, more than double the whole of 2025, and JP Morgan’s estimate that about $4.1T of AI spending will be paid for with debt between 2026 and 2030. The big AI builders sold 47% of all the company bonds issued in sterling this year. The committee gives two warnings. The borrowing is hard to see into and partly circular, with AI companies financing one another, so losses could be larger if AI disappoints. And governments’ growth forecasts and budgets now count on AI raising productivity, so a disappointment would hurt government bonds as well as AI shares.

The Treasury argument, from the other side. Ruchir Sharma, further down this page, says the US government’s borrowing makes the builders’ money dearer. The chair of the Federal Reserve now says the builders’ borrowing also makes the government’s money dearer. Asked on 16 September why long-term rates had risen, Kevin Warsh gave three reasons, and the second was the builders: “the competition for capital is real”, he said, and it partly explains the rise. The FT reports that the Treasury secretary agrees, and one bond investor says some governments now time their own bond sales to avoid the builders’. Others doubt it. HSBC’s head of debt markets tells the FT that governments borrow so much more than the five that the five cannot be what moves government rates. PIMCO looked for investors switching from Treasuries into the five’s bonds and found little sign of it. On the FT’s Unhedged podcast of 1 October, Robert Armstrong made the same case: the biggest buyers of Treasuries, such as central banks and insurers, do not choose between the two.

The one forecast underneath. Torsten Sløk, Apollo’s chief economist, wrote on 21 September that lenders’ confidence in the five rests on a single forecast that analysts share: the cash the five’s businesses bring in each year (their operating cash flow) rises from $600B in 2025 to $2T in 2030. If it does not, he expects lenders to charge more, the companies to cut their spending, and US growth to slow.

What it tests: the strain chart on the tracking page, which already measures the number lenders are counting on. That chart sets each company’s capital spending against its operating cash flow. Its own figures put the five’s operating cash flow at $603B for 2025, which is Sløk’s starting point. To reach $2T in 2030 it must grow 27% a year for five years. In the latest twelve months each company has reported, it was $707B, up 37% on the twelve months before. So for now the pace is there. What the chart adds is the other line. Over the same twelve months the five’s capital spending rose 84%, from $319B to $586B. It took 83 cents of every dollar of that cash, against 62 cents a year earlier. The analysts’ forecasts in Sløk’s chart have spending pass cash flow in 2026 and stay above it through 2027. Above the chart’s 100% line the five as a group spend more than they bring in, and the rest is borrowed.

It tests the revenue floor in §2, and agrees with it. $2T is $1.4T a year more cash than the five made in 2025. The floor asks the same five for $1.3T a year of new revenue by the same date. The two are not the same measure. His is cash after running costs, from all their businesses, and part of the rise would come from the older businesses growing as before. They are the same order, though. Lenders are counting on something close to the floor being met. Against that, the five’s bond sales this year, about $200B, are nearly three times the new cloud revenue Goldman Sachs counts in the entry below, about $70B a year.

It tests the spending slider in §2, and finds it low. Spending is what drives the borrowing, and the forecasts in this reporting are above the slider’s range. The slider covers the six years 2025 to 2030. It runs from $3T to $4.5T, which is the BIS’s range, and starts at $3.5T. About $1.3T is already spent or planned for 2025 and 2026, and analysts expect $1.1T in 2027 (the entry below). If spending only held at that level, the five would reach $3.5T by the end of 2028 and pass $4.5T in 2029. The chart below draws that path. Sløk’s chart has analysts expecting spending to rise each year to 2030, not hold. A higher total raises the floor: each extra $1T adds about $300B a year at the front page’s settings. The front page keeps the BIS’s range, and says under the slider that newer forecasts run above it.

How to read it: each bar is the five’s total spending from 2025 to the end of that year, if they spend the $1.1T analysts expect in 2027 and the same again in each year after. The band is the range the front page’s slider covers, and the dashed line is where it starts. On this path the total is $5.7T by the end of 2030. With the front page’s other settings unchanged, that would make the floor $1.96T a year. The floor is $1.30T at the slider’s start and $1.60T at its top. This is the cautious path: it holds spending level after 2027, where analysts expect it to keep rising.

What the higher path would ask of workers, or of growth. The front page turns the floor into a sum for the whole economy. A firm that buys AI keeps most of what it saves, and §4 assumes the sellers capture 25% of it. So for the builders to receive the floor, their customers must gain four times the floor. That gain is either wages no longer paid or output that would not otherwise exist. At the slider’s start, customers must gain $5.2T a year by 2030. On the higher path, with a floor of $1.96T, they must gain $7.8T, which is half as much again. If all of it came out of wages, that is 98 million jobs’ worth of pay, not 65 million, and workers’ share of income would fall from 60% to 50%, not 53%. If all of it came from growth, the advanced economies would have to grow 4.1% a year to 2030, not 3.4%. They are growing 2.1% a year, and averaged 2.0% over the last ten years. Any mix of the two has to add up to the same $7.8T.

How to read it: three totals for the five’s spending from 2025 to 2030, and what each would ask. The left bars are the pay that would have to go if growth supplied none of the gain, counted in jobs at $80,000 each. The right bars are the growth needed if wages supplied none of it; the grey part is today’s pace. Workers’ share of income is 60% today. Output is that of the 43 advanced economies: $69.6T in 2025, and $77.2T by 2030 at today’s pace. These are the front page’s own sums from §4 and §6, run on a larger floor. The Growth Saviour runs them for the floor as it stands. Nothing here feeds a number on the front page.

It tests both sliders on the debt page. The first slider there is the Treasury yield, and Sharma’s line on it has been crossed: the ten-year yield closed above 5% on every trading day from 23 September to 2 October. The second is the premium the builders pay above that yield. This reporting says it has risen as well, and for a reason the page did not list. It is not that lenders trust the five less. It is the supply of their bonds, which follows their spending plans. That ties the premium to a number the tracking page already carries: the five’s plans for 2026, $842.5B between them. The FT’s $745B for four of the five agrees with this site’s $750B for the same four, which is the $842.5B less Oracle’s share (the ledger has the split). Goldman Sachs, in the entry below, expects the upward revisions to get smaller. The five’s results in late October will show whether they do.

It also questions the direction of the debt page’s argument. Sharma’s trigger sits outside the industry, in Washington. If Warsh is right, part of it sits inside: the builders’ borrowing raises the rate their own borrowing is priced from. The Bank of England’s second warning ties the two borrowers together from the far end. A government that has budgeted for AI’s productivity gains is exposed to the same disappointment as the builders’ lenders.

And it questions the debt clock on the tracking page. That clock says the debt raised in 2025 and 2026 falls due around 2028 to 2030 and must then be borrowed again. The FT describes something else. The five are borrowing for longer than companies have before, for 100 years in the case of one Alphabet bond, and other companies are borrowing for shorter periods to stay out of their way. None of these reports breaks the debt down by repayment date, so the clock is not settled here. But for the five’s own bonds, the pressure in this reporting is not old debt falling due. It is new debt, more than $1T of it, that has to be sold quarter after quarter, into a market that KKR says can lend the largest builders $1.7T at most.

Last, it repeats the warning in the FT’s August piece further down this page. Meta borrows through its own accounts and through separate companies tied to single data centres, and KKR counts leases and guarantees that no bond index shows. The strain chart sees only what is on the five’s own books.

Euan Healy, Emily Herbert and Michelle Chan, “AI hyperscalers are transforming debt”, FT, The Big Read, 28 September 2026, read in full. Herbert discusses it on Unhedged, “Hyperscale my debt!”, 1 October 2026, read in the FT’s transcript. Michelle Chan, “KKR warns of growing credit market risks from AI borrowing spree”, FT, 30 September 2026, read in full. Bank of England, Record of the Financial Policy Committee meeting on 25 September 2026, published 30 September, read at source. Kevin Warsh: the Federal Reserve’s transcript of his press conference of 16 September 2026, read at source. Torsten Sløk, “Hyperscaler Credit Rests on One Consensus Assumption”, Apollo, 21 September 2026, read at source; the years and the five companies are from his chart and its note. PIMCO, “Hyperscalers Are Repricing, Not Displacing (So Far)”, 28 September 2026. Sourcing note: Goldman Sachs’s $500B, $200B and $1T reach this page through the FT alone, in the article and in its reporter’s account on the podcast. Morgan Stanley’s $450B, in the Bank of England’s record, and KKR’s $600B outstanding are separate counts of the same order. KKR’s report was read through the FT’s account of it. A second report, in the Globe and Mail, carries the $600B, the share of about 6% and the fifth of the index that could end up exposed; the $1.7T, the 2.6% and the $8T rest on the FT alone. The ten-year yield is the US Treasury’s daily series, as on the debt page. The cash flow and capital spending totals, their growth rates, the 27% a year, the 83 and 62 cents, the $1.4T, the “nearly three times”, the years in which spending would pass the slider’s values, the chart with the three floors in its note, and the second chart and its table are this page’s arithmetic. They use the strain chart’s data at the last quarter each company had reported by 5 October 2026, and the ledger shows each. As always here, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.

Goldman’s note of 25 September covers five US builders, reported as the same five the front page models. They are on course to spend $800B on capital equipment in 2026. Analysts on average expect $1.1T in 2027, and Goldman expects more. Goldman takes the average of those two years and asks what AI revenue a year would justify it. Its chart gives four answers, depending on the return on invested capital the builders want (profit as a share of the money put in): $308B for none, $417B for 10%, $526B for 20% and $636B for 30%. A second Goldman chart shows why 30% is the test: the five earned about 40% on their capital before the build-out and earn about 25% now. Against those targets Goldman sets two facts. The five’s cloud revenue was running about $70B a year above its pre-AI trend in the second quarter of 2026, and their customers have signed more than $1.5T of orders that are not yet delivered. Wigglesworth does the division: revenue must rise about four times to break even and about ten times to earn 30%. He adds that he trusts the order book less than Goldman does.

That is only what the builders must earn from renting out computing power. Most of their customers are firms that sell AI products, and those firms must charge more than they pay in rent if they are to make a profit themselves. Goldman’s table assumes computing is 75% of such a firm’s running costs. On that basis the people who finally pay, the companies and households that buy AI products, must spend about $0.9T to $1.3T a year if the builders settle for a 10–20% return and the product firms keep 30% of their sales as profit. If the builders want 30% and the product firms 40%, the figure is $1.9T. The table is not built on the chart’s bars. The chart charges the builders’ return on one year’s average spending, and the table charges it on the two years together. So a 10% return means $417B a year for the builders in the chart and about $520B in the table. Wigglesworth sets both beside the roughly $1.5T the world spends on software each year. He calls the first “ambitious but not fantastical” and the second less feasible, because people who are used to getting a product free or cheap resist paying what it costs to make.

Van Nieuwerburgh’s paper, written for the Brookings Papers on Economic Activity, counts something wider: every AI data centre in the US, whoever owns it. A data centre’s size is measured by the electrical power it draws. He puts the cost of a 200-megawatt site at $8.2B and expects 183 gigawatts of them to be finished by 2032. Building them takes $10.3T of investment between 2025 and 2032, or 3.6% of US GDP a year. That is a larger share of the economy than the canal, railway, electrification, highway or telecoms booms took. For the owners to earn 10% a year, with chips that last six years and half of every dollar of revenue going on running costs, the data centres must bring in $3.7T a year by 2032. That is 9.2% of the US GDP he projects for that year. If the chips last three years and not six, the figure is $6.0T.

Bernstein and Cummings, both former White House economists, count six firms: the front page’s five plus SpaceX. By their estimate the six have spent about $1.2T on AI since 2024 and earned $277B from it. To pay for that spending the six need between $2.4T of AI revenue over the next six years, which is enough to break even, and $3.8T, which is enough to match what their existing businesses return. In the coming year alone they need $520B to $850B, three to four and a half times what the authors estimate they made this year. Add the spending planned to 2030, $5.7T over five years, and the requirement becomes $13.1T to $18.7T over ten years. A footnote gives the cost of a slow year. If revenue grows by two-thirds next year, to $313B, the six then need $612B to $1.2T in 2028 to stay on course.

Wigglesworth opens with the case on the other side. In August Dan Davies worked through a deal the front page draws in §5: Anthropic renting SpaceX’s Colossus 1 data centre for a reported $1.25B a month, or $15B a year. Davies allows $6B a year to run it and writes off a building cost of $13B over five years, and he regards both as high estimates. That leaves SpaceX $6.4B a year, a return of about 50% on the money it put in. His own caveat is that a return like that lasts only while the chips are in short supply.

What it tests: the revenue floor in §2. The floor is this site’s own arithmetic, not the BIS’s, so outside sums that reach the same place by another route are the best check it can get. These three do. Their totals look far apart, from $308B to $3.7T a year, but most of the distance is in how much spending each one counts, not in how much revenue each dollar of spending has to earn.

Van Nieuwerburgh is the closest comparison, because he asks for the same 10% return the §2 slider starts at. His method is different. He charges the chips over six years and the buildings over twenty, the way a mortgage is repaid, then doubles the result because half of revenue goes on running costs. It comes to 38.7 cents of revenue a year for every dollar of capital. The §2 floor writes everything off over five years, adds $250B a year of running costs and a 10% profit, and reaches $1.3T on $3.5T of spending: 37 cents for every dollar. His total is nearly three times the front page’s because he counts $9.6T of capital, every US data centre to 2032, against five companies’ $3.5T to 2030.

Goldman’s figure is smaller than the floor for the opposite reason. It is sized on one year’s spending, where the floor is sized on six. Goldman’s note is not public and the FT does not reproduce the capital figure behind the chart, so the match here is rough. If the base is what the chart’s title says, the average of 2026 and 2027, it is about $1T, and $417B a year at a 10% return is a little over 40 cents for every dollar. Bernstein and Cummings’ ten-year figure, spread evenly, is $1.3T to $1.9T a year. That is for six firms and $5.7T of spending, where the floor is $1.3T a year for five firms and $3.5T, so the two are not the same sum. They are the same order of answer: more than a trillion dollars a year. None of the three finds the floor too high.

It also tests the two pools in §3, and finds the comparison there kinder to the bet than it should be. §3 sets the floor against the $1.2T the world spends on business software, as if software buyers paid the builders directly. Mostly they do not. They pay a firm that sells an AI product, that firm rents computing from a builder, and it needs its own margin on top. Goldman’s table is the first outside estimate of that mark-up to appear on this page, and in every cell the final buyers’ bill is well above what the builders need. So where §3 says AI would have to replace all the world’s business software to reach the floor, the fuller statement is that it would have to replace more than all of it. The exception is a builder that sells to the final buyer itself, as Meta does with advertising; Goldman’s table applies no mark-up to Meta. The site now carries that mark-up as the buyers’ bill, on its own page. It uses Goldman’s method but not its dollar figures, which are built on different years of spending, and its profit slider starts at 10%, the bottom of Goldman’s range. The front page keeps to the floor and says in §3 that its drawing is kind to the bet for this reason. The FT’s $1.5T for world software spending is also a wider measure than §3’s $1.2T for business software. Neither figure replaces the other.

On the revenue that has arrived so far, it agrees with the strict view of the gap chart. Both readings are for the second quarter of 2026. Goldman counts cloud revenue above its pre-AI trend: $70B a year against the $308B needed to break even, which is under a quarter. The gap chart on the tracking page, in its trend-adjusted view, counts all of the five’s revenue above their 2022–23 pace: $149B a year against $319B of the bill that had come due by then, which is under half. The two measure different things. Goldman’s leaves out advertising and retail, and each is set against a different requirement. They reach the same verdict: counted strictly, the new revenue is a fraction of what the spending already needs.

What none of this settles is the price. Davies and Van Nieuwerburgh make the same point from two sides. At today’s rents a data centre pays, and pays well. Van Nieuwerburgh works out that his $3.7T needs about $5.50 for each hour a chip is installed, if every chip is rented all the time, and such chips rent for $6 to more than $10 an hour now. So his sum does not need prices to rise. It needs buyers for about 180 gigawatts at the prices a market short of chips pays today, and whether those prices hold once there is several times as much capacity to rent is the question both men leave open. The front page reads the SpaceX lease in §5 as one firm in the boom paying another. Davies reads the same lease as a very profitable let. Both are true now. Bernstein and Cummings add that their requirement rises if revenue from firms the builders part-own is left out, which is the §5 circuit again. And the floor is a still picture, as §2 says of itself. Bernstein and Cummings supply the part that moves: chips are written off in five or six years, so each year revenue falls short is added to the years that remain.

Robin Wigglesworth, “What is the AI capex breakeven rate?”, FT Alphaville, 28 September 2026, read in full. Goldman Sachs, “More AI capex, more volatility”, 25 September 2026, by strategists led by Ryan Hammond; the title and the five companies’ names are as reported by TechTimes. Sourcing note: the Goldman note is not public. Its headline figures ($800B, $1.1T, about $300B, $70B, $1.5T, and about $1T a year of spending on AI products) are confirmed against two reports written the day it came out, by Investing.com and Yahoo Finance. The four-bar chart of required revenue, the chart of past returns and the sixteen-cell table are read from Goldman’s own exhibits as the FT reproduces them, and have no second source. Stijn Van Nieuwerburgh, “Financing the AI Buildout”, conference draft for the Brookings Papers on Economic Activity, Fall 2026, read at source; the required-revenue figures are its Table 6 and the text around it. Jared Bernstein and Ryan Cummings, “AI Companies Are In a Race Against Time”, 25 September 2026, read at source. Dan Davies, “if it floats, flies or megaFLOPs”, Back of Mind, 14 August 2026, read at source. The cents-per-dollar comparisons, the $520B behind the table’s 10% row, the yearly average of the ten-year figure and the two fractions in the gap-chart paragraph are this page’s arithmetic on the authors’ figures, and the ledger shows each. As always here, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.

The “since 2024” part can be checked in the underlying data, and it holds. The BLS publishes the labour share for the nonfarm business sector (private businesses outside farming) as an index. That index barely moved in the two years to the end of 2024: 96.900 in the last quarter of 2022 against 96.940 in the last quarter of 2024, a difference of four hundredths of a point. It has fallen in five of the six quarters since, to 93.446. Converted to the percentage the release publishes, that is roughly 54.8% at the end of 2024, falling to 52.8% now. About two percentage points of output has moved from workers to the owners of capital in eighteen months, after two years in which nothing moved.

Quarterly readings, 318 of them, from the first quarter of 1947 to the second quarter of 2026, taken from the BLS labour-share index for the nonfarm business sector and rescaled so the last point equals the 52.8% the release publishes. Both panels start their vertical axis at 52%, not at zero: the whole 80-year range is thirteen points wide, and an axis starting at zero would flatten it into a straight line. Read the panels for direction and turning points, not for how steep the drop looks. The lower panel enlarges the tinted window in the upper one. Nothing here feeds a number on the front page.

Two other figures in the same release point the same way. Real hourly compensation, which is pay after allowing for consumer prices, fell 3.3% in the quarter and is down 0.1% on a year earlier: pay is standing still while output grows. Profits per unit of output at nonfinancial companies rose in the quarter at a pace that works out to 43.0% a year, the fastest since the second quarter of 2021. Over four quarters they rose 17.8%, the fastest since the last quarter of 2021. Output per hour rose 1.4% in the quarter and 2.2% over the year. The productivity gain is real; it is not reaching pay.

What the release does not show is a productivity boom running away from wages. Output per hour has grown at 2.1% a year since the end of 2019. That is above the 1.5% a year of the previous business cycle, but it is exactly the average since 1947. On the BLS’s own numbers, output per hour is growing at its normal historical pace while the share of that output going to workers is the lowest ever recorded.

What it tests: the starting assumption behind the §6 dial. Until this release, the front page had no measured series for the labour share at all. The dial was arithmetic with nothing to check it against; the two labour-share rows in the watch list on the tracking page now give it that check. The §6 dial rests on a specific premise, taken from the BIS: in the scenarios where AI does not replace workers, labour stays an essential input and its share of income does not move at all, whatever the growth rate. On the dial the share falls only as far as the bet is paid for out of wages. This release is the awkward case for that premise. By the front page’s own reading in §4, AI is not yet replacing workers. Today’s AI revenue comes from tools and from the loop (firms in the boom buying from and investing in each other), and the jobs data show no one being displaced. Yet the labour share is falling anyway, by about 1.3 points a year.

Three things could be causing the fall, and more than one can be true at once. It may have nothing to do with AI. The US labour share moves with profit margins, with how much machinery is used per worker, with energy prices and with the business cycle. It was falling for twenty years before anyone bought a GPU, the kind of chip AI runs on. It could be AI already replacing workers, which is what the bet needs, and which §4 says is not visible, since the jobs data show no displacement. Or AI could be changing the wage bargain without replacing anyone yet. What sets your pay is not whether your employer has replaced you but what your employer’s alternative to paying you is worth, and a credible replacement changes that the moment it becomes credible. On this third route the share falls in anticipation, and nobody has to lose a job for it to work.

That third route is the one the front page had not allowed for, and it is the one that breaks the dial’s starting assumption. The dial assumes the share stays put until AI actually replaces workers, so the front page has been watching the jobs data for displacement as the thing that starts the clock. If the bargaining route is real, the clock starts earlier and runs silently: the transfer from workers to capital begins with the threat, and would never show up in the jobs figures. §6 already argues the reverse direction: a shrinking share weakens workers’ bargaining power. Plainly the cause runs both ways, which makes it a circle rather than a sequence. That is a stronger objection to the dial’s premise than “the share moves for other reasons”. The honest position is that this release cannot tell the three routes apart. It rules none of them in.

One check bears on the two AI routes, and it is geographic. The measure that can be compared across regions is the whole-economy wage share from the European Commission’s AMECO database, carried in the tracking page’s watch-list row next to the US one. On that measure the euro area’s labour share has been rising since 2022 while the US share falls. If AI were replacing thinking work, or shifting the wage bargain, across the developed world, the European half of that world is showing the opposite. That narrows the claim. The record low is real, and the break in the series is real, but on present evidence it is a US development rather than a developed-world one, and the front page’s argument is about the developed economies as a group. It does not close off the two AI routes, and the reason is worth stating: AI is used more widely in the US than in Europe, so a real AI effect arriving in the US first would produce exactly this shape. At this stage both readings predict a gap between the two, which is why it settles neither. What would tell them apart is Europe’s series turning down later without a European capex boom of its own. That is a thing to watch, not a thing this release shows.

So the release is not evidence that the bet is working. A record-low labour share is exactly what the replacement story eventually predicts, which makes it the easiest number on this page to over-claim. Nothing in the release puts any of the fall down to AI, and this page does not either. What the release does is remove the reason for assuming the share holds still while the question is being decided. It also explains why §4’s count in jobs keeps finding nothing: the same loss of wages can arrive as a smaller pay rise rather than as anyone losing a job.

One thing it is not is a correction to a number in the model. The 60% labour share used in §4 and §6 is the BIS’s figure for whole economies across the developed world. The 52.8% here is for the US nonfarm business sector, which leaves out government, households, non-profits and farms, and has always read lower. They measure different things, and neither updates the other. What the US series adds is direction, and the direction is down, faster than before. The comparable whole-economy series is the tracking page’s row beside it, and on that measure no developed region reads 60%. Whether the front page’s 60% should move is recorded as an open question in the ledger, not quietly changed.

US Bureau of Labor Statistics, “Productivity and Costs, Second Quarter 2026, Revised”, USDL 26-1434, released 8:30 a.m. ET, 3 September 2026, read at source. Flagged by Mike Konczal, Bluesky, 3 September 2026, whose post quotes the release verbatim and adds the post-2024 reading. The quarterly path is this page’s own arithmetic on the BLS labour-share index for the nonfarm business sector (series PRS85006173, 2017 = 100), retrieved from the BLS public API and rescaled so that the Q2 2026 index of 93.446 equals the 52.8% the release publishes; the ledger records the conversion. The other figures — real hourly compensation, unit profits, productivity by cycle — are quoted from the release text. Whole-economy comparison from European Commission AMECO, series ALCD0 (adjusted wage share, total economy), spring 2026 vintage. Sourcing note: Konczal’s two posts are a thread — the second is his own reply to the first — and the labour-share figure was confirmed against the BLS release rather than taken from the posts. As always here, every figure is the cited author’s claim, feeds no model above, and the ledger records it that way.

On jobs he is blunt: the displacement is not in the data. The two occupations you would have picked to go first, he says, are long-haul truckers and junior coders, and hiring in both is still growing. He gives two reasons. One is Erik Brynjolfsson’s J-curve: for a period, probably under ten years, firms have the technology but have not yet reorganised themselves around it, and during that period few jobs are lost. The other is Luis Garicano’s messy jobs argument, which Posen calls compelling. Almost every job involves more specific knowledge and more working relationships than its description suggests, so the lists of “most exposed occupations” published by consulting firms and international bodies are probably misleading. His example from history is the handloom weaver, who really was replaced, while the wider pattern of job losses in the Industrial Revolution looked nothing like what a study of “exposure” written beforehand would have predicted. Where he does allow that AI may already be visible is in reduced hiring of younger workers, though he warns against blaming AI for that, since it is tangled up with the reshuffling of jobs after Covid.

On productivity he calls the debate legitimate and open, and puts himself just on the pessimistic side. He does not think AI gave growth any measurable productivity boost until the last year and a half. As for the largest claims made for it: we should be so lucky, but we are not there yet.

The third strand is about measurement. Weisenthal raised it, and the sources below trace it to its origin. US GDP misses most of the value Nvidia adds by designing chips that are made and sold abroad. No goods leave the US, so no export is recorded; no foreign buyer pays separately for the chip design, so no export of intellectual property is recorded either. Posen’s answer has two halves. The accounting question (whether you measure output or income, and what counts as an import or an export) does not change the underlying path. Nor does it change Nvidia’s profits: the money arrives regardless and is split between shareholders, reinvestment and workers. But if 0.3% a year really was missed, and really belongs to the AI sector, he grants that it tells a different story about productivity.

What it tests: the jobs arithmetic in §4, and the growth dial in §6. On jobs, Posen agrees with the front page’s current reading, and makes its future harder to observe. §4 says today’s AI revenue comes from tools and from the loop, not from replacing workers. The watch list read July’s first-reported jobs figure of −23,000 as a shrinking workforce rather than displacement. The BLS revised it to +21,000 on 4 September. That removes the fall without changing the reading, since August’s +162,000 is still no displacement on the scale the bet needs. Posen reaches the same place from the labour side. But the messy-jobs argument goes further than the front page does. It says the lists of exposed occupations that any estimate of the wage pool leans on are the wrong shape, so the pool of wages the bet must reach may not be reachable job by job in the way the arithmetic assumes. That is a caution about the target, not about the bill.

On the measurement point, note what it does and does not touch. It does not add a dollar of revenue to the bet. Nvidia’s contribution is already on the front page, on the bill side: what Nvidia earns is what the five biggest AI builders spend on chips, seen from the other side of the invoice. And the gap chart is built from the five’s own reported revenue, which no national-accounting rule changes. What it touches is §6, where the dial asks what growth in the developed economies the build-out is being measured against. If US growth is understated by roughly the amount claimed, the dial’s recent readings are low by that much. And the understatement is itself a product of the boom, so it grows as the boom grows.

Bloomberg Odd Lots, “Adam Posen Thinks Things Could Get Very ‘Messy’ for the Fed”, 1 September 2026. Sourcing note: Bloomberg’s own transcript is behind its bot protection, so Posen’s remarks here are read from a third-party automatic transcript and are paraphrased rather than quoted; that transcript misspells several of the names it reports, so no wording is presented as verbatim. The two substantive claims are cited to their own sources instead. The messy-jobs argument is Luis Garicano, Jin Li and Yanhui Wu, “Messy Jobs: The Work That AI Cannot Reach” (June 2026). The 0.3% figure is not Posen’s and not the Peterson Institute’s: it comes from Isabel Juniewicz, Daniel Carey, Phil Trammell and Anson Ho, “The Nvidia-sized hole in US GDP statistics”, Epoch AI, 24 August 2026, which puts the understatement at about 0.3 percentage points of US GDP growth over the past year — a gap averaging some $30bn a quarter through 2025 — and projects it toward two percentage points by the end of 2028 if Nvidia keeps growing at its current pace. Epoch says it confirmed the absence with the Bureau of Economic Analysis.

Has any of it shown up yet? A qualified yes. Bick and his co-authors compared how far each industry had adopted AI with how much its productivity growth had picked up against its own pre-2020 trend. They did this for the US from the fourth quarter of 2022 to the third quarter of 2025, and for Europe over 2022 to 2024. Where the share of an industry’s workers using AI rose by 10 percentage points, output per hour grew by an extra 2.9 percentage points in total over the period. “We find this strong evidence for AI having positive effects,” Bick says, while noting that the estimates do not prove cause and effect, and that other spending on automation could explain some of the result. The same exercise run on employment found no clear job gains or losses at industry level either way.

On wages, the claim is that pay follows the gain: “typically, wage increases reflect productivity and inflation”. Bick cites a speech by Fed Governor Christopher Waller from October 2025, in which productivity growth held above 2% “will tend to support rising real incomes and living standards without inflation pressure”. On jobs he accepts that the adjustment has costs and calls the net effect “incredibly hard to forecast”, reaching for farming as the example: about 41% of US workers in 1900, under 2% today.

He also names the short-term force pulling the other way, which is this site’s subject. Building the data centres is itself pushing prices up: the parts are not arriving fast enough, and electricity supply cannot keep up with demand in the short run. Before the productivity gain lowers prices, the spending needed to buy it raises them.

What it tests: the two pools of money in §3, and the capture rate in §4. Read carefully, the Fed’s mechanism is not the bet’s mechanism. In the Fed’s account the productivity gain leaves the seller: it shows up as lower prices for buyers and higher pay for workers. The gain is passed on, not kept. The bet needs the opposite: a quarter of the wages AI displaces arriving as revenue for the AI companies. That share is the capture rate, which §4 turns into a count of jobs. Both can hold at once only if the gain is large enough to split. So this piece does not soften the arithmetic; it says what the arithmetic has to compete with for the same dollars.

The measured effect is worth sizing carefully, because it is easy to get wrong. The 2.9 points is a total over the whole period, not a yearly rate. An industry with 10 points more AI adoption grew output per hour about 2.9% more in total over the eleven quarters to the third quarter of 2025. That is roughly one point a year. Set against the 2% a year that Waller treats as the threshold for rising real incomes, that is not a small number. If AI is the cause, which the authors are careful to say they have not shown, it is a large one.

What it is not is an answer to the question this site asks. Productivity growth and AI companies’ revenue are different things. One is output per hour across an industry; the other is money arriving in five companies’ accounts. The Fed’s own mechanism is that the gain leaves the seller, so the finding can be large and still say nothing about whether the bill gets paid. What limits the bet is not the size of the productivity gain but how much of it the sellers keep, and this study does not measure that.

The employment finding matches the front page’s jobs row from the other side. No displacement is visible at industry level, just as the monthly jobs figures read as slow hiring rather than workers being replaced on the scale the bet needs (July’s first-reported −23,000 was revised to +21,000 on 4 September, and August added 162,000). Neither is evidence that the displacement is not coming. Both are evidence that it has not started.

It also lengthens the patience clock on the tracking page (how long investors and policymakers will wait for the revenue), in the same way the IMF entry does. If a productivity boom is the expected outcome, and the Fed treats productivity growth above 2% as something that holds prices down, then policy stays friendly to the build-out while the revenue question is still open.

St. Louis Fed, “How Does Productivity Affect Inflation, Jobs and Pay?”, Open Vault, 26 August 2026, public. The adoption-and-productivity figure comes from Bick, Blandin, Deming, Fuchs-Schündeln and Jessen, “Mind the Gap: AI Adoption in Europe and the U.S.” (30 March 2026), drawn from a paper prepared for the Brookings Papers on Economic Activity, Spring 2026. The Waller quotation is from his speech of 15 October 2025.

The heart of the piece is where the risk ends up. Banks are moving the risk on their data-centre loans off their own books through deals called synthetic risk transfers, in which another investor agrees to absorb the losses. Separate companies set up for the purpose (special-purpose vehicles) keep projects off the tech companies’ own accounts. Broadcom is borrowing against chips it designs with Google. Morgan Stanley expects pension funds and insurers to buy the deals that are sold privately rather than on the open market. And the price of this debt is rising. A QTS bond backing a data centre leased to Microsoft sold in April at an interest rate of 5.7%, with three times more demand than supply. In August the same company’s new debt cost over 7.2%, the rate a company with a junk credit rating pays. Insurance has not kept up either. A single data centre now routinely costs over $10bn to build, while the largest single natural-disaster exposure at Munich Re, the world’s largest reinsurer (an insurer of insurers), was €8.5bn. Only a fraction of Meta’s $14bn Texas project is insured.

The article’s deepest worry is that the chips and buildings may not hold their value for as long as the debt that pays for them. “It’s like you’re financing a fax machine and then someone invented email,” says Carlos Mendez of Crayhill Capital, which has turned down debt deals backed by chips. Many lenders now want loans backed by chips repaid in full before the lease on the data centre ends. Oracle, one of the five companies in the front page’s model, is where lenders are already pushing back. It has had financing trouble at its Michigan campus, a demand from Wisconsin’s utility regulator for $7bn of collateral (assets pledged as security), and a downgrade by the rating agency S&P to triple B minus, one notch above junk.

What it tests: the strain chart on the tracking page. That chart measures each company’s capital spending (money spent on things that last for years: chips, buildings, power) against its operating cash flow (the cash its normal business brings in). This piece describes financing that moves spending off those books: special-purpose vehicles, leases, risk transfers, loans backed by chips. If more of the build-out is financed that way, the strain chart will understate the real strain: the ratio can look steady while the debt piles up where the chart cannot see it. The early warning would then come from the interest rates lenders charge, not from cash-flow statements. That is what the QTS move is: the cost of the same debt rose from 5.7% to over 7.2% in four months, before any answer on revenue.

It also adds to the patience clock. The IMF entry below says the build-out buys political patience. This piece shows lenders’ patience running out first, through higher borrowing costs. The bill stays the same size; what changes is who is stuck holding it if the revenue comes late.

One detail matters on its own: the QTS debt runs for longer than the first tenant’s lease, so the financing only works if a second tenant turns up. That is the same revenue question this site asks, in different words.

FT · The Big Read, edition of 26 August 2026, subscriber access. The FT’s data centres topic page collects its follow-ups. Figures are credited within the article to Infralogic, DC Byte, Gallup and S&P among others.

Her reading of the AI half is explicit: “What started out as a US phenomenon with AI is now becoming a growth engine for the global economy.” The Fund says the world economy has withstood the closure of the Strait of Hormuz for four reasons: countries drawing down their oil reserves, supply from outside the Gulf, lower demand and a partial return to coal. And AI investment held demand up while energy supply shrank.

What it tests: the patience clock on the tracking page. This site treats AI capital spending as a bill that AI revenue must eventually pay. The Fund treats the same spending as a boost to demand that counts as growth. Both are true, and that is the trap: spending on data centres counts as GDP while it is being spent, whatever it earns later. A build-out big enough to make up for a closed strait is also big enough that stopping it would show up as a global slowdown. That buys the build-out political patience. It does not buy it a dollar of revenue.

Reported by The National; see also the IMF media centre.

He puts this in terms of Daron Acemoglu’s idea of skill-biased technical change: the technology of the late twentieth century made already-skilled workers far more productive and left the rest behind. Anthropic’s own finding points the same way: its coding tools help most the users who already have expertise. Shaw concludes that investors have little reason to automate work where wages are low, and that only government policy changes where the money goes. He cites a 2026 study led by Erik Brynjolfsson which found that a 10% higher legal minimum wage went with 8% higher use of robots in factories, and the records of China, Singapore and South Korea in steering their industries by policy.

What it tests: the two pools of money in §3, and the jobs arithmetic in §4. Both ask whether AI revenue will come from software budgets or from wages. Shaw’s claim cuts across that question. If money keeps flowing to desk work because that is where the returns are, the pool of wages the bet must reach is not the whole white-collar total. It is the part nearest to what already-productive workers do, while low-wage physical work stays out of reach of either pool. The target narrows; the bill does not. That is the uncomfortable reading, not the reassuring one.

It also points at a lever the §5 diagram of the loop does not draw: minimum wages and industrial policy decide which kind of automation gets built, not just how fast.

FT · Free Lunch, newsletter edition of 23 August 2026, subscriber access. This edition went out email-first and carries no separate public URL. The study behind its central figure is public: Brynjolfsson, Li, Miranda, Seamans and Wang, “Minimum Wages and the Rise of the Robots”, Stanford Digital Economy Lab.