Turning the boom chapter’s three questions into three live numbers
The boom chapter ended with a claim it could not, on the page, make good: that the unanswerable question “is it a bubble?” hides three answerable ones. Is the spending financed in a way that can absorb disappointment? Are the machines earning enough to justify what they cost? And how much of what is being built would still be useful if demand halved? Those questions can be measured, from public filings and peer-reviewed work, and watched as they move. This appendix sketches how, and points to a live dashboard, The Return on Thinking, where the numbers are kept up to date.
Start with why the headline comparison everyone reaches for is the wrong one. It is tempting to set the hundreds of billions spent on AI infrastructure against the modest sums consumers pay for chatbot subscriptions, gasp at the gap, and call it a bubble. That gasp is cheap. Much of what AI earns is hidden inside other products: better advertising, an assistant bundled into an office subscription, work done faster inside the firms themselves. Count only the subscriptions and you miss it. The honest number is profit rather than revenue: the money left after the electricity, the staff and the running costs are paid. A data centre, like a busy restaurant, can take a great deal of money and still lose it. So the first number worth having is a coverage ratio: what the AI capital stock earns each year, counted in both directions, divided by what it costs to hold. Above one, the machines pay their way. Below one, someone is quietly losing money, however impressive the technology.
The second number asks a stranger question, and it is the one nobody yet answers: is the world getting value even if investors are not? A technology can enrich society and ruin its backers at the same time. Britain’s railway mania bankrupted thousands of shareholders and left Britain with railways; the 1990s telecoms bust burned through fortunes and left behind the fibre that later carried the internet for a song. The reason is that only some value can be charged for. The hours a tool saves you, and the things it lets you do that you could not do before, mostly escape into the world free, and free value cannot repay a loan. This is why “AI is obviously useful” and “AI is a bubble” can both be true at once. Usefulness does not prevent a crash; it decides what is salvaged after one. So the second number estimates the benefit the technology delivers to the people who use it and those around them, valued from measured usage and controlled experiments rather than from anecdote.
The third number is the one financial history says matters most on the morning it matters: whose money is at risk? Two families buy identical houses and prices halve. The one that paid cash grumbles and stays put; the one that borrowed nine-tenths is ruined, and the forced sale drags down the street. The houses are identical and the fall is identical. The damage is not, because of how each purchase was financed. The dot-com crash destroyed trillions in shares and barely dented the economy, because it was mostly investors’ own money; the smaller housing crash nearly broke the world, because it was built on debt. So the third number tracks where the borrowed money sits: how much of the build-out is funded from the tech giants’ own cash flow, painful to lose but dangerous to no one else, and how much from debt, leasing and the circular deals at the industry’s edge, where one firm’s disappointment becomes another’s default. As of mid-2026 the honest reading is that the boom began in the first column and is sliding towards the second.
Put the first two numbers on a pair of axes and every boom in history lands in one of four corners. The machinery pays and the world benefits: mature electricity, cloud computing. Investors lose but the world wins: the railways, the fibre. Investors lose and little of use survives: the machines and power burned mining cryptocurrency. Or the owners profit while the world pays for harms no one is charged for. The point is direction, not position. Every young technology starts in the noble-failure corner, spending far ahead of its earnings; the healthy ones travel towards the corner where earnings catch up, quarter by quarter, while the doomed ones sit still as the borrowed money piles up. The dashboard plots AI’s path against the railways, electrification and the fibre boom that ran before it.
Three honest warnings come with the numbers, and the dashboard prints them on the page. It will never name the date of a crash; nothing can, which is the point of bubbles. Every figure is published as a range, because pretending to a precision the data cannot support would be its own kind of dishonesty. And a coverage ratio stuck below one still has innocent explanations: buyers learning what the tool is worth, competition handing the value to users instead of shareholders, or giants spending to insure the businesses they already have rather than to earn a direct return. The measurements narrow the argument. They do not end it. Anyone who claims a single number settles the bubble question, either way, is selling something.
That is the whole ambition: to move the AI-bubble debate, conducted today almost entirely by anecdote, a little way towards accounting. The result is meant to be the thing economics provides at its best: an honest public ledger that anyone can check, doubt and improve.
The live dashboard is at returns.priceofthinking.com: three dials and a moving map, updated from primary sources, with the method and the data open. The figures are released quarterly and archived with a citable DOI, and the methodology page links every number to its source.