The bottleneck was never the motor.

Ecologists who study why ecosystems suddenly flip into a different state draw a line between two kinds of variable. Brian Walker, Stephen Carpenter and their colleagues, who work on the resilience of rangelands and lakes, call the ones people watch fast variables: the size of this year's harvest, say. Underneath sit slow variables, like the amount of organic matter in the soil, which barely move from one year to the next but decide what a dry summer does to the harvest. A fast variable swings with every shock. A slow one decides whether the shock is absorbed or tips the whole system into a different state.

AI has an abundance of fast variables. Benchmark scores, model launches and the length of task a model can finish all move monthly and all make the news. The slow variables are the ones that decide what those headlines amount to: how many firms use the technology, how much work is rebuilt around it, whether output per hour moves, which part of the industry keeps the profit, and how much of the value ends up with the people using it rather than the people selling it. This site is named after them, and it tracks five.

American factories could buy electric power from the early 1880s. It did not show up in manufacturing productivity until the 1920s. Paul David, an economic historian at Stanford, went looking for the reason in 1990, when computers were in every office and still absent from the productivity figures, and found it in the floor plan.

A steam-era factory ran off one engine. A shaft ran the length of the building, belts dropped from it to every machine, and buildings went up several storeys to keep the shafts short. The first electric factories mostly kept this arrangement, with a motor turning the shaft where the engine had been. The large gain came later, from a small motor on each machine. With no shaft to feed, a factory could be a single storey, lighter to build, and laid out around the flow of materials instead of the flow of power. But that was a different building, and David found that the change waited on the old buildings wearing out and on an investment boom in the 1920s to pay for new ones. The motor was ready decades before the factory was.

Everett Rogers came at the same problem from a farm. He grew up in Carroll, Iowa, where his father refused to plant hybrid seed corn for years, though it plainly out-yielded the old seed, until a drought in the 1930s made the difference impossible to miss. Rogers became a rural sociologist, built on a study of how that corn spread through two Iowa towns, and spent his career on why good ideas travel slowly. His answer was that speed depends on how an innovation looks to the person deciding: whether its advantage is obvious, whether it fits the way they already work, how hard it is to learn, whether it can be tried on a small patch first, and whether they can see it working for someone else.

Put the two together and the shape of the question changes. Asking how fast AI is spreading is like asking how fast electricity spread: the answer depends on which link of the chain you time. So this site splits the chain into its links, the methods, the products built from them, the firms that try the products, and the work that finally gets rearranged, and runs a separate clock on each. METR, an independent research group that tests what AI models can do, measures the length of software task the best model can finish half the time, counted in how long the task takes a skilled person. That length has been doubling every 108 days. The share of American firms using AI stands at 24.6%, and the share of all work hours done with it at 5.3%. Those are not contradictory readings. They are clocks on different machines, and the gap between them is the time it takes to rebuild the factory.

Carlota Perez, a Venezuelan economist who studies how technologies get paid for, went through five technological revolutions: the cotton mills that began in Britain in 1771, steam and railways, steel and electricity, oil and the car, and the computer, which she dates to the microprocessor in 1971. Each unfolded in two halves. In the first, installation, financial capital leads: money that cares about the return rather than about what it builds pours into the new technology, and that indifference is exactly what lets it fund infrastructure whose worth nobody can yet calculate. Canals, railways and fibre-optic cable were all overbuilt this way. The installation half usually ends in a crash, and then comes deployment: the technology becomes ordinary, the institutions catch up, production rather than finance sets the pace, and the gains spread out of the new industry and into everyone else's.

That makes "are we in a bubble" the wrong question and "which half are we in" the right one, and the second has evidence. The clearest is how far spending on the infrastructure runs ahead of the revenue it is meant to earn. Over the past year Microsoft, Alphabet, Amazon, Meta and Oracle spent 4.38× as much on buildings and equipment as the AI revenue we can measure, which is the leading labs' sales plus one survey of what American businesses spend on generative AI, by the venture firm Menlo Ventures. Not all of that spending is for AI, and not all AI revenue is counted, so the ratio is rough; the direction is not. The second piece of evidence is who is paying whom. $998B of commitments were signed over the same four quarters in deals where the seller also funds the buyer: cloud contracts with companies that own a stake in the lab buying the computing, and chip deals in which the chip maker lends to, guarantees or invests in its customer. Demand that the seller is helping to fund is real demand, but it is also demand that stops when the funding does.

The site places itself on Perez's curve by a published rule rather than by feel. Spending at least a dollar on buildings and equipment for each dollar of AI revenue, while such deals are still being signed, is installation. Spending under half a dollar per dollar while fewer such deals are signed is deployment. Anything else is a turning point, and a change counts only when two quarters in a row agree. By that rule this is installation, which is a statement about where we are, not a forecast of how it ends.

David Teece, a business economist at Berkeley, asked in 1986 why the firm that brings an invention to market so often watches customers and imitators do better out of it. His answer had little to do with the invention itself and a great deal to do with two things around it. The first is how easily it can be copied. The second is who owns what the invention needs in order to reach a customer, which he called complementary assets: the manufacturing, the distribution, the service network, the relationships. When copying is easy and those assets are scarce, the profit walks past the inventor and settles with whoever holds them. Being first is an event. Owning the part nobody can route around is a position.

William Nordhaus, a Yale economist later given the Nobel memorial prize in economics for bringing climate change into the study of long-run growth, asked the next question: of all the value an innovation creates, how much do the innovators keep? For the American economy from 1948 to 2001 his estimate was about two cents in the dollar. The rest went to buyers, as lower prices and better products, because profits from being ahead are competed away as others catch up.

The record of earlier waves reads like a worked example of both. The inventors seldom kept much, and when they did it was through a legal monopoly rather than the business itself. James Watt and his partner Matthew Boulton made their money from a patent, which Parliament extended to 1800, and from royalties set at a third of the value of the coal each engine saved over the older design; by one economic historian's reckoning, most of the partnership's profit came from those royalties rather than from building engines. The people who carried an invention into a working business often did better, and then often lost it. William Shockley, who shared a Nobel prize for the transistor, gathered the talent that went on to build the silicon industry and kept none of it: eight of his staff left in 1957 to found Fairchild Semiconductor, and two of them, Robert Noyce and Gordon Moore, left again to found Intel in 1968. Samuel Insull, who came to America as Thomas Edison's private secretary, built one of the country's great electric utility empires and saw it go into receivership, run by a court-appointed manager for its creditors, in 1932.

Those who paid for shared infrastructure often paid for everyone else's gains. Andrew Odlyzko, a mathematician who studies financial manias, estimates that roughly a third of the money put into British railways by 1850 was lost, while the mania of the 1840s tripled the country's track and probably left the country better off; the smaller railway boom of the 1830s, by contrast, paid its investors well. He tells a story much like the 1840s about the fibre-optic networks laid in the late 1990s, far ahead of the traffic meant to fill them. In this site's reading, the durable fortunes went to firms that used a new technology to reorganise something else: Ford's moving assembly line, in a plant built around electric power it generated itself; Standard Oil, which turned railway freight deals into control of American refining; Walmart, whose edge the economist Emek Basker traces to logistics and inventory systems and the scale they allowed.

There are two standing exceptions, and the site watches for both. The first is the seller of equipment to a building boom. In March 2000, as the network build-out neared its peak, Cisco, which sold the routers, was briefly the most valuable company in the world. The obvious modern parallel is the chip makers, whose share of the industry's gross profit the plate below tracks; the parallel is this site's, not the historians'. The second is the chokepoint, a position every user has to pass through, which can hold for decades: the Bell telephone system until the government broke it up in the early 1980s; Microsoft's operating system, which a federal court found held more than nine in ten of the market on Intel-compatible personal computers for a decade; Google's search-ad auction, its own refinement of a pay-per-click auction that a rival, Overture, had pioneered, and which Overture later copied back. If the models themselves become a chokepoint, the historical pattern breaks, which is why the gap between open and closed models is among the readings the site follows most closely. At the level of nations, the political scientist Jeffrey Ding argues that the great powers that rose to economic leadership on general-purpose technologies, the kind, like steam or electricity, that spread into every industry, were those that spread them widely through their economies, not those that invented them first. And the economist and investor William Janeway argues that speculation, bubbles included, has paid for much of the build-out of new technologies, alongside the state.

Taken together, the theory and the record say that a spectacular frontier is entirely compatible with thin and short-lived profits, and that the thing worth tracking is not how clever the models are but which part of the industry holds its margin, and for how long. The site measures that with gross profit, what is left of revenue after paying to deliver the product. It can measure three pieces: the two big chip makers, NVIDIA and AMD, from their filings; Microsoft's cloud business, the one cloud division that reports its costs; and an estimate for OpenAI and Anthropic. A year ago the chip makers took 65.7% of that gross profit. They now take 72.6%. On the other side of the ledger, Americans, added up, say they would need to be paid $172B a year to give up AI chatbots: value users keep rather than value anyone sells. Teece would recognise the first number, and Nordhaus the second.

Five numbers carry the argument, and each is graded by a rule the site publishes and can change only in public, with a written reason. Firms using AI: 24.6%. Work hours done with AI: 5.3%. Output per hour, against a year before: 2.2%. The chip makers' share of the gross profit we can measure: 72.6%. And the value users keep: $172B. Each links to its sources, its history and the rule that grades it. Read together, they describe a technology that nothing yet shows spreading faster than an ordinary one, being paid for well ahead of its revenue, and so far rewarding the companies that sell the scarcest input.

An argument like this one is easy to keep forever, because every slow reading can be called early. The only defence is to name in advance what would change the conclusion and to check it on a schedule. The site's conclusions each come with that list, re-tested every night against the same data, and the list follows this essay.

One of them has already failed. The expectation that profit would drift up from the chip makers toward the model labs, as models became interchangeable, is contradicted by its own series: the labs' share grew, but the chip makers' share grew too, and the labs' gain came out of the cloud business instead. The Who profits page says so in its headline, where anyone can check it against the data. That is what a tracker is for.