Intelligence Starts with Electrons

If you have sat through an AI strategy presentation this year, you have probably met the cake. Energy sits at the bottom, then chips, then the data centres Nvidia calls AI factories, then models, with applications on top. Jensen Huang presented it at Davos in January and set it out on Nvidia's blog in March. His foundation is physics: every token, he writes, is "the result of electrons moving". huang-five-layer-cake

He is right, and the cake is complete. My argument is with how it gets read. It is a list of what has to be built, and it is being treated as a timetable for when the value arrives. The last time electrons reached industry, the value came decades after the power, and only once people had rearranged the work around it.

A Supplier's Map

Huang is open about where he stands. At Davos he told Larry Fink that the top layer is "where economic benefit will happen". huang-davos-fink Below it, he puts the bill for the buildout in the trillions of dollars. Nvidia sells the chips on the second layer and much of the machinery on the third. Fink closed the session by asking whether the world was investing enough. He also chairs the AI Infrastructure Partnership, a BlackRock-led investor in data centres, in which Nvidia is a partner and technical adviser. aip-aligned blackrock-aip

None of this makes the cake wrong. Huang's own post says AI will change how work is organised, and concedes that much of the workforce has not yet been trained. But a supplier draws the stack in the order things are bought. Capital goes in from the bottom, and the headlines follow it: power plants, chips, data-centre deals. Rearranging the work makes no headline, so readers of the cake tend to skip it.

The Bottom Layer Is Real

Start with what the cake gets right. The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of the world's total. iea-energy-ai Their demand rose by a further 17% in 2025, and the agency expects it to double by 2030. The bottlenecks it names are physical: slow grid connections, and short supplies of gas turbines and transformers. iea-data-centres-2025

Nobody can arrange a factory around a motor they have not got. On sequence, Huang has the order right.

Forty-Five Years from Shafts to Wires

Until late in the nineteenth century, most American factories ran on a single steam engine or water wheel. It turned iron line shafts along the ceiling of every floor, and each machine was belted to a shaft overhead. Electric motors first drove factory machinery around 1883. They supplied less than 5% of the power for American machinery in 1899, just over half in 1919, and about 78% by 1929. Output per worker-hour in manufacturing grew by 1.3% a year before 1919 and by 3.1% a year after. devine-shafts-wires

The early adopters did the obvious thing. They took out the steam engine, put an electric motor in its place, and kept every shaft, belt and pulley. In some textile mills one motor drove more than a thousand looms. The gains came when each machine got its own motor. Machines could then stand in the order the work flowed, cranes could run overhead, and daylight could reach the floor. Paul David used this history in 1990 to explain why computers were not yet visible in the productivity figures. david-dynamo

The Gains Were in the Floor Plan

The electricity was too cheap to be the point. Fuel or bought-in power usually came to between half a per cent and 3% of the cost of making something. A superintendent at the Baldwin Locomotive Works said he would have electrified even if it had taken several times the power. In 1904 the engineer F. B. Duncan called for "a complete re-design of present machine tools with motor operation alone in view".

In 1905 the Detroit Edison Company began lending motors to manufacturers and doing the engineering free. The gains from the new layout were not automatic. Its sales department, run by Sarah Sheridan, sold a complete drive service with higher output as part of the package. The power company had worked out that selling electrons meant selling the rearrangement too. Factories called that work engineering and management. In my trade it is called design. On 13 October I argued that this is Understanding Is the Job: knowing why the work should be arranged as it is.

What the History Does Not Say

Paul David and Gavin Wright, who know this history best, warn that a purely technological explanation of the 1920s surge is inadequate. When immigration was closed by law after the First World War, factory wages rose, and managers had new reasons to use labour carefully. New management practices arrived alongside the new motors. Measured across labour and capital together, manufacturing productivity grew by more than 5% a year between 1919 and 1929. david-wright-gpt

So the history does not show that redesign alone did it. It shows that the power alone did not. Erik Brynjolfsson, Daniel Rock and Chad Syverson find the same shape in newer technologies. The investments that make a general-purpose technology pay are mostly intangible, such as new processes and skills, and national accounts barely see them. brynjolfsson-j-curve The same split showed up in watchmaking. The Swiss won the accuracy trials and still lost the jobs, as I set out in They Won the Accuracy Trials.

Who Does the Rearranging

This time the rearranging is already under way, and it is not shared evenly. Anders Humlum and Emilie Vestergaard linked large Danish surveys of chatbot use to official earnings records. Two years after ChatGPT arrived, they could rule out any effect on earnings or hours larger than 2%. Employers were absorbing the tools by reorganising tasks instead. About 8% of users had taken on new tasks, and about 17% where employers backed the tools; the commonest was fitting AI into the workflow. About one in ten teachers who had never used a chatbot reported new work from them, mostly policing students' use. humlum-nber

In the United States the rearranging has a sharper edge. Brynjolfsson, Bharat Chandar and Ruyu Chen track employment through payroll records. By mid-2026, employment of 22- to 25-year-olds in AI-exposed jobs was 19% below trend, measured against less exposed peers. The gap came mostly through fewer hires, in jobs where AI substitutes for the work. The authors call these early, descriptive findings, not proof of cause. brynjolfsson-canaries My reading is that the first redesign many firms have made is the junior job they did not advertise. I described that stall in March in The Unemployment Rate Is Fine. That Is the Trap.

The New Grip

A tennis coach watches you hit for ten minutes and then moves your hand. Your knuckle slides a little way round the handle, and the racquet face meets the ball at an angle your wrist does not recognise.

For a month you play worse than you did before the lesson. Shots you used to put away go into the net. Between points your hand drifts back to the old grip, because it still works. It will never work any better than it does now.

This Time the Motor Arrives in a Browser

No shaft has to come down this time. The motor arrives in a browser tab, and the worker installs it herself. By Nvidia's account of the Davos session, AI tools reached nearly a billion people in two or three years. In the United States, nearly 40% of people aged 18 to 64 were using generative AI by late 2024. Adoption at work kept pace with the personal computer. bick-rapid-adoption In a study of 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14%, and by 34% for novices. brynjolfsson-gen-ai-work Agents may soon redesign the workflow themselves. On 9 October I wrote, in He Stopped Reading the Plans, that the agent who built the page still draws the route. On this view, the decades between 1883 and the 1920s shrink to a few quarters.

Much of that is right, and the speed is real. Humlum and Vestergaard's working paper, revised for the National Bureau of Economic Research in March 2026, is the test of it. Their conclusion: "Technological change reshapes work well before it surfaces in earnings or hours". The support agents worked inside one company that had built the assistant into their job; across Denmark, use spread fast and pay did not move. What spreads fast is the motor. If agents shorten the rearranging, the wait will be shorter, but someone still has to decide where the machines stand.

Where the Value Gets Made

The last time intelligence started with electrons, the value waited more than thirty years for people to move the machines. Huang's cake tells you what to build, but not when it pays, or who does the rearranging in between. In Denmark, a teacher who never opened a chatbot polices her students' use of one, and no layer of the cake has her in it.