Stocks for Thoughts · Nobody taxes a machine

Nobody holds a position that long.

How many years passed between a general-purpose technology being invented and the year it did the most for productivity. Every one of these was real. None of them paid on anyone's schedule.

years of waiting the window of peak productivity effect

Technology Years from invention → Lag

Steam

Watt's separate condenser, 1769 → peak contribution to British productivity in the 1850s–70s

~90 yrsa century

Electricity

Edison's first central station, 1882 → the factory productivity surge of the 1920s

~40 yrsrebuilt plants

Computers

The microprocessor, 1971 → the US productivity acceleration of 1995–2004

~25 yrsSolow's paradox

Artificial intelligence

The transformer paper, 2017 → the capital is committed; the payoff window is the open question

9 yrs inyou are here

9 years in — AI today

Years elapsed →

0 25 50 75 100

Is it halving each time? No.

The lag depends entirely on which moment you call "the invention." Here are the same three technologies under three consistent rules. Every rule shortens. No two agree on by how much.

Rule for dating the invention	Steam	Electricity	Computers	Ratio
First working machineNewcomen 1712 · Faraday 1831 · ENIAC 1945	~148 yrs	~95 yrs	~50 yrs	1.6× / 1.9×
First commercial deploymentWatt 1769 · Edison 1882 · microprocessor 1971 — used above	~90 yrs	~40 yrs	~25 yrs	2.3× / 1.6×
Mass-market formTrevithick 1801 · AC grid 1888 · IBM PC 1981	~55 yrs	~37 yrs	~14 yrs	1.5× / 2.6×

And if you extrapolate to AI anyway

Four defensible ways to run the trend forward from the transformer paper in 2017. They land sixteen years apart, and the range straddles today.

Comin & Hobijn's measured rate, 0.43 yrs faster per year

2022already passed

Straight line through all three points

2027next year

Each lag halves the one before

2030four years out

The most recent leg only, 1882 to 1971

2034eight years out

A sixteen-year spread out of three data points. That is not a forecast — it is the absence of one.

Read the dashed line, not the bars. Nine years after the invention, steam was adding one or two hundredths of a per cent a year to British productivity, electricity had rewired almost nothing, and the computer was still forty years from showing up in the statistics. At the same point on the clock, every one of these looked exactly like a technology that wasn't working.

The shortening is real. Comin and Hobijn measured adoption across two centuries and found that a technology invented ten years later is adopted 4.3 years faster, at much the same rate before and after 1950. What is not real is the ratio — and the series above cannot tell you when. It is three points, hand-dated, drawn from three different measurements in three different economies, and selected for having worked at all. Nuclear power was invented in 1954 and has never produced a measurable productivity peak.

So the honest reading is directional and nothing more: later technologies pay off sooner than earlier ones, and nine years in, the numbers still cannot separate a real revolution running late from a story that never arrives.

Method: "invention" is dated to the enabling patent or first commercial installation; "peak effect" to the period in which economic historians measure the largest contribution to total factor productivity growth. These are windows, not events, and reasonable historians place them a decade either way. Sources: Crafts, Understanding Productivity Growth in the Industrial Revolution (Economic History Review, 2020), for steam's contribution rising from 0.01–0.02pp a year before 1830 to roughly 0.4pp after 1850; David, The Dynamo and the Computer (AER, 1990), for the electricity lag; Solow's 1987 remark and the BLS multifactor productivity series for the 1995–2004 acceleration; Vaswani et al., Attention Is All You Need (2017), for the transformer; Comin & Hobijn, An Exploration of Technology Diffusion (HBS working paper 08-093; AER 2010), for the 4.3-years-per-decade acceleration in adoption lags, a mean lag of 47 years and a range of 5 to 118 — note their measure is the lag to adoption, not to peak productivity effect, a related but not identical quantity, which is one more reason the two should not be blended into a single trend line. The sensitivity table's alternative dating rules are applied consistently within each row; the peak-effect windows are unchanged. The AI row shows elapsed time only — no forecast is implied by the length of the faded bar, and the four extrapolations are shown precisely because they disagree. Not investment advice.