← Back to the library

    The Picturephone and the S-curve

    holds up

    This is documented history rather than a laboratory result. The events are well recorded. What they teach is an argument, and the entry makes it openly.

    AT&T had believed in video telephony for decades. Bell Labs demonstrated a live video call as early as 1927, with the US Secretary of Commerce on one end, and by the 1964 World's Fair in New York the company was ready to show the public its future. The Picturephone booths drew queues, Lady Bird Johnson took part in the inaugural call, and the demonstration sat alongside IBM mainframes and models of the planned World Trade Center as an exhibit of what was coming. AT&T then opened public Picturephone rooms in New York, Chicago and Washington, where two rooms could be rented for between 16 and 27 dollars for the first three minutes. In the following six months, 71 calls were made. The rooms stayed open until 1970, and in their final year not a single call was placed.

    The company read the early silence as a pricing and packaging problem rather than a verdict. It pressed on, projecting more than one million Picturephones in use by 1980, with revenues to match, and launched commercial service in Pittsburgh on the first of July 1970 at 160 dollars a month. Chicago followed in 1971. The service peaked in early 1973 at 453 subscribers. By the end of July 1974 five units remained in service in Pittsburgh, and AT&T, which had spent more than 500 million dollars over some fifteen years by the most repeated accounting, shut the project down and described its own product as a concept looking for a market. The company tried again in 1992 with a colour videophone for ordinary lines, and discontinued it in 1995. Video calling finally arrived decades later through a different door entirely, as software riding on the internet.

    The curve AT&T's forecast assumed is real. Adoption of successful technologies does follow an S: slow at first, then explosive, then flat as the market saturates. The shape appears because growth depends on two quantities at once. The number of people who already have the thing drives the spread, because they are the ones who recommend it, while the number of people left to convert shrinks as adoption rises. Early on the first number is small, so growth is slow. In the middle both are large, so growth peaks. Near the end the second number runs out, so growth flattens, and the total traces an S. The most rigorous recent test, by Doyne Farmer's group at Oxford, collected data on 47 technologies from canals to mobile phones and found the shape remarkably universal, with a single curve family explaining more than half the variance in diffusion at the point of fastest growth.

    Where the evidence stops, and for this entry the limits are the finding. The curve describes the past of technologies that succeeded, and it has two blind spots that sit exactly where forecasts need it most. First, the ceiling cannot be seen from early data: before the bend, a curve that is about to explode and one that is about to stall are mathematically indistinguishable, because the running-out-of-room term has not yet begun to bite. Second, the record of completed substitutions hides the stalled ones. 3D television, the Segway and the first wave of virtual reality headsets all showed an early rise of the kind the classic substitution model said guarantees completion, and stopped. The Picturephone is the cleanest of these cases because the company in question was among the most sophisticated technical organisations on earth, the technology worked, and the demand it planned for never arrived at all. The same Oxford group also documents that fitting these curves is statistically treacherous even when the shape is right, so that naive extrapolations come out overconfident. The curve describes how successful diffusion looked after the fact, and nothing in it says which technologies will diffuse.

    The S-curve in a pitch deck is an assumption wearing the clothes of a forecast, and the way to interrogate it is not to argue about the drawing. It is to ask what engine is supposed to be underneath it, because the technologies that actually climbed the curve were driven up it by identifiable machinery, and the ones that stalled were not. Three engines cover the record.

    The first is cost that falls with cumulative production. Solar panels became about 20 percent cheaper with every doubling of installed capacity, and held that rate for four decades, which is why that curve never stalled: every unit built made the next unit more affordable, recruiting buyers the last price could not reach. The second is users who bring users. A telephone became more valuable with every new subscriber, since each one was somebody you could now call. The Picturephone ran this engine in reverse: at 160 dollars a month almost nobody had one, so there was almost nobody to call, and each person who declined it made it slightly more useless to everyone else. The third is infrastructure that already exists. ChatGPT reached users in nearly every country within days of release, the fastest arrival on record, because the infrastructure it needed, the internet and billions of connected devices, had already been built for other reasons. The Picturephone needed network capacity that did not exist at scale, so its diffusion was gated by construction.

    The test is three questions asked of any technology bet: does the cost fall as production scales, does each new user make it better for the next one, and does the infrastructure already exist. The telephone eventually had all three. The Picturephone had none, and the technology itself was never the main problem. A bet with three yeses has machinery under its curve. A bet with none is a plan for a million users from a product that may not find its first thousand, and the drawing on the slide cannot tell you which one you are holding, because at the early stage both look the same.

    The deeper discipline is about what the chart you are shown leaves out. The successful substitutions that populate every S-curve illustration are survivorship bias in its purest form, since the stalled curves are quietly dropped. A forecast built from your own launch plan rather than from what happened to comparable products is the planning fallacy. And whether the freed capacity of a genuinely rising technology turns into results is a separate question from its arrival, which is where Solow's productivity paradox begins.

    Source: The Picturephone: Engineering and Technology History Wiki, Picturephone, and Britannica's entry on the device, for the 1964 World's Fair debut, the 500 million dollar development cost, the one million by 1980 projection, the 1970 Pittsburgh launch at 160 dollars a month, the 453 subscriber peak in early 1973, and the five units remaining by July 1974. The public rooms, their 71 calls in six months and the callless final year are as reported in retrospectives drawing on AT&T's archives. The "concept looking for a market" line is AT&T's own, widely quoted. The universality of the S-curve: Wagenvoort, Lafond, Dyer and Farmer, The Universality and Predictability of Technology Diffusion, INET Oxford Working Paper 2026-19. The solar learning rate: Our World in Data, Learning Curves and Wright's Law, at 20.2 percent per doubling of cumulative capacity. Generative AI's arrival speed: WIPO, World Intellectual Property Report 2026. The stalled cases and the statistical difficulties of curve fitting: the same Oxford paper and the sources gathered in the review this entry draws on.

    The book, if you want to go further

    Diffusion of Innovations

    Everett M. Rogers, 1962

    The book that founded the study of how new things spread, first published in 1962 and revised through five editions. Rogers synthesised hundreds of diffusion studies into the questions this entry keeps asking: what makes an innovation spread, why adoption follows the S, and which attributes of a technology, relative advantage, compatibility, complexity, trialability and observability, govern its speed. The Picturephone fails his checklist on almost every line, decades before anyone could have run the numbers, which is the best argument for reading him now: the framework would have caught it in 1964.

    Draw your own card. It does not take long, and it rewards taking your time.