The Railroad Tracks Survived

The Railroad Tracks Survived

Why AI may be a bubble and still change everything

Over the past few weeks, several conversations about AI kept connecting in my head.

Terence Tao asked what happens to mathematics when machines can solve problems humans no longer understand. Ben Thompson asked what happens when frontier capability outruns our ability to turn it into useful, profitable products. And a historical framework I came across offered a way to connect them: transformative technologies run on separate clocks for capability, profitability and adoption.

AI may be advancing on all of them, just at very different speeds. That gap, between what AI can do and what our economies and institutions can absorb, explains why AI can be a bubble and still change everything.

Three clocks

History offers three instructive cases.

Airships once looked like the future of long-distance travel, but they eventually lost the capability race to airplanes.

Concorde had the opposite problem. It crossed the Atlantic in roughly half the time of a conventional jet, yet it struggled to make that capability broadly economic.

Railroads are the more interesting case, because they worked and eventually paid off. The trouble was timing. Tracks, bridges and stations needed enormous capital up front, while towns and factories took decades to reorganize around them. Capital moved faster than adoption, and the railway mania that followed destroyed fortunes. The railways themselves survived and became more important than their early promoters imagined.

The dot-com era repeated the pattern. Vast amounts of fibre were laid before anyone needed it. Investors got the timing badly wrong while being broadly right about the technology.

Applying the clocks to AI

Most arguments about an AI bubble blur these clocks into a single question. Separating them helps.

Capability. Increasingly, AI can do useful work. Coding and mathematics make the clearest case because their outputs can be checked. Models that stumbled on basic tasks a few years ago now write substantial software and solve difficult problems.

But capability is really many clocks running at different speeds. A model can be excellent at code and unreliable at long, autonomous tasks. It can reason well in mathematics and confidently invent facts where truth is hard to verify. The trajectory is impressive and very uneven.

Profitability. The public conversation usually asks whether OpenAI or Anthropic will turn a profit. The question that decides adoption is whether using AI is profitable for the customer.

Suppose an AI system costs ₹10 lakh a year and removes ₹30 lakh of work. That looks compelling. Now suppose it also creates ₹10 lakh of supervision, ₹8 lakh of rework and ₹15 lakh in the cost of errors. The saving has quietly become a loss.

The unit that matters, then, is the total cost of producing a sufficiently reliable outcome. That number varies enormously by domain, because the cost of a single error does. A flawed marketing image costs almost nothing. A flawed contract clause or credit decision can erase the value of thousands of successful interactions. AI will likely become profitable first in low-stakes work that is easy to verify, and move into high-stakes work only as verification gets cheaper.

Adoption. This may be the clock we misread most. Hundreds of millions of people use AI chatbots, which looks like extraordinarily fast adoption. Most of it is superficial: someone asking a model to polish an email. Integrated adoption, where a company redesigns a workflow around AI, requires new data architecture, governance, training and incentives. That takes years.

We have seen this before. In 1987 Robert Solow observed that you could see the computer age everywhere but in the productivity statistics. Computers were widespread long before organizations learned to restructure around them. AI may go through its own version of that paradox.

A fourth clock, and what a bubble really is

Thompson's writing suggests a fourth clock: productization. Having intelligence is different from having a product people rely on. If frontier models become good enough, competition may shift towards whoever builds the most useful product around a sufficiently smart model. As intelligence becomes cheaper and more interchangeable, value moves outward to context, workflow, distribution and trust. Apple is one illustration: it may never own the leading model, but it owns the devices, the personal context and the trust of its users.

Where does the bubble fit? I don't think capital is a fifth clock. It is better understood as the gap between what investors expect and where the four clocks actually stand. A bubble forms when capital prices in capability, profitability, adoption and products that have not yet arrived.

That definition says nothing about whether the technology is real. Railroads were revolutionary, and railway stocks formed a bubble. The internet reshaped civilization, and dot-com valuations became absurd. It is entirely consistent to believe that AI is transformative and that the current investment cycle contains a large bubble.

The same gap at every scale

Tao pushes the question further. If an AI proves ten thousand theorems that no human understands, has mathematics advanced? In a narrow sense, yes. Yet mathematics exists to produce understanding as well as answers.

I feel this most sharply in chess. When I run my own games through an engine, it tells me a position is +0.8 in an instant. What it cannot give me is the reason, in a form I can carry into the next game. The number is cheap. Understanding why the number is what it is still has to be built by me, move by move. As answers become abundant, the understanding behind them may become more valuable, and harder to come by if we stop doing the work.

Yuval Noah Harari sees the same gap at the scale of society. Contracts, laws, money and governments are all built from language, and for all of history humans held a monopoly on reading and writing that code. We no longer do. His concern is AI quietly embedding itself in the systems through which banks lend, courts decide and states govern, before those institutions have adapted. You don't need to accept his most dramatic scenarios to take the question seriously.

Seen together, these perspectives describe one phenomenon at different scales. Tao worries capability will outrun human understanding, Harari that it will outrun institutional control. Thompson argues it is already outrunning productization, and the bubble framework says it may outrun economic adoption. In each case, capability is arriving faster than the systems around it can absorb.

Where the bottleneck is moving

For the past few years, the central question has been how intelligent we can make the model. The next decade will be shaped more by questions of absorption: whether these systems can be trusted in high-stakes work, and whether organizations and institutions can redesign themselves quickly enough.

Which analogy fits, then? AI is clearly not an airship; the capability is real. Nor does it look like Concorde, since its costs are falling rather than rising. It looks more like a general-purpose technology in the lineage of electricity, computing and the internet. If so, we should expect the familiar sequence of overinvestment, painful corrections and consolidation, followed by uses no one predicted at the start. A major correction, if it comes, would tell us a lot about timing and very little about AI's eventual importance.

The railroad tracks survived the railroad bubble.

Perhaps the more interesting question for the coming phase is what becomes scarce when intelligence becomes abundant. My guess is judgment, trust and the ability to decide what is worth doing. Those may become the real bottlenecks of the AI age, and that is where I suspect the most interesting work begins.

Sources and further reading

• Terence Tao on AI and mathematics: [link]

• Yuval Noah Harari on AI and language: [link]

• Ben Thompson (Stratechery) on AI capability and productization: [link]

• The three-clocks framework for technology bubbles: [link]

• Robert Solow, "We'd better watch out," New York Times Book Review, 1987

• On railway mania and the dot-com fibre overbuild: [link]