A Bubble with Four Clocks Out of Sync: AI Could Change the World and Still Be a Catastrophic Investment

The biggest mistake in the increasingly heated debate over artificial intelligence—at least in economic terms (which will be our main concern today, regardless of the fact that someone at Anthropic claims there is a 10 percent chance AI will destroy humanity, we can save those chills for another article)—is the assumption that there are only two possible outcomes. Either AI is already a revolution that will transform the economy from the ground up, or it is a bubble destined to end in spectacular losses. But we have seen this kind of thing before, which means there is clearly a third possibility: the technology can succeed while the investment in it proves disastrous.

The future may arrive exactly as the optimists predicted (though they may in fact be techno-feudalists—but that, too, is a subject for another day), only several years too late for those who financed it. The infrastructure and useful products may survive while companies collapse and shareholders lose their money. Technological truth and market price are not the same thing. Britain learned that lesson in the mid-19th century.

The railway was not a fraud, of course. It transported goods and passengers faster and more cheaply than the alternatives. That was precisely why the bubble known as Railway Mania seemed so convincing—until it burst. The index of railway shares doubled between 1843 and 1845, while investment at the peak reached roughly seven percent of Britain’s national income.

Then costs exceeded projections, revenues fell short and credit tightened. By the end of the decade, the index stood roughly 65 percent below its peak. The railway lines remained, but many of their original owners did not.

Three clocks that never tell the same time

One useful way to understand technological bubbles is to imagine three clocks. This is not a new invention of ours, merely our interpretation of a well-established conceptual framework used within the industry.

The first is the Capability Clock. It measures whether a technology can actually do what is claimed for it. Consider the airship. It once appeared to be the future of intercontinental travel, but its structural fragility, vulnerability to weather and safety problems were never adequately resolved. The Hindenburg became the symbol of its demise, while conventional aircraft continued to become faster, safer and more practical.

The second is the Profitability Clock. Here we can take another example from aviation. Concorde proved that supersonic passenger flight was possible, but it consumed enormous quantities of fuel, carried relatively few passengers, cost a fortune to maintain and, because of its sonic boom, could operate on only a limited number of routes. A technological triumph ended up as a commercial niche. It is not enough for a machine to be capable of doing something. It must be able to do it at a price someone is prepared to pay.

The third is the Adoption Clock. Railways were capable and eventually profitable, but they had to wait for cities, factories, warehouses and sufficient traffic to emerge along their routes. The same happened with computers and the internet (although the "early internet" had a charm of its own, see: When the Web Was Young and Belonged to Everyone). A technology can be purchased quickly. Reorganising an institution around it is much harder. These three clocks provide a useful mental model for interpreting today’s AI euphoria. But we should add one more.

The fourth clock: capital does not measure time in decades

Political economy requires another clock: the Capital Clock.

This fourth clock does not ask when a technology will mature, but when the investment will begin producing an acceptable return. It measures quarters, debt maturities, share prices and the patience of investment funds. If the Capability Clock points to 2032 and the Adoption Clock to 2038, while investors must prove the business case by 2028, the conditions for a financial collapse are already in place—even if the long-term optimists are right.

Economist Carlota Perez describes technological revolutions as processes in which a period of installation—the construction of infrastructure, financial euphoria and experimentation—precedes broader deployment. A crisis often lies between the two, destroying inflated valuations but, crucially for our story and all the similar stories that came before it, not the technology itself. In other words, a bubble may be the mechanism through which the market mobilises capital too quickly, builds too much and discovers only after the crash what was genuinely useful.

You have surely noticed that "AI" is now mentioned quite literally everywhere, even in advertisements that sound, to put it mildly, ridiculous. Every major online platform is suddenly pushing an "AI assistant" that perhaps no one ever asked for. Google’s summary, which now appears whenever you search for something, irritates many users and, as if that were not enough, is beginning to make the "search engine" itself seem somewhat irrelevant. In other words, AI saturation can already be felt everywhere. All of this may suggest that the bubble is approaching its breaking point.

Society may need a decade or two to extract the full benefit of a new technology, but the capital financing it often cannot wait that long. The first investors pay for the learning process, the infrastructure and the failed business models. The next owners buy whatever survived, at a price that finally makes sense.

The difference this time is that so much money has been poured into this initial phase—let us call it the inflation of the bubble—that the resulting explosion could shake the entire world.

AI in 2026: one clock is racing while the others lag behind

AI’s Capability Clock is clearly moving. Systems that produced clumsy prose only a few years ago can now write code, analyse documents, translate and perform some forms of research. Some can build entire applications on their own. But the progress is uneven. A model can excel at a task with a clearly defined answer, then confidently fail in a perfectly mundane situation. It can accelerate the work of an expert, but it can also produce an error that is extremely costly to identify.

That is why the Profitability Clock is more complicated than the number of users might suggest. AI becomes profitable for a business only when the additional revenue or savings exceed the cost of the model, infrastructure, data cleaning, integration, supervision, legal risk and error correction. In a cheap piece of marketing copy, an error means a few minutes of revision. In medicine, law, finance or an industrial system, a single error can wipe out the savings generated by hundreds of previous successes.

Private profitability, moreover, is not the same thing as social productivity. A company can increase its profits by using AI to dismiss workers, intensify surveillance of those who remain and shift part of the workload onto its customers. That does not mean society has produced more useful value. It may simply mean that wages have been converted into profit margins, while the cost of insecurity has been transferred to households and ultimately to the state.

The Adoption Clock may move more slowly still. Opening a chatbot takes seconds. Reorganising a hospital, bank, factory or government ministry requires cleaning up its data, changing procedures, training staff, assigning responsibility and connecting legacy systems—which quite often means very old systems that "resist" change. The OECD (Organisation for Economic Co-operation and Development) therefore stresses that widespread adoption depends on skills, organisational change, infrastructure and institutions, not merely on access to models. This is both entirely understandable and entirely correct.

This is a new version of the so-called Solow paradox: the technology is everywhere, yet its full impact remains difficult to detect in productivity data. In its more recent figures, the OECD sees only early and uncertain indications of a possible AI effect. The familiar "productivity J-curve" explains why general-purpose technologies first demand vast, largely invisible investments in software, training and organisational change, producing measurable gains only later.

The Capital Clock, however, is already ticking loudly and nervously. JLL, one of the world’s leading investment firms, estimates that the four largest hyperscale cloud providers will spend roughly $725 billion in 2026, 77 percent more than the previous year, mostly on AI computing and data centres. Meanwhile, a Reuters analysis of five major companies suggests that their capital expenditure could exceed their combined free cash flow in 2027. In other words, the question is not merely whether AI can make money one day, but whether it can make enough of it, quickly enough, to justify the infrastructure being built now.

That infrastructure, of course, is not quite the same as a railway line. A data-centre building and its power connection may last for decades, but the most expensive computing layer becomes obsolete far more quickly. Goldman Sachs estimates the economic life of AI chips at roughly four to six years, compared with about twenty years for buildings and even longer for energy infrastructure. Older chips will not necessarily become worthless, but a new generation may dramatically alter the cost of computing before the previous one has even paid for itself.

You may have encountered this problem of obsolescence and progress yourself, not through chips but through the capabilities of AI systems. If, for example, you used AI to help with programming in 2024, you will remember what that was like. The assistance was real and tangible: AI could write a function much faster than you could type it. But as soon as the task became more complex, the chances of error began to accumulate. Still, you persisted and finished the project in, say, six months. Had you waited until 2026 and used today’s considerably more advanced models, the entire project might have taken only a month, perhaps even less.

Who buys the future at a clearance sale?

If the AI bubble bursts—and it would be very unusual if it did not—that will certainly not prove that artificial intelligence was a fraud. But it may mean that prices were wrong (although powerful models are not expensive, except that they are if you derive no benefit from them), expectations were premature and investment horizons were too short. It may also mean that the market built more capacity than current demand could profitably absorb.

But here we arrive at the greatest problem: the crash would not be socially neutral. Workers dismissed in the name of automation will not recover their lost wages because the project later failed. Local communities will not be relieved of the costs imposed on their power grids, water supplies and land simply because a data-centre owner miscalculated the return. Pension funds and small shareholders may bear the losses, while the largest technology companies use their enormous reserves to acquire the chips, patents, experts and infrastructure of bankrupt competitors.

And that is before we even consider the companies using AI merely as an alibi for mass layoffs. Under ordinary circumstances, such cuts might trigger panic among shareholders and send the company’s share price tumbling. But now those companies can exploit the moment and declare: "Everything is going well—we are laying people off because we are introducing AI." Shareholders swallow the story without hesitation.

AI looks like the future, but the past still has something to teach us. After Railway Mania, the tracks did not disappear. They were consolidated in the hands of stronger companies that bought them cheaply. An AI correction would therefore not shatter technological power but concentrate it even further. The first generation of capital finances the future, the public pays part of the cost of its physical infrastructure and energy consumption, and the winners acquire the system at a discount after the crash.

In economic and financial terms, AI may therefore prove to be a little of everything: a dead end in some areas, an expensive niche in others, essential infrastructure elsewhere and a platform for businesses we cannot yet imagine.

One thing is far more certain. The four clocks will not tell the same time. Technology will advance at its own pace, companies will pursue profits, institutions will adapt slowly, and capital will demand its return before history has finished conducting the experiment.

Leaving aside the exceptions to the rule—the ones that always make for such good stories—in capitalism, the future rarely belongs to those who imagined it first. More often, it belongs to those who still have enough cash after the crash to buy the ruins.