Why Revolutionary Technologies Often Produce Terrible Investments
Aug 5, 2026
One of the most stubborn myths in investing is the belief that spotting a revolutionary technology early is enough to produce extraordinary returns, when history suggests something far less comforting, because some of the greatest innovations humanity has ever produced destroyed enormous amounts of shareholder capital in their early phases not because the technology failed to matter, but because investors badly overestimated how quickly a transformative idea could become an economically self-sustaining industry.
Railroads eventually redrew the commercial map of entire nations, yet thousands of railway companies disappeared through bankruptcy before the network matured, electricity transformed factories and households but only after decades of speculation, consolidation and financial wreckage, and the internet permanently changed communication, retail, advertising and entertainment while the first generation of dot-com investors learned the hard way that technological inevitability and investment success are not remotely the same thing.
This distinction matters because markets have a bad habit of confusing the destination with the journey required to reach it, and investors rarely lose money simply because they identify the wrong future, since more often they lose because they assume the path toward that future will be smooth, orderly and economically efficient, an assumption for which history offers remarkably little support.
Every technological revolution begins with genuine innovation, attracts extraordinary optimism, draws in unprecedented amounts of capital and eventually reaches a point where expectations become detached from the financial realities required to justify them, so that the technology often survives and even flourishes while much of the capital chasing it does not.
Artificial Intelligence Is Entering the Same Dangerous Zone
Artificial intelligence increasingly appears to be moving into precisely this stage of development, and that should not be read as an argument against AI itself or as another lazy attempt to declare the entire sector a bubble simply because valuations look elevated, because those arguments usually generate more heat than light and quickly collapse into debates over whether current frontier models constitute real intelligence or whether artificial general intelligence is five years away, fifty years away or simply unknowable.
Those questions are intellectually fascinating, but they are not the questions investors should be asking, because markets do not price philosophical definitions and instead price expectations, future cash flows and the probability that today’s investment will generate tomorrow’s economic return.
The more useful question is not whether artificial intelligence will transform the economy, because it almost certainly will in ways that are still difficult to fully appreciate, but whether financial markets are pricing that transformation as though it will unfold in a relatively smooth and uninterrupted fashion, despite overwhelming historical evidence showing that technological revolutions almost never develop that neatly.
Every major transition in economic history has moved through an uneven sequence of optimism, overinvestment, temporary excess capacity, financial retrenchment and eventual consolidation before the technology finally matured into the sort of indispensable infrastructure that later generations simply take for granted.
The Real Pattern Is Not Technology, But Capital Behaviour
The invariant hiding beneath these episodes has surprisingly little to do with the technology itself and almost everything to do with the behaviour of capital, because financial markets are often brilliant at recognising genuine innovation while remaining remarkably poor at judging the economics required to commercialise it.
Once investors become convinced that a technology will reshape civilisation, money begins moving toward that future with increasing speed, early successes reinforce confidence, rising valuations lower the cost of raising more capital, easier financing accelerates expansion, and that expansion produces impressive revenue growth for suppliers, which then appears to validate the original optimism and encourages still greater investment.
What begins as a rational response to genuine innovation gradually turns into something more fragile, not because investors suddenly become stupid, but because the financial system starts extrapolating present momentum far into the future before the underlying economics have fully matured.
Railways Proved That Being Right About the Future Is Not Enough
The railway boom of the nineteenth century remains one of the clearest examples of this process, because there was never any serious doubt that rail transport represented a revolutionary improvement over canals, horse-drawn freight and coastal shipping, since railways dramatically reduced transportation costs, connected previously isolated markets and fundamentally changed the economics of agriculture, manufacturing and trade.
Investors who understood that future were absolutely right about the technology, yet many were catastrophically wrong about the investments, because the ease with which capital flooded into railway construction produced an enormous expansion in track laying, often far beyond the immediate economic demand required to support profitable operations.
Many routes proved commercially weak, numerous companies failed and shareholders absorbed painful losses, even though the rail network itself ultimately became one of the defining achievements of industrial civilisation.
The Internet Repeated the Same Lesson in Modern Clothing
A remarkably similar pattern appeared more than a century later during the construction of the internet, because few serious observers today would claim fibre-optic networks were a poor technological investment for society, given that they formed the physical foundation on which the modern digital economy was eventually built.
The investment experience of the late 1990s told a very different story, however, because companies laid extraordinary amounts of fibre based on projections of exponential demand growth, all financed by capital markets willing to fund almost anything attached to the internet narrative.
When demand inevitably grew more slowly than those assumptions required, the result was not the disappearance of the technology but years of excess capacity, widespread bankruptcies and brutal valuation collapses, and the irony is that much of the fibre installed during that speculative period later became immensely valuable after ownership had changed hands and much of the original capital had already been destroyed.
Cloud Computing Was More Disciplined, But Still Not Smooth
Cloud computing followed a more disciplined path, yet even there the relationship between technological success and investment returns proved far more complicated than the simple stories told after the fact, because building hyperscale infrastructure required enormous upfront investment before enterprise adoption could fully justify the spending.
Many participants underestimated both the time required for businesses to migrate critical workloads and the ongoing operational costs attached to maintaining global cloud infrastructure, and although the technology succeeded because the economics eventually matured, that maturation occurred over many years rather than according to the smooth growth curves investors initially preferred to imagine.
AI Has Not Yet Reached Broad Diversification
Artificial intelligence shares important features with each of these earlier transitions, but it also introduces structural complications that make the current cycle unusually difficult to analyse, because earlier technological revolutions eventually broadened into highly diversified sources of demand, with railroads serving countless industries, electricity powering nearly every factory, the internet connecting billions of individuals and millions of businesses, and cloud computing spreading gradually from technology companies into almost every major sector of the global economy.
Diversification became the mechanism through which extraordinary capital expenditure was eventually justified, but the AI economy has not yet fully reached that stage.
This should not be misunderstood as a criticism of artificial intelligence or as proof that widespread adoption will fail to arrive, because enterprise use continues to accelerate, new applications appear almost weekly and productivity gains are becoming increasingly visible across a growing number of industries.
Nevertheless, much of today’s infrastructure spending remains concentrated inside a surprisingly small ecosystem of organisations whose investment decisions exert enormous influence over the semiconductor, networking, power and data centre supply chains, and that concentration is one of the least discussed but potentially most important features of the current investment cycle.
Growth Is Not the Same as Resilience
Most investors instinctively associate rapid growth with increasing diversification, but history suggests the relationship is often far more complicated, because a market can expand at extraordinary speed while remaining heavily dependent on a narrow group of participants, and when that happens the apparent strength of aggregate demand can hide a much more fragile underlying structure.
The difference between a growing market and a diversified market is subtle during periods of expansion, because both produce impressive revenue growth, rising valuations and optimistic forecasts, but the distinction becomes painfully visible once the pace of investment begins to change.
This is why the current AI cycle deserves closer examination, because its future may depend less on whether artificial intelligence succeeds as a technology and more on how capital continues to move through the financial architecture supporting its expansion.
The Four Risks Markets Are Quietly Pricing as Certainties
Every investment cycle develops a dominant narrative that slowly hardens into accepted wisdom, and once that happens the market often stops asking whether its assumptions are reasonable and instead begins arguing only about the speed with which those assumptions will become reality.
During the railway boom, the assumption was not merely that rail transport would reshape commerce, but that every new line would eventually become economically viable, and during the internet boom investors stopped debating whether online commerce would transform business and instead assumed that nearly every company attached to the digital economy would ultimately justify its valuation.
The current AI discussion increasingly shows the same pattern, because the debate has moved beyond whether artificial intelligence will matter and into a set of assumptions about how rapidly adoption will broaden, how profitable deployment will become and how long the current pace of infrastructure investment can continue without interruption.
Markets rarely make one enormous forecasting mistake in isolation, because they usually make several smaller assumptions that appear reasonable individually but together create an expectation structure with very little room for disappointment.
Risk One: Demand Is Large, But Still Concentrated
The first assumption concerns demand concentration, and it may be the most underappreciated variable in the entire AI ecosystem because investors often confuse the scale of spending with the breadth of its origin.
Headlines about hundreds of billions of dollars in planned capital expenditure naturally create the impression that demand has already become broad, diversified and deeply embedded across the economy, yet a closer look suggests something more concentrated, with Microsoft, Alphabet, Amazon, Meta, Oracle, OpenAI, Anthropic and xAI occupying central positions in a network that increasingly determines investment decisions across semiconductors, networking, power and data centres.
That distinction matters because concentrated demand behaves very differently from diversified demand, since millions of independent businesses making separate investment decisions create resilience through offsetting behaviour, while a relatively small number of exceptionally large organisations can transmit changes in sentiment, financing conditions or strategic priorities through the supply chain with surprising speed.
Growth can appear extraordinarily robust while confidence remains high, but the stability may come less from diversification than from the continued willingness of a handful of dominant decision makers to keep allocating capital at an unprecedented pace.
History shows that concentrated demand amplifies both expansion and contraction, as commodity markets have repeatedly discovered when mining companies increased production during periods of intense Chinese infrastructure spending only to see modest shifts in construction activity turn shortages into oversupply.
Memory semiconductors have displayed the same tendency, with periods of heavy investment giving way to sharp corrections once a limited number of large buyers reduced inventory accumulation, and the pattern is not unique to technology because it is a recurring feature of capital-intensive industries whose economics depend heavily on a small group of participants.
Risk Two: Capability Is Not the Same as Return
The second assumption concerns operating economics, and this discussion often becomes needlessly polarised because critics and advocates tend to argue past one another, with one side insisting frontier models are little more than sophisticated autocomplete systems and the other insisting artificial intelligence will transform every industry and justify virtually unlimited investment.
Reality is far more nuanced, because large language models have already demonstrated remarkable usefulness across coding, research, content generation, customer support, software development and numerous other applications, which means their capability is no longer seriously in question.
The more relevant issue for investors is whether those capabilities are translating into sustainable improvements in profitability across the broader economy at a pace sufficient to justify the enormous infrastructure investments currently being made.
There is a crucial distinction between demonstrating capability and generating economic return, because businesses often adopt technologies whose strategic value becomes visible long before their financial contribution can be measured with confidence.
Enterprise software, cloud computing and industrial automation all passed through long periods in which adoption accelerated while measurable returns remained uneven across industries, and artificial intelligence may follow a similar path.
That should not trouble engineers or scientists, because technological progress rarely depends on quarterly earnings, but investors operate under a different constraint, since valuations incorporate future cash flows long before those cash flows materialise, allowing expectations to drift away from operating economics even while the underlying technology continues advancing exactly as expected.
Risk Three: Intelligence Has to Be Maintained
The third assumption receives remarkably little attention despite its potential influence on long-term industry economics, because much of today’s discussion focuses almost entirely on the cost of training larger models while underestimating the ongoing costs of owning, monitoring and maintaining deployed AI systems.
Once artificial intelligence becomes embedded inside business processes, the economics extend far beyond initial deployment, because models require continuous monitoring as the world they represent refuses to sit still, with consumer behaviour evolving, regulations changing, markets adapting, scientific knowledge expanding, fraud tactics improving and language itself shifting over time.
Systems that once produced highly reliable outputs gradually become less dependable unless they are continually evaluated, retrained, fine-tuned and governed, which is why engineers speak of model drift even though the broader economic implications go well beyond a technical maintenance issue.
Traditional software usually becomes cheaper to operate once the original development is complete, because maintenance costs are often modest relative to the initial build, but artificial intelligence increasingly appears to follow a different model in which intelligence is not merely created once and left alone but continuously maintained, recalibrated and supplied with computational resources simply to preserve its usefulness.
Inference consumes energy, monitoring requires specialised personnel, compliance demands governance frameworks, security adds complexity and data pipelines require constant updating, and while none of those activities generate the excitement associated with a more capable frontier model, together they may determine whether AI ultimately resembles conventional software or critical infrastructure from an economic perspective.
That distinction matters enormously because markets assign very different valuation frameworks to software businesses and infrastructure businesses, since software receives premium valuations partly because investors assume high operating leverage, recurring revenue and low marginal costs, while infrastructure requires continuous capital expenditure just to preserve operational capacity.
If AI increasingly exhibits infrastructure-like economics rather than software-like economics, then many valuation assumptions may need revision, not because the technology disappoints but because its financial characteristics differ from what markets currently appear to expect.
Risk Four: Capital Is Abundant Until It Becomes Selective
The fourth assumption concerns financing, and it may be the operator connecting all the others, because artificial intelligence is becoming one of the most capital-intensive technological build-outs in modern economic history, requiring extraordinary investment not only in semiconductors but also in power generation, electrical transmission, cooling systems, networking equipment, specialised memory, land acquisition and dense data centre construction.
This expansion remains entirely feasible while capital markets are willing to provide financing on attractive terms, and rising valuations often reinforce the cycle by lowering the effective cost of raising additional capital, encouraging further expansion and strengthening the impression that demand will remain virtually unlimited.
Yet capital has always possessed one trait that markets repeatedly underestimate, which is that it remains abundant until it suddenly becomes selective.
That shift rarely happens because investors collectively abandon belief in the underlying technology, and more often it occurs because uncertainty rises around the timing of future returns.
Executives seldom reduce capital expenditure in neat, gradual fashion over many consecutive quarters, because while confidence remains high investment often accelerates as nobody wants to surrender competitive advantage, but once uncertainty begins to rise, projects are postponed, financing becomes more disciplined and assumptions about future demand receive much greater scrutiny.
Capital expenditure therefore tends to move in bursts rather than smooth curves, creating feedback loops that amplify both optimism and caution.
This pattern has appeared across telecommunications, liquefied natural gas, mining, solar manufacturing, memory semiconductors and commercial real estate, where expansion continued until the economic narrative supporting further investment became only marginally less convincing, at which point capital allocation adjusted with surprising speed.
The resulting slowdown often appears abrupt, not because demand vanished overnight, but because markets had previously assumed that exceptionally favourable conditions would persist indefinitely.
Artificial intelligence may ultimately prove no different, because the technology can keep advancing, enterprise adoption can keep broadening and productivity gains can keep accumulating while the financial architecture funding today’s expansion still moves through periods of overinvestment and retrenchment.
History suggests those cycles are not evidence of technological failure, but an almost unavoidable consequence of how capital behaves during periods of profound structural change.
The Transition Is the Investment
Every generation of investors eventually convinces itself that history has become less relevant because the technology under consideration seems too important, too transformative or too unprecedented to be constrained by the lessons of earlier investment cycles.
Railroads were going to change civilisation, electricity was going to redefine industry, the automobile was going to reshape society and the internet was going to connect the world, and none of those predictions proved wrong.
The mistake lay elsewhere, because investors gradually stopped distinguishing between the certainty of the destination and the uncertainty of the journey, assuming that because a technology would eventually transform the economy, the financial path toward that future would unfold with the same inevitability.
Artificial intelligence appears to be nearing that same crossroads, because the technology continues advancing at extraordinary speed, frontier models become more capable with each generation and businesses across nearly every sector are experimenting with applications that would have seemed implausible only a few years ago.
None of that should be dismissed or underestimated, since the probability that artificial intelligence becomes one of the defining technologies of the twenty-first century appears remarkably high.
Yet history suggests that technological importance and investment returns rarely move in perfect synchrony, because the capital required to build the future almost always arrives long before the economics required to sustain that future have fully matured.
Markets Are Pricing a Smoother Journey Than History Usually Allows
This distinction becomes especially important when markets begin pricing not only eventual success but also the speed, efficiency and smoothness with which that success is expected to occur.
The current AI investment cycle increasingly assumes that enterprise adoption will broaden steadily, operating economics will improve predictably, maintenance costs will remain manageable and capital markets will continue financing one of the most ambitious infrastructure build-outs in modern history without meaningful interruption.
Each assumption may prove correct over the long term, but together they leave surprisingly little room for the delays, recalibrations and retrenchments that have accompanied every previous technological revolution.
Markets usually describe those interruptions as disappointments, but history suggests they are better understood as the normal process through which revolutionary technologies evolve from impressive demonstrations into economically sustainable industries.
Railroads required decades of consolidation before they became consistently profitable, fibre-optic networks spent years operating with excess capacity before internet traffic finally filled them, and cloud computing demanded enormous upfront investment before enterprises moved enough workloads to justify the infrastructure already in place.
The pattern repeats because technological capability almost always advances faster than the institutions, business models and economic incentives required to exploit it fully.
The Real Question for AI Investors
That may be the central lesson of the current AI cycle, because investors are not really being asked to decide whether artificial intelligence matters, since that question has increasingly been answered.
The more important question is whether today’s valuations properly reflect the economic journey still ahead, or whether they assume a level of continuity that history rarely grants.
Financial markets are excellent at recognising transformative technologies, but they are just as capable of underestimating the complexity of converting those technologies into durable, widespread and consistently profitable economic systems.
The real investment decision, therefore, is not whether artificial intelligence will change civilisation, because it almost certainly will, but whether the financial architecture supporting that transition has already discounted a future that is likely to be far more uneven than current market expectations appear willing to admit.
History does not suggest transformative technologies fail, and in fact it suggests the opposite, because they often reshape the world in ways few people initially imagine.
What history also suggests, however, is that the first great wave of capital pursuing those transformations often mistakes inevitability for immediacy, confusing the certainty of long-term progress with the assumption that the path toward it will be smooth, uninterrupted and financially rewarding for everyone involved.
History offers remarkably little support for that belief, because technological revolutions usually fulfil their promise, but rarely according to the timetable, valuations or narratives investors first imagine, and there is little reason to believe artificial intelligence will become the first exception.
