AI companies may fail. Investors may lose fortunes. Vast quantities of computing infrastructure may prove uneconomic. None of that means artificial intelligence is going away.

Artificial intelligence is apparently collapsing.

The evidence, depending on the week, is that corporate pilots are failing, consumers are tiring of synthetic content, publishers are rejecting machine-generated submissions, infrastructure spending is outrunning revenue, or investors have finally noticed that some of the world’s most valuable companies are making extraordinarily expensive bets whose economic returns remain uncertain.

Some of these concerns are justified. The investment cycle may be overheated. Many AI products are undifferentiated. Corporate adoption is frequently superficial. Enormous amounts of capital are being committed to chips, data centres and energy infrastructure before anyone can say with confidence how much of that capacity will earn an acceptable return.

But the conclusion does not follow from the evidence.

A collapse in AI valuations would not constitute the collapse of AI. It would constitute the repricing of companies, assets and expectations built around it. Markets can decide that investors paid too much for a technological revolution without deciding that the revolution itself was imaginary.

Financial bubbles and technological transformations are not opposites. They often arrive together.

Markets collapse. Useful technologies remain.

The dot-com crash destroyed companies, capital and careers. It did not reverse the internet. Railway speculation repeatedly ruined investors while rail transport continued to reorganise national economies. Telecommunications bubbles burst, but global connectivity expanded. Nokia’s dominance disappeared, not because mobile communication proved temporary, but because Apple changed the definition of the device.

Kodak offers an even more instructive example. A Kodak engineer, Steven Sasson, built the first self-contained digital camera in 1975. The company understood digital imaging and helped create it, but its existing economics remained tied to film. Kodak’s problem was not that it failed to glimpse the future. It was that moving decisively towards that future threatened the business it already had.

That is the pattern worth watching now. The danger is not only that businesses may believe too much in AI. It is that they may confuse disappointment with irrelevance, retreat after poorly designed experiments and leave the difficult work of integration to competitors.

The adoption figures do not describe a technology being abandoned. Stanford’s 2026 AI Index found that 88 per cent of surveyed organisations used AI in 2025, while 70 per cent used generative AI in at least one business function. Yet the deployment of AI agents remained in the single digits across nearly every function. The combination is revealing: experimentation is widespread, but deep operational integration remains early.

AI is therefore not mature enough to justify every valuation attached to it. It is also too widely embedded to dismiss as a passing fashion.

The 95 per cent problem

The strongest argument against enterprise AI is not that people dislike machine-written emails. It is that companies are spending heavily and often receiving little measurable value in return.

MIT’s NANDA initiative reported in 2025 that only about 5 per cent of the integrated generative-AI pilots in its study were producing substantial financial value. The remaining 95 per cent were generating no measurable effect on profit and loss. The study reviewed more than 300 public AI initiatives and drew on interviews and organisational surveys.

The number deserves to be taken seriously. It also supports a more complicated conclusion than the headlines around it suggested.

The report did not principally blame the capabilities of the models. It identified a divide between accessible general-purpose tools and systems that could learn from context, fit existing operations and improve through use. Organisations were buying tools without redesigning workflows, defining success clearly or addressing the data and integration work required to make the technology useful.

This is not evidence that AI has no economic value. It is evidence that purchasing access to intelligence is not the same as changing how a company operates.

A personal computer did not transform an organisation merely because one appeared on every desk. The internet did not create a digital business merely because a company launched a website. Cloud software did not repair a broken process simply because the process was moved online.

Technology produces productivity only after institutions reorganise around it.

That work is slower, less photogenic and more difficult than buying licences or announcing a pilot.

The 95 per cent figure should concern AI proponents. It should also concern business leaders who treat adoption as a procurement exercise. The models may work. The organisation often does not.

This bubble really could be different

Historical analogies are useful, but they can become a substitute for analysis. AI is not simply the internet again, nor is it electricity with better marketing.

One important difference is the pace at which its physical infrastructure can lose relative economic value. Railways, electricity grids and fibre networks were expensive, but their useful lives could extend across decades. Advanced computing hardware operates under a faster technological rhythm.

A 2026 study of Nvidia data-centre GPUs found that FP16 computational performance had historically doubled approximately every 1.44 years and FP32 performance every 1.69 years. Memory capacity and bandwidth improved more slowly, but the central point remains: equipment need not cease functioning to become commercially inferior.

That creates a credible bear case. If demand disappoints, utilisation remains low or newer processors make earlier generations uneconomic faster than expected, the financial correction could be abrupt. Data centres may remain standing while the equipment inside them earns far less than was assumed when the capital was committed.

The accounting treatment is especially revealing. Alphabet reassessed its servers and certain network equipment in January 2023 and extended their estimated useful lives to six years. Microsoft had already moved both server and network equipment from four years to six, effective from its 2023 financial year. Meta followed in January 2025, increasing the estimated life of certain servers and network assets to five and a half years. Alphabet said its change would reduce 2023 depreciation expense by approximately $3.4bn. Meta estimated a reduction of approximately $2.9bn in 2025.

The changes may reflect better maintenance, improved software, more efficient asset management and the reuse of equipment across workloads. They are not evidence of wrongdoing. But the cross-industry pattern exposes a genuine tension: hyperscalers are extending the accounting lives of computing assets at the same time that technical performance is improving at extraordinary speed.

The economics of the AI build-out therefore depend partly on two competing propositions being true at once. The equipment must remain useful for longer, while new equipment becomes much more capable.

If those assumptions fail to coexist, the consequences could include overcapacity, asset impairments, failed infrastructure providers and severe losses for investors. None of this should be minimised.

Yet it would still demonstrate that capital markets financed too much infrastructure, at the wrong price or too early. That is not the same as demonstrating that the underlying technology lacks utility.

Detection is solving two different problems

The anxiety surrounding AI-generated content provides another example of categories being confused.

Schools want to determine whether students used generative tools. Publishers want to identify synthetic submissions. Employers want to know whether applicants wrote their own materials. Software vendors promise to recognise whether an article, report or email originated with a machine.

There are circumstances in which provenance is essential. A fabricated video of a political leader, forged authorship, impersonation, academic fraud and manipulated evidence are not merely questions of quality. They are questions of identity, consent and authenticity. In such cases, asking whether a human created, approved or appeared in the material is entirely legitimate.

But that does not justify treating detection as a general test of ordinary knowledge work.

Imagine that it is 1998 and Microsoft Word is replacing the typewriter. A company launches software capable of detecting whether a document was produced on a computer rather than typed mechanically.

As part of a fraud investigation, the question might occasionally matter. As a universal measure of whether the document was good, it would be absurd.

The same distinction applies now. A quarterly report should be judged on whether its figures are accurate, its reasoning is sound and its conclusions are defensible. An email should be judged on whether it communicates clearly and truthfully. A policy paper should be judged by its evidence. The fact that software helped shape the sentences tells us little about the quality of the underlying thought.

AI does not remove responsibility from the author. It makes responsibility more important.

A person can produce misinformation without assistance. A model can produce misinformation with extraordinary fluency. A person can also use AI to examine evidence, challenge assumptions and communicate an idea more clearly than they could alone.

The appropriate response is not to pretend the tool was absent. It is to make verification, accountability and provenance proportional to the risk.

The wrong question is not always, “Was AI involved?” The wrong question is asking it when what we actually need to know is, “Is this true, useful and accountable?”

Why this transition feels different

Human beings have spent centuries building machines that reduce physical effort. Steam power displaced muscle. Industrial machinery increased output. Vehicles compressed distance. Robots absorbed repetitive manufacturing tasks.

Artificial intelligence reaches into work people associate with the mind: writing, translation, software development, analysis, planning, design and decision support.

That is why the reaction is more visceral.

A machine lifting a heavy object is understood as a tool. A machine producing a credible paragraph feels like competition. We are comfortable augmenting the body. We are less comfortable augmenting cognition because cognition is entangled with status, expertise and identity.

Yet the boundary has never been as clean as we pretend. Calculators changed arithmetic. Spreadsheets changed financial analysis. Search engines changed research. Navigation systems changed spatial memory. Computers did not merely help people type faster. They changed what could be created, measured, stored and coordinated by one person.

AI continues that development at greater breadth. It can interpret, generate and compare across domains rather than merely executing a predefined sequence. That makes it unusually consequential, but it does not make it supernatural.

It remains technology: fallible, unevenly distributed, shaped by incentives and dependent on human institutions.

Software should adapt to people

For decades, organisations have purchased software and then redesigned human behaviour around its limitations. Employees attend courses to learn where functions are hidden. Teams alter their processes to fit rigid systems. Specialists connect applications that were never designed to work together. People repeatedly enter the same information because the systems around them cannot share context.

This arrangement became so familiar that it stopped appearing unusual.

AI offers the possibility of changing that relationship. Software can increasingly interpret natural language, retain relevant context and translate an intended outcome into operations across several systems. Instead of requiring users to understand the software’s internal structure, the interface can begin with what they are trying to accomplish.

The transition will not eliminate the need for training, expertise or disciplined processes. Nor will every interface become conversational. But it may shift where human attention is most valuable.

Less time may be spent learning how a machine expects a task to be expressed. More can be spent defining the objective, evaluating the result and deciding what should happen next.

That is not a manifesto against software. It is a reasonable test of whether software is improving: does the system reduce the effort required to achieve an objective, or merely transfer its complexity to the user?

From product to infrastructure

The most important shift may be conceptual.

AI is still discussed as though it were a discrete product category: an assistant, a chatbot, a writing tool, a model or a feature. New technologies often arrive as visible objects. The most consequential ones eventually disappear into infrastructure.

Companies no longer advertise that their offices use electricity. Websites do not boast that they are connected to the internet. Businesses rarely describe databases as an innovation. These technologies remain essential precisely because they have become ordinary.

AI is moving in the same direction.

Models will be embedded in accounting systems, industrial equipment, customer service, logistics, medicine, education, engineering and public administration. In many cases, users will neither know nor care which model performed the work, just as few people know which database processed their last payment.

Eventually, the phrase “AI company” may become as uninformative as “internet company”. The relevant distinctions will be how well a business applies intelligence, what data and processes it controls, what decisions remain human, and whether the resulting system creates measurable value.

That future does not require every current AI company to survive. It does not require today’s market leaders to remain leaders. It certainly does not require every investor to have paid the right price.

The argument we should be having

The present cycle contains plenty that deserves to collapse: products without customers, companies without defensible advantages, wasteful implementations, dubious benchmarks, industrial quantities of synthetic content produced without judgment, and infrastructure projects based on heroic assumptions about demand.

A financial correction could force the market to separate genuine utility from theatre. The disappearance of easy capital might even accelerate the transition from impressive demonstrations to systems that solve narrow, expensive and persistent problems.

But the serious debate is no longer whether AI will remain part of modern life. Adoption has already moved too far, and the potential applications are too broad, for that to be the most useful question.

The serious questions concern control, concentration, accountability and distribution.

Who owns the infrastructure? Who controls the models? Whose data improves them? Which decisions may be delegated, and which must remain human? Who receives the productivity gains? What happens to people whose work changes faster than institutions can adapt?

Can AI help address problems in climate, health, energy, education and public administration without simultaneously amplifying surveillance, disinformation and inequality?

These are not arguments against the technology. They are arguments about the society and economy constructed around it.

Artificial intelligence has the potential to help address problems human beings created and have repeatedly failed to solve. It will not do so automatically. Intelligence without governance can increase capability without improving judgment. Efficiency without direction can help us make the wrong decisions faster.

AI companies will fail. Valuations will fall. Investors will lose money. Some infrastructure will prove uneconomic, and some promised productivity gains will never materialise.

None of that would establish that AI was temporary. It would establish that technological progress does not protect companies from poor economics, organisations from poor execution or investors from paying the wrong price.

Humanity rarely abandons a technology that meaningfully expands what people can accomplish. The technology becomes cheaper, more reliable and less visible until, eventually, it stops being treated as a separate category.

One day, we may no longer speak about AI-generated software, AI-supported work or AI-enabled companies.

We will simply call it software.

And the supposed collapse of artificial intelligence will be remembered for what it really was: the moment when speculation met reality, and the technology began the slower, harder work of becoming useful.

Sources and notes

  1. Stanford Institute for Human-Centered Artificial Intelligence, The 2026 AI Index Report.
  2. MIT NANDA, The GenAI Divide: State of AI in Business 2025.
  3. Emanuele Del Sozzo, Martin Fleming, Kenneth Flamm and Neil Thompson, How Much Progress Has There Been in NVIDIA Datacenter GPUs?, 2026.
  4. Alphabet, Microsoft and Meta public filings concerning revised useful-life assumptions for servers and network equipment.