Inside the AI Bubble: The Trillion-Dollar Bet That Must Not Fail

Artificial intelligence may transform the world. That does not mean the companies spending trillions to build it will recover their money.

The AI boom is no longer principally a software story. It has become an industrial construction programme built on debt, scarce hardware and the assumption that future demand will justify almost any amount of spending today.

A technology does not have to be fraudulent for the investment surrounding it to become a bubble. Railways, electricity, telecommunications and the internet all changed civilisation. They also produced manias in which investors built too much, paid too much and discovered too late that a revolutionary technology is not necessarily a profitable investment.

The largest technology companies are now committing sums that appear to outrun the evidence that customers will pay for what is being built. Businesses once prized for being asset-light and cash-rich are becoming owners and long-term tenants of data centres, power systems, cooling equipment, specialised memory and rapidly depreciating processors.

This is no longer principally a software boom. It is an industrial construction boom.

Its logic is defensive. Each company spends because it fears that restraint will surrender the market to a rival. Investment is therefore governed not only by expected returns but by the terror of being left behind. What may appear rational for each company can become ruinous for the industry as a whole.

The financing is becoming more dangerous as the commercial assumptions remain unproved. Technology companies are issuing debt, entering long leases and using special-purpose vehicles and private-credit structures to fund data centres. Some chipmakers and cloud companies are also financing laboratories and infrastructure providers that then spend the money on their investors’ products.

The transactions may be legitimate. But they make demand harder to interpret. Revenue generated by a customer financed by its supplier is not the same as demand produced by an independent, profitable market.

The central question is not whether money is changing hands. It is whether there is a final customer capable of paying enough to support the entire chain.

That remains uncertain. Frontier AI companies must make their products cheaper while generating enough revenue to fund more expensive infrastructure. They must subsidise use today while promising vastly greater demand tomorrow. More usage does not solve the problem if every additional query, image or line of code remains costly to provide.

The strongest answer is that the companies involved are already enormously profitable. That is true. They possess dominant businesses in advertising, cloud computing, retail and software, and they can absorb losses for years. AI also has real uses. It can accelerate coding, research, administration and creative work.

But profitability does not abolish valuation risk. A company can survive while its shareholders are ruined. A useful technology can endure while the capital invested at the height of enthusiasm is destroyed.

The same problem applies to the infrastructure. A data centre is valuable only if it remains occupied, powered and commercially relevant. AI processors may function for years while losing competitive value within a few product cycles. Yet the assets and leases financing them may last far longer.

If the equipment becomes obsolete faster than the accounting assumptions suggest, the consequences will arrive through weaker margins, asset write-downs and cancelled construction.

The wider danger is concentration. A small group of AI-linked companies now occupies an extraordinary share of the American stock market. Millions of investors believe they own diversified index funds, yet a disproportionate part of their savings depends on the same companies, the same suppliers and the same investment thesis.

A correction would therefore not remain confined to speculative laboratories. It could reduce chip orders, data-centre construction, memory prices, private-credit valuations and the expected earnings of the largest companies in the market at the same time.

The technology industry would not disappear. The largest companies would probably survive. Useful models would remain. Data centres would continue operating. But survival is not the same as vindication.

Partially completed facilities would be cancelled or sold. Processors would enter the secondary market at lower prices. Suppliers would write down inventory. Highly leveraged cloud providers would fail or be absorbed. AI laboratories would be acquired by their principal financiers. Venture capital would contract as investors discovered that scale was not the same as a viable business.

The collapse would, paradoxically, make AI cheaper.

That is what bubbles often do. They mobilise immense capital, create excess capacity and then destroy the owners of that capacity. Railways remained after railway investors were ruined. Fibre remained after telecommunications companies failed. The internet remained after the dot-com crash.

AI will remain after an AI crash.

The question is not whether artificial intelligence is real. It is whether its economic value will emerge quickly enough, and accrue to the right companies in sufficient amounts, to justify the money already committed in its name.

The industry’s defenders argue that so much capital cannot be wrong. History suggests the opposite. Large sums do not prove that an investment thesis is sound. They may show only that institutions have become too frightened to stand apart from it.

Chief executives cannot slow spending without appearing to concede defeat. Investors cannot abandon the trade because the companies driving it dominate the indices against which they are judged. Suppliers expand because customers promise extraordinary demand. Lenders continue financing the buildout because the borrowers remain among the strongest corporations in the world.

Each participant may be behaving rationally. Together, they may be building far more than the economy can profitably use.

The likely reckoning will not produce a world without artificial intelligence. It will produce a world forced to discover what artificial intelligence is actually worth.

Useful models will survive. Useful infrastructure will remain. AI will become part of the economy rather than an excuse for suspending economic judgment.

The technology may change the world. The bubble is the belief that this guarantees today’s investors will profit from it.