AI Doesn’t Have to Fail for the Boom to Break

The dangerous question is not whether artificial intelligence will change the world.

It is whether the businesses spending heavily on it will earn enough, soon enough, to justify the bill. Those are different questions. Mixing them up can turn a sensible belief about technology into a very expensive investment.

Imagine a gym owner who notices that everyone suddenly wants to get fit. His existing machines are busy. Membership is growing. So he orders another hundred treadmills, signs a larger lease and borrows money for the renovation.

“People will always need exercise,” he says.

Probably. But the bank has a less philosophical question: when is the next payment arriving?

Now imagine a rival gym cuts membership prices. Some new members stop coming after their free trial. The building’s electrical upgrade takes longer than expected. The treadmills still work beautifully. None of that pays the rent.

That is a useful way to examine the AI boom. We do not need to predict that AI will become useless. We need to understand what could interrupt the journey from exciting technology to dependable cash returns.

First decide which boom you mean

Three things often get bundled into one sentence: AI usage, spending on AI infrastructure, and the prices investors pay for AI-related shares.

They can move in different directions.

More people could use AI while providers charge less per task. Infrastructure spending could slow while existing data centres remain busy. Share prices could fall because growth disappoints expectations, even when sales keep rising.

The first crack may therefore look surprisingly ordinary: a project moves back two quarters, a customer negotiates a lower price, or management becomes more careful about its next expansion.

Nothing has to stop completely. The numbers only have to become less attractive than the numbers investors were counting on.

A useful technology can survive a disappointing investment cycle. Shareholders do not get paid simply for recognising that the technology is useful.

When customers do not pay enough

Usage is encouraging. Paying, renewing customers are more useful when expensive infrastructure needs funding.

The critical question is whether AI services produce enough additional revenue or savings to cover computing costs, implementation, support and the capital needed to deliver them. A popular demonstration is not the same as a repeatable commercial return.

Consider a hypothetical gym with 1,000 members paying $100 monthly. Revenue is $100,000. Suppose monthly operating costs are $70,000, leaving $30,000 before financing, tax and equipment replacement.

A competitor forces the monthly price down to $80. Membership grows 10% to 1,100, but revenue becomes $88,000. Even if operating costs stay unchanged, the remaining amount falls to $18,000, a 40% decline.

More customers. Less money left over. Those numbers are illustrative, not a forecast for any AI company.

For an AI provider, the equivalent danger is growing usage that does not translate into adequate margins and cash. For an enterprise customer, it is an impressive tool that fails to deliver measurable savings after integration and supervision costs.

But the evidence does not support simply declaring that demand is imaginary. In its October 2026 earnings call, Alphabet reported Google Cloud revenue of $30.7 billion and operating income of $11.1 billion. Cloud includes more than AI, so neither figure should be presented as pure AI revenue or profit.

The real test is whether commercial returns keep pace with the investment required, not whether anybody wants to use the products.

When the bill arrives before the revenue

Buying infrastructure and receiving customer cash rarely happen on the same day.

Alphabet reported $51.4 billion of capital expenditure against $42.3 billion of operating cash flow in that quarter, resulting in negative $9.1 billion free cash flow. The company also identified effects from inventory for TPU sales and a $5.2 billion European Commission fine payment. It would be misleading to attribute the entire cash-flow result to AI.

Still, the mechanism matters. Expansion can consume cash even while the underlying business grows.

Accounting adds another layer. A server purchase is not usually charged entirely against earnings immediately; its cost is allocated over its useful life through depreciation. Meanwhile, cash may already have left the business.

Microsoft’s FY2026 filing reported depreciation expense of $34.3 billion, up from $22 billion the previous year. These are company-wide figures, not a separate AI expense account. Its disclosures also describe investing in AI infrastructure before revenue streams are fully developed.

Slowing new purchases would not instantly erase the expense associated with equipment already installed. Nor does reported depreciation tell us precisely when a particular chip becomes commercially obsolete.

The warning sign is a persistent widening gap between commitments and the cash those commitments can reasonably produce. One unusually weak quarter needs explanation, not an automatic funeral announcement.

When progress makes yesterday’s investment less valuable

Here is the awkward possibility: AI could become better and cheaper, and some AI investors could still lose.

The Stanford 2025 AI Index documented a more than 280-fold reduction in inference cost for systems performing at GPT-3.5 level between November 2022 and October 2024. That is a historical, benchmark-specific finding, not a claim that every frontier service’s price is collapsing today.

The economic implication is worth considering. If customers can obtain similar output more cheaply, suppliers may have to lower prices, upgrade equipment or offer more service for the same money.

Customers benefit. A supplier that paid too much for yesterday’s advantage may not.

There is a strong counterargument: cheaper computing can unlock new applications and much greater demand. Lower cost per task could also improve a provider’s margins if selling prices fall more slowly.

Which effect wins depends on pricing power, utilisation and how much extra usage materialises. “More efficient” is not automatically bullish or bearish for every company in the chain.

The gym analogy still works. Better machines that allow more customers per hour can improve the business. Better machines sold cheaply to every rival can also make an expensive refurbishment much harder to recover.

When physical delivery falls behind the presentation

A spending announcement does not connect a data centre to electricity.

Microsoft’s FY2027 first-quarter earnings discussion described third-party capacity deliveries slipping, with some expected GPU capacity moving into the second half of its fiscal year. Management also said demand exceeded supply. A delay can therefore reflect genuine demand meeting a physical constraint, rather than demand disappearing.

Even so, a delay can damage investment returns. Financing, leases or other commitments may continue while the expected revenue arrives later.

The IEA’s 2026 energy and AI analysis identifies electricity and supply-chain bottlenecks as constraints on aggressive near-term expansion. The analysis considers data centres broadly as well as AI-specific demand.

There may be a strange result: scarcity helps an operator with available capacity charge more, while hurting a developer waiting to switch its equipment on. The same constraint can create a winner and a loser.

That is why “AI infrastructure” is too broad a label for an investment conclusion. Ask whether the company owns usable capacity, supplies a bottleneck component, or is still paying to build something that cannot yet operate.

When financing becomes less forgiving

A cash-rich company and a heavily financed specialist do not have the same ability to wait.

CoreWeave’s June 2026 quarterly filing reported $35.6 billion of total indebtedness. Three customers represented 36%, 26% and 10% of quarterly revenue, or 72% combined. The filing does not identify those customers in that disclosure.

That creates a question about resilience, not proof of imminent failure. The company also reported substantial contracted future revenue. Contracts can improve visibility, but future revenue is not the same as cash already collected or profit already earned.

If financing becomes more expensive, a delivery slips and a major customer changes its plans, the pressures can reinforce each other. Selling assets to raise cash could then reduce the prices other owners can obtain for similar equipment.

This is a possible transmission mechanism, not a claim that it is currently unfolding everywhere. Customer concentration also does not make revenue fictitious. Accusations of fraud require evidence, not a diagram with arrows between company logos.

The fragile part of a boom is often the assumption that money, customers and infrastructure will all arrive on schedule.

What would make the concern more convincing

Do not collect frightening headlines and call that analysis. Look for several pieces of evidence pointing in the same direction.

• Are customers renewing and expanding paid use, or mainly consuming discounted trials?
• Are prices and utilisation supporting returns after equipment, energy and financing costs?
• Are projects being delivered and converted into revenue on the promised timetable?
• Are cash generation and funding capacity keeping up with commitments?

A growing backlog alongside stronger cash collections and improving margins would weaken the pessimistic case. Repeated delays, weaker prices and rising funding needs would strengthen it.

For a stock, add one more question: how much disappointment is already reflected in its price? Identifying a genuine business risk does not automatically identify an overpriced share. A dramatic fall may already incorporate a very gloomy outcome.

Back at the gym

The gym owner may be right that more people want to exercise. He may even own excellent machines.

But he still needs members who pay enough, equipment that earns its keep and enough cash to survive the wait. A packed opening day cannot answer all three.

AI’s future may be extraordinary. The price of participating in it still deserves ordinary business arithmetic.