The largest technology companies are spending several times what AI currently earns, and the argument over whether that amounts to a bubble has moved from commentators to fund managers. For a company that builds on these services, the more useful question is what happens to its costs, its supply and its products if the economics change, in either direction.
Nothing in this article constitutes investment, financial or legal advice. It is offered to support discussion and further review with qualified advisers.
How large is the gap between AI spending and AI revenue?
In September MIT Technology Review put the hyperscalers’ capital spending at about $750bn for 2026, against total AI revenue of roughly $150bn to $200bn. The same piece cited a Morgan Stanley estimate that more than half of the $2.9tn the hyperscalers are expected to spend between 2025 and 2028 will be financed from outside the companies, through debt, leases and partners. Alphabet’s second quarter results, filed in July, showed negative free cash flow of about $5.9bn, its first deficit since it listed in 2004. Meta has guided to capital expenditure of $130bn to $145bn for 2026.
The revenue needed to justify that spending is a matter of estimate, and the estimates are large. Bain & Company calculated in September 2025 that AI would need around $2tn of new annual revenue by 2030 to fund the computing it requires, and that even on generous assumptions about savings the industry would remain some $800bn short. Such projections depend heavily on their inputs, and none should be read as a forecast. Taken together, though, they describe an industry building capacity well ahead of the income that is meant to pay for it.
Is it a bubble?
Investors are increasingly uneasy. In Bank of America’s global fund manager survey of November 2025, a net 20% of respondents said companies were overinvesting, the first such reading in two decades. By August 2026, 32% named an AI bubble as the largest tail risk to markets, yet 71% did not expect the hyperscalers to cut their spending. A Wharton School economist told MIT Technology Review that a failed boom would be ‘the largest misallocation of capital in history’.
The opposing case rests on history and on usage. Railways in the 1840s and fibre networks in the late 1990s were overbuilt and ruined many of those who financed them, yet the capacity survived and later carried businesses that could not have existed without it. Usage of AI services continues to grow, and the companies spending most heavily are among the most profitable in the world. Nobody can say with confidence whether a correction is coming, or when, and a board that bases its AI plans on a view of the market’s timing is taking a bet it is poorly placed to judge.
Why should a company that only buys AI care?
Most businesses are not investors in data centres. They are customers of the models those data centres run, and the prices they pay today were set by providers competing for market share while capital was plentiful. If the economics shift, that price is the variable most likely to move, and there are several ways it could move.
- Providers under pressure to show a return may raise prices, withdraw discounted tiers, or charge more for the larger models that many products now depend on.
- Smaller providers and resellers that sit between a company and the model may be acquired, change their terms or close, taking integrations and stored data with them.
- Older models are retired on the provider’s timetable, and a product tuned to one model’s behaviour may need rework when it goes.
- Where capacity is short, providers may prioritise their largest customers, leaving smaller ones with slower service or waiting lists.
A correction could equally push prices down, if capacity built for demand that does not arrive is sold off cheaply. That outcome favours buyers who can move between providers and penalises those who cannot. The common thread is lock-in. A company that can switch models in weeks is exposed to the market’s direction only modestly, whereas one that would need a year to switch is exposed to whatever its provider decides.
What does the evidence on returns add?
The revenue the providers need has to come from customers, and customers have so far found returns hard to show. MIT’s NANDA initiative reported in July 2025 that 95% of the organisations it studied were getting no measurable return from generative AI, drawing on 52 interviews, a survey of 153 leaders and more than 300 public initiatives. The sample is modest and the definition of return was strict, so the figure is better treated as a warning than as a census. MIT Technology Review’s September article cited a survey of about 6,000 executives in which roughly 90% reported no productivity gain.
This matters to a buyer for a reason beyond its own business case. If the customers paying for AI cannot show value, the providers’ route to the revenue they need runs through higher prices for those customers who can. A company whose AI use is measured and profitable is better placed to absorb that than one whose use is still justified by expectation.
What is a sensible test for an AI-reliant business?
We argued in September that the durable advantage lies in the learning a company keeps, since the model itself is rented. The same principle gives a short test that a board, an acquirer or an investor can apply to any business whose products or operations depend on a third-party model.
- Does the product or process still make money if the cost of model access doubles, and if it triples?
- How long would it take to move to a different model or provider, and has anyone tried it?
- Does the company hold its own prompts, evaluation sets, fine-tuning data and logs, in a form it could take elsewhere?
- What notice do the contracts give of price changes and model retirement, and what rights do they give to export data?
- What share of revenue-critical work depends on a single provider?
- Is the value of each AI use measured as an outcome, such as cost per resolved case or revenue per adviser, or only as usage?
- What does the product do if the model is unavailable for a week?
A business that can answer these questions well is largely indifferent to how the infrastructure debate resolves. One that cannot is carrying a risk that does not appear in its accounts.
What does this mean for investors in AI-reliant companies?
Valuations of companies built on third-party models often assume that today’s model prices hold or fall. That is a reasonable central case, and it is still an assumption that should be stated and stress-tested in diligence like any other input cost. Gross margin at doubled inference cost, the time and cost of switching providers, and the ownership of evaluation data are all answerable questions. Management teams that have already asked them of themselves tend to be the ones worth backing.
The same applies in reverse to companies whose advantage depends on access to a particular model. Where the differentiation is the model, a competitor can buy the same access next quarter. Where it lies in proprietary data, workflow and measured results, it is likely to survive both a price rise and a price collapse.
Frequently asked questions
Will the price of AI models go up or down?
Nobody knows. Competition and falling hardware costs push prices down, while the providers’ need for revenue pushes the other way, and the balance may differ between frontier models and cheaper ones. The prudent response is to be able to move, so that the direction matters less.
Should we sign long-term committed-spend contracts to fix prices?
Commitments can secure discounts and capacity, but they deepen dependence on one provider and can leave a company paying above market if prices fall. They suit stable, high-volume workloads with measured returns. For anything still being proven, shorter terms and portable architecture are usually the better trade.
Does the bubble debate mean we should slow AI adoption?
It argues for adopting with measurement and portability in place. Uses that show a measurable return at today’s prices, and would still do so at higher ones, are worth pursuing whatever happens to the market.
What single figure should a board ask for?
Gross margin, for each AI-dependent product or process, at twice the current cost of model access. It is quick to calculate and reveals at once whether the business case depends on prices that nobody controls except the provider.
