Is there an AI bubble

The question is usually put wrongly. "Is AI a bubble" sounds like a question about whether the technology works. It is not that.

In economics, a bubble means a situation where more money goes into something than can be got back out of it in the next few years. It is a claim about the gap between investment and returns, not a claim that a model cannot write a summary. Railways were a bubble in the 1840s and fibre optics were one in 2000. Both collapsed as investments and stayed in the ground, where they are still in use.

This piece has two sets of numbers. The first says how much money goes in. The second says how much has so far come out.

What goes in

The combined investment of Microsoft, Alphabet, Meta and Amazon is expected to reach 760 billion dollars this year. In 2025 the same figure was 413 billion, so that is 84 per cent growth in one year.

The number is not stable but rising. Alongside second-quarter results published at the end of July, Amazon, Alphabet and Meta each raised their own estimate. Individual company figures are worth reading with care, because Microsoft's financial year does not follow the calendar year, which is how different sources end up with different numbers for the same company.

The money does not go into models but into concrete, copper and silicon. It is data centres, accelerators, grid connections and cooling, precisely the bottom layer of the stack that is nowhere visible to a user. Part of the growth is also not new capacity but chips getting more expensive.

What has come out so far

There are two answers, and they point in different directions.

On the seller's side, growth is fast. In second-quarter results, Amazon's cloud business grew 37 per cent year on year, its fastest pace in 18 quarters. Microsoft's AI business grew 123 per cent and Alphabet's cloud business 82 per cent. Demand exists, and it is paying its bills.

On the buyer's side, the result is more modest. The GenAI Divide: State of AI in Business 2025, from MIT's NANDA project, went through 300 publicly disclosed enterprise deployments of generative AI and interviewed and surveyed managers and employees. The conclusion: after some 30 to 40 billion dollars of spending, around 95 per cent of organisations had no measurable effect on their results.

That figure circulates online as "95 per cent of AI projects fail", which is not the same claim. What was measured was the effect on results, not whether the tool was useful to anybody. A hundred-person organisation can have a hundred people whose work got easier without a single line of the income statement moving.

Why both sets of numbers can be right

Four explanations, none of which excludes the others.

A pilot is not production. A pilot is run on the best available material with the most enthusiastic team. Production brings permissions, integrations, exceptions and the users who did not want this.

The benefit shows up in the wrong place. If ten people save an hour a week, it appears in no metric unless somebody decides to do something else with that hour. A saving nobody collects is not a saving but slack.

Measuring is genuinely hard. Without a baseline you cannot demonstrate a change. Few measured anything before starting.

Investment comes first and returns come later. A data centre is built in years and depreciated over ten. Comparing it against one year of returns is the wrong sum. That does not, however, mean any level of investment is justified.

What would burst it

The decisive thing is not whether the models are good. The decisive thing is whether there is enough paying demand for the capacity already ordered. Three things are worth watching:

1. Cloud business growth. Built capacity is either used or it is not, and that shows up in cloud companies' revenue faster than in any forecast.
2. Electricity and grid connections. Capacity cannot be brought into use without power, and the connection queues are long. This slows the boom down, but it also stops everyone building everything at once.
3. The price of compute. If the price falls faster than usage grows, the returns will not cover it. If the price holds and usage grows, the investment paid off.

What an ordinary organisation should conclude

Almost nothing directly. Your decision to buy or not to buy does not change because of how much Amazon invests. Three questions are more useful than any macro debate:

1. What does this cost us per month based on usage, rather than as a one-off purchase?
2. Which number is this supposed to move, and what is that number now?
3. What happens if the supplier's price rises materially next year?

The third is the one the bubble debate makes topical. The scale of the investment and the returns accumulated so far are not in balance, and at some point one of them moves. When signing a contract it is worth knowing which way your own bill flexes then.

What to take away

In one sentence

The question is not whether AI works but whether it returns, in the next few years, what is being poured into it now, and those two questions can get different answers.

Disclosure

I work in the AI field and have financial interests connected to a company in it, set out on the disclaimer page. This piece deals only with figures published by publicly listed companies themselves, and it takes no position on any individual company's prospects. It is not investment advice.