Is AI a bubble? It depends where you look

Every week someone asks me: “Is AI a bubble?”. I get the question, but I think it’s badly framed. There is no single “AI industry”: there are companies already earning a lot, others living on promises, and investments that only hold if everything goes right. Saying “it’s a bubble” or “it isn’t” for all of them together is misleading either way.

Where the numbers hold up

Let’s start with what a bubble doesn’t have: real revenue growing fast.

• Anthropic reported $16.3 billion in revenue for the first half of 2026, up from $1.1 billion a year earlier. Of the $80 billion-plus loss that made headlines, about $72 billion is an accounting charge, not cash going out.
• SpaceX, which controls Grok, reported $7.56 billion in AI revenue in a single quarter, nearly triple the previous one.
• Use is mass and daily: hundreds of millions of people come back every week to ChatGPT, Claude, Gemini and Grok, and companies pay for subscriptions and APIs because they save hours of work.

And costs are collapsing: according to Artificial Analysis, GPT-6.1 Sol solves one of their benchmark tasks for 72 cents, a few points behind the most capable model. A technology that becomes more capable and cheaper at the same time doesn’t disappear: it spreads.

Where I see fragility

Some areas of the industry, though, really do remind me of past bubbles.

Debt-financed investment. While the giants paid for data centres out of their profits, the risk stayed in their own coffers. Now a growing share is financed with debt, often off the balance sheet. If expected revenue is late, the problem shifts to the lenders.

Money going round in circles. The chipmaker invests in the lab, which uses that money to buy chips from the chipmaker, which books revenue. The cloud provider funds the customer that buys its cloud. Not necessarily wrong, but it inflates the numbers.

Huge commitments on the future. Labs have signed contracts for hundreds of billions of compute over the next ten years. They hold if growth continues at this pace; if it slows, someone will be left with half-empty data centres.

“Thin” companies. Many startups resell someone else’s model with an interface on top. Every time the big labs add a feature, some of them lose their reason to exist.

So?

I think we’ll see corrections, even hard ones, in some of these areas: some startups will close, some investors will lose a lot, some data-centre projects will be postponed. But I don’t expect the whole industry to burst, because the revenue of those at the centre is real and growing.

The useful precedent is the internet: in 2000 the dot-com bubble burst, but companies with a real product survived and the internet became the infrastructure of everything. The difference with AI is that the main players’ revenue is arriving much earlier.

What it means for people who use AI

1. Don’t tie yourself to one provider. Keep data, instructions and skills portable, and know an alternative for every important tool.
2. Expect prices to change. Some services are sold below cost to win users; promotions end. Plan as if costs could also rise.
3. Invest in skills. Knowing how to make agents work well will stay useful whatever happens on the stock market.
4. A correction might even help you. If part of the compute built today ends up in surplus, AI will get even cheaper for its users.

In one line

There isn’t a single bubble: there’s a solid centre with real revenue and a fragile periphery made of debt and promises. People who use AI to work better don’t need to guess the next crash, just avoid depending on a single horse.

This is a personal opinion, not investment advice.

Sources

• Reuters — Anthropic, first-half 2026 results
• SEC — SpaceX quarterly report (AI revenue)
• Artificial Analysis — prices and cost per task

Informational content, current as of the date shown. Prices, models and rules change often: always check official sources. Written with the support of AI tools and reviewed by the author.