Is AI a bubble?

The question is usually asked as a yes or no, and the numbers do not answer it that way. Two things are true at the same time: the capability is real and in daily use, and the financing structure behind its build-out carries risks that do not depend on the technology working.

What the spending numbers say

Capital expenditure is where the bubble argument gets its force.

Analyst estimates put combined 2026 capex for the five largest hyperscalers between roughly $700 billion and $900 billion, about 36 percent above 2025. Company guidance during 2026 tracked that: Amazon near $220 billion, Alphabet at $175 to $185 billion, Microsoft above $120 billion for its fiscal year, and Meta raising its range toward $125 to $145 billion, citing higher memory prices and added data center costs.

Commitments run well past annual budgets. Meta disclosed $279 billion in future data center leases, and formed a $14 billion venture with BlackRock for a single El Paso campus. Those are obligations that persist whether or not the demand arrives.

The financing has also moved off balance sheets. Hyperscalers began tapping external financing as capex outran operating cash flow, which is a structural change from a build-out funded entirely from profits. It is the single most reliable historical marker of a capital cycle turning fragile, because debt service does not adjust when revenue disappoints.

What the return numbers say

This is the weaker half of the picture for the bulls.

The most cited evidence is a July 2025 MIT report, The GenAI Divide: State of AI in Business, which examined 300 public deployments alongside 150 leader interviews and a survey of 350 employees. It concluded that despite $30 to $40 billion in enterprise generative AI spending, roughly 95 percent of organizations were seeing no measurable profit and loss impact, with about 5 percent capturing most of the value.

That number deserves the caveats it has attracted. The report defines success narrowly, as measurable return within roughly six months, the interview base is small, and it was not peer reviewed. The criticism goes to precision rather than direction, and its central finding, that most organizations never baselined the workflow they were automating, is not disturbed by the methodological complaints.

Independent signals point the same way. In August 2026 a survey found 49 percent of enterprises had scaled back AI agent deployments over cost. Canva cut its growth target because AI compute costs rose faster than the revenue attached to them. Amazon reportedly burned $1.8 million on a single Claude project. GPT-5.6 Sol ran a real business for a day and burned through the cash.

Against that, usage is not in doubt. Gemini reached 950 million users. The infrastructure suppliers are selling everything they can make, to the point that AI buyers absorbed the DRAM supply and pushed DDR5 prices up 500 percent, which in turn raised consumer device prices. That is demand, whatever its eventual profitability.

Where the bubble comparison holds and where it breaks

The dot-com analogy gets reached for automatically and only half fits.

It holds on the spending-to-revenue gap and on circular financing. Money moving from a chip vendor into a startup that then buys that vendor’s chips books as revenue on one side and investment on the other, which flatters both until the capital stops. That pattern was present in 1999 and is present now.

It breaks on what the money buys. Dot-com capital went heavily into customer acquisition, which vanishes, and into fiber that sat dark for a decade. AI capital goes into data centers, power interconnections and chips that are running at capacity today rather than waiting for demand to show up. The closer analogy is a railway or telecom build-out: the assets outlive the companies that financed them, the capacity eventually gets used, and the equity holders who paid for it are frequently wiped out along the way.

That is the distinction worth holding. “Is AI a bubble” and “will AI be useful” are separate questions, and the answer to the second one being yes has never protected anyone from the first.

What to watch

Four indicators, in roughly the order they would show up.

Capex guidance direction. Raising guidance is the current pattern. The first quarter in which several hyperscalers cut rather than raise is the signal, and it will arrive before any public acknowledgment.

GPU rental prices. They reflect the balance of supply and demand for compute more honestly than any company statement. Sustained falling rental prices mean supply caught up.

The circular deals. Whether the arrangements in which AI companies fund their own customers keep expanding or start unwinding tells you how much of the revenue is genuinely external.

Enterprise renewals. Pilots are cheap and easy to start. The second paid year is where the MIT finding either reverses or hardens.

One more marker is now on the calendar. Anthropic expects to raise roughly $75 billion at its debut, $86.2 billion including the overallotment, in what would match or top SpaceX’s record. A public listing at that scale forces quarterly disclosure of the unit economics that have so far been private, and the numbers behind the argument stop being estimates.

Related coverage

• Anthropic expects an IPO matching or topping SpaceX’s record, the listing that will make the numbers public.
• Microsoft and Meta earnings deliver Wall Street’s AI capex verdict, how the market reads the spending.
• Meta disclosed $279 billion in future data center leases, commitments beyond the annual budget.
• 49 percent of enterprises scaled back AI agents over costs, the demand side wobbling.
• Canva cut its growth target as AI compute costs rose, what happens when usage outruns margin.
• Amazon reportedly burned $1.8 million on a single Claude project, the cost of one enterprise experiment.
• AI buyers took the DRAM supply and DDR5 is up 500 percent, demand showing up in component prices.
• How much electricity does AI actually use?, the physical constraint on the build-out.