The AI bubble question, and the numbers that would answer it

Asking whether AI is a bubble is usually a way of not saying anything. So here’s a version of the question with an answer attached: what specifically would have to break, and what would you see first?

Three numbers from the past fortnight are enough to frame it.

Signal	Figure	What it rests on
Databricks valuation	$190B, up 42% in six months	Revenue run-rate above $7B, growth above 80%
Nvidia financing plan	Up to $500B	Non-binding MOUs, collateral guaranteed by Nvidia
Training compute growth	5x per year since 2020	Spending that has to be funded every year

Databricks is the cleanest case

Databricks is the cleanest case, because the underlying business is real. Run-rate revenue passed $7 billion and second-quarter growth was above 80%, so this is not a company with no customers.

But the valuation rose 42% in six months while revenue grew 80% annually, which means the multiple is expanding roughly in line with the business rather than being rescued by it. At $190 billion that’s about 27 times run-rate.

That multiple is defensible while growth holds. It is not defensible at 30% growth, and nothing in the current numbers tells you which of those two futures you’re buying.

A bubble is not high prices. It is prices that require a specific future, held by people who have stopped checking whether that future is arriving.

The Nvidia structure is the fragile one

The Nvidia structure is where it gets more interesting, and more fragile. Six private capital firms signed non-binding agreements toward up to $500 billion of data centre financing, and Nvidia made it work by guaranteeing the collateral value of its own chips.

Read that twice. The company selling the asset is underwriting what the asset will be worth at resale, against hardware whose obsolescence schedule it also sets.

Goldman Sachs puts usable GPU lifespans at four to six years, and CNBC has flagged China exposure as a separate risk to the same plan. So the guarantee is a bet that demand stays strong enough that nobody calls on it, which is exactly the kind of arrangement that works until it is tested and then does not.

The case against calling it an AI bubble

Now the case against calling any of this a bubble, given its best hearing. The revenue is real, the customers are paying, and Epoch AI reports pre-training compute efficiency improving around 3x per year, which means capability keeps rising even if spending flattens.

Dot-com comparisons mostly fail on that point. Those companies had traffic and no revenue. These ones have revenue, margins under pressure, and infrastructure costs that are visible on a balance sheet.

Three things that would signal the AI bubble turning

So what would actually signal a turn? Watch for a down round at a company with genuine revenue growth, because that separates a repricing from a business failure. Watch whether the Nvidia MOUs convert into binding commitments or quietly expire, since non-binding is doing real work in that headline.

And watch whether Epoch’s 5x annual compute growth breaks and stays broken for two years, because that would be the industry answering the affordability question itself. CNBC noted the Databricks round was the second this year, and cadence like that is itself a data point.

Until one of those three moves, “bubble” is a mood rather than a finding. The numbers are stretched and the businesses are real, and both of those can stay true for a long time.