Did Nvidia Just Prove There Is No AI Bubble?

Take a wild guess.

Photo by Ryan Braxton on Unsplash

“There is a lot of talk about an AI bubble,” Nvidia CEO Jensen Huang said during Nvidia’s recent Q3 earnings call. “We see something different.” It’s easy to see why Huang might think that. After all, this earnings call was basically a litmus test for the AI industry, which they aced. Their revenue for the quarter jumped 62% from last year to $57 billion, higher than Wall Street’s predictions. I’m sure from atop Huang’s mountain of gold, everything seems fine. Indeed, it seems like everyone has agreed that fears of the AI bubble catastrophically bursting are unfounded. But I am calling bullshit, because this is no mountain of gold — it is a tower of cards.

What everyone seems to be missing is that demand for trowels or spades doesn’t mean tulip mania or the gold rush bubble isn’t about to pop. Likewise, demand for the AI chips Nvidia sells doesn’t mean there isn’t an AI bubble that is about to ruinously collapse.

Case closed. But everyone also seems to be missing the damning context of this earnings report. In a similar way to pulling one thread and having your jumper fall apart, when you follow the money here, the horrifying extent of the AI bubble is laid bare.

Strap in — this is a long one!

The Fragile Cascade Pt. 1 — Data Centres

90% of Nvidia’s Q3 revenue came from AI data centre chip purchases. This shouldn’t be surprising as the price of data centre compute time is at an all-time high due to skyrocketing demand from AI. Data centre operators, such as Microsoft, Amazon, Oracle and CoreWave, are scrambling to increase supply to meet this demand and are buying millions of Nvidia chips to do so.

With demand so dramatically outstripping supply, you’d think these data centres would be insanely profitable. In fact, arguably, they should be at their peak profitability right now. But, no. They are losing money hand over fist.

Praetorian Capital CIO Harris Kupperman recently revealed that the AI data centres being built today will incur $40 billion in annual depreciation while generating somewhere between $15 billion and $20 billion in revenue. So, before we even take into account operational costs like maintenance and energy, AI data centres are losing more money than they make in revenue!

This is supported by a recent report, which found that the AI industry will need to generate $2 trillion in annual revenue by 2030 to break even on the data centres they plan to build by 2030. Even the rosiest, most blindly optimistic predictions estimate the AI industry falling $800 billion a year short of that, meaning these data centres are unprofitable on a scale never seen before.

But, believe it or not, this problem could actually be even worse! Michael Burry, of “The Big Short” fame, has accused data centre operators of artificially boosting earnings by underreporting depreciation of the chips. If true, then these predictions are far, far too optimistic.

Let’s also not forget that as more data centres are built, supply will begin to meet demand, reducing prices and making data centres even less profitable.

Now, is 90% of your revenue coming from an unsustainable industry that is miles away from profitability a good thing? No, not at all! That alone should utterly terrify investors!

The Fragile Cascade Pt. 2 — AI Operators

Okay, but what about the AI operators like OpenAI? Data centres being so unprofitable isn’t a problem if they can cover these losses with their profits. That way, Nvidia’s customers don’t disappear into bankruptcy.

That could work, if not for the fact that AI operators are even less profitable than data centre operators.