The Wrong Question: Is AI a Bubble?

How to Position for the Inevitable Volatility. A framework for investing in AI.

Everyone is still arguing about AI as if it has to be either a bubble or not a bubble, as if reality owes us a clean binary. There are very smart people on both sides.

And yes, one thing seems obvious: AI will be massively transformative, and the use cases will go well beyond what any of us can list today. But that statement is exactly what people said about the internet in 2000, and it was correct then too. The internet was transformative and also in a bubble at the same time.

So the better question is not whether AI matters. It does. The better question is where to position yourself within the inevitable volatility.

Some companies say they cannot build fast enough. Researchers warn about a bullwhip effect across different timescales: capital gets committed quickly, GPU orders pile up quickly, hype spreads instantly, but electricity, grid connections, transformers, permits, and data centers do not move on internet time. When the system has mismatched clocks, you do not get a smooth S-curve. You get volatility and surprise. Short sellers point out the discrepancy between Capex and earnings.

A Simple Example of Embedded Volatility: NVIDIA

Look at NVIDIA as a mental model for how volatility can be structural rather than accidental. The story sounds great until you look at who is paying, and why. If you can sell silicon at 75 percent gross margin, the obvious question is whether that margin is actually durable when a meaningful share of demand is coming from customers that are, to put it politely, not exactly printing money.

If one of the biggest customers is a loss-making company that has announced $1.4 trillion of investment in data centers (with GPUs as the biggest cost factor) while having virtually no revenue relative to that $1.4 trillion and no proven business model, then the supply chain is not anchored on stable end-demand. It is anchored on funding cycles, narrative cycles, and competitive panic.

And there is substitution. The Googles, Alibabas and Amazons of the world do not enjoy paying 75 percent gross margins to anyone. They build their own chips, they optimize around their own workloads, and they use their distribution to make their economics work.

None of this says NVIDIA is a bad company. It says the ecosystem can swing from shortage to digestion faster than investors want to admit, especially when power and deployment bottlenecks can delay real consumption even if the orders were placed.

Three Ways to Invest in a Technology Wave

Historically, every major technology upgrade gives you three broad ways to invest.

One is to build the core infrastructure itself: the CapEx-heavy backbone, the fiber cable, the railroad. In AI terms, that is the model layer and the training-and-inference buildout that burns capital to create capability.

The second is the pick-and-shovel trade: you sell the tools, inputs, and enabling infrastructure to the builders, like chips, cooling, power equipment, and data-center construction.

The third is the application layer: once built out you use the infrastructure cheaply to build products and business models that were not feasible before.

The pattern is not guaranteed, but it has a rhyme. The capex-heavy builders often struggle because returns get competed away and cycles are brutal. The pick-and-shovel players often win early because everyone rushes to build at once, then growth normalizes when the buildout saturates. Once the technology is ready the application layer tends to capture the durable value because it turns cheap infrastructure into new profit pools.

Caveat: AI might differ in one important way