The Profit Is in AI's Underground

While the surface enchants, infrastructure charges rent.

In March 2022, AI can still be discussed as promise, tool, model, research, or laboratory. The market begins to feel that something large is there, but still tries to fit everything into familiar categories. Intelligent applications. Software with automation. Better chatbots. Generated images. Assisted code. All of this matters. But the profitable thesis may be somewhere else: underground.

AI's underground is made of GPU, server, network, energy, cooling, memory, storage, data center, connectivity, and infrastructure software. It is ugly, expensive, physical, and indispensable. While investors search for the winning application, infrastructure suppliers receive orders. While executives discuss strategy, engineers ask for capacity. While users are enchanted by answers, racks consume energy.

Nvidia and AMD are the obvious names. Nvidia because its acceleration ecosystem has become the center of the race. AMD because demand for alternatives and competition always appears when one supplier becomes too dominant. But Vertiv, Arista, Broadcom, and Supermicro may explain the thesis better. They show that AI is not only a model. It is a chain of constraints.

Vertiv captures energy and cooling. Arista captures data center networking. Broadcom captures silicon and connectivity. Supermicro captures rapid server integration. With every model advance, the physical stack must respond. The common investor asks which app will have the most users. The infrastructure investor asks which supplier has backlog before the user pays.

Perhaps in 2023 the market wakes up violently. Perhaps guidance from a leading GPU company reveals that data center demand exceeded old mental models. When that happens, many will call it a surprise. But the surprise will already have been built. Large models require expensive training. Expensive training requires hardware. Hardware requires data center. Data center requires energy. The chain was visible to anyone who did not confuse product with infrastructure.

The reader's profit lies in looking where inevitability meets limited capacity. GPU can be limited. Optimized server can be limited. Network can be limited. Local energy can be limited. Cooling for high density can be limited. Data center construction can be limited. Companies that solve limits capture margin while euphoria argues over slogans.

The counter-thesis is that underground also becomes a bubble. When everyone discovers infrastructure, any company touching a rack receives a premium. Supermicro can be excellent in the rush phase and problematic in the audit phase. Vertiv can grow and still suffer if capex pauses. Arista can be excellent and expensive. Broadcom can be necessary but not pure. AMD can capture narrative without capturing enough margin. Nvidia can be the best operating asset and a dangerous stock depending on price.

But avoiding the thesis for fear of bubble is also laziness. The correct move is to separate necessity from valuation. The need for infrastructure can be real even if some stock prices become insane. The mature investor does not turn a good thesis into a license to pay any multiple. He also does not turn a high multiple into proof that the thesis is false.

AI will be sold as software because software is easy to explain. But at the beginning of the race, money will go to physical capital. This happened before. Railways before national networks. Towers before broad mobile telephony. Fiber before reliable streaming. Data centers before elastic cloud. Now, accelerated infrastructure before omnipresent AI.

The market likes what appears. Cash likes what is missing.

And what is missing, for now, is underground.

Leo Bentier