The AI build-out is a power story, not just a chip story

By Thomas | financial enthusiast

---

For two years my AI watchlist looked like everybody else's. Nvidia, a couple of hyperscalers, done. Then I read a few data center earnings calls in a row and one word kept coming up. Not "GPU". Power.

The bottleneck moved

A modern AI data center is basically a giant electricity bill with a roof on it. Training and running large models needs racks that draw many times what a classic server rack did. The chips are only useful if somebody can feed them energy and get the heat back out.

So the question I now ask about any AI headline is simple: what does this need physically? Electricity, grid connections, transformers, cooling, memory. Every one of those is a business.

Three things I watch

Memory. Large models are hungry for high bandwidth memory. That is why I keep Micron on my radar (you can follow my bot on it in the Lab). Memory is a cyclical business, and that matters. Great years are followed by ugly ones.

Energy and grid. Utilities and grid equipment makers suddenly get questions from investors that used to be reserved for tech companies. Long planning cycles, regulated returns, slow but steady.

The buyers. Alphabet, Microsoft, Amazon and Meta spend enormous sums on capex. As long as they keep spending, the suppliers do well. The moment they slow down, the whole chain feels it.

What I am careful about

Picks and shovels sound safe. They are not automatically safe. If AI demand disappoints, the shovel sellers built capacity for a gold rush that did not show up. I have seen that movie in other cycles.

My own approach: no single bet on "the winner". A broad ETF as the base, and only a small part in individual names where I understand what physically gets sold.

What about you: are you invested in the chips themselves, or in the stuff around them?