The first phase of the AI trade was simple: buy the chip designer. That trade is crowded, expensive and largely understood. The more interesting question in the second half of 2026 is where the constraint moves next — because in every buildout, the bottleneck migrates, and the money follows the bottleneck. Right now the constraint is not compute design. It's the three things that let compute actually run: power, memory and cooling.
Follow the bottleneck, not the headline
A data center full of the fastest accelerators is useless without three inputs the market spent 2024 ignoring: enough electricity to run them, enough high-bandwidth memory to feed them, and enough cooling to keep them from throttling. As accelerator density has exploded, each of these has become a genuine physical limit. That is where pricing power now lives.
| Bottleneck | Why it binds | Names in the conversation |
|---|---|---|
| Power | AI data centers are electricity hogs; grid and generation are the hard limit | NextEra, Williams, Cameco |
| Memory | Accelerators are starved for high-bandwidth memory | Micron, TSMC, Credo |
| Cooling | Chip density makes air cooling insufficient; liquid is required | Vertiv |
Power: the constraint nobody can code around
You can design a faster chip in eighteen months. You cannot build a power plant or upgrade a regional grid on that timeline. The electricity draw of AI data centers has turned power into the tightest constraint in the whole stack — and one that money can't simply accelerate. That's why utilities and independent power names like NextEra, pipeline operators like Williams, and even uranium via Cameco have entered a conversation that used to be purely a technology story. When the incremental gigawatt is the scarce input, the owners of generation get pricing power they haven't had in a generation.
Memory: starved accelerators
Raw compute has outrun the memory that feeds it. High-bandwidth memory has become a gating item — you can have the fastest processor in the world and still bottleneck on getting data in and out of it. That structural shortage is why Micron has become a core AI name rather than a commodity-cycle afterthought, and why the connectivity layer — names like Credo — matters more than its size suggests. TSMC's advanced packaging sits in this bucket too, stitching memory and logic together.
Cooling: the unglamorous winner
The least discussed and most physical constraint. As chips pack more transistors into the same footprint, air cooling stops being enough. Liquid cooling stops being an option and becomes a requirement. Vertiv is the name most cited in advanced liquid-cooling solutions — deeply unglamorous plumbing and thermal engineering that happens to be non-optional for the next generation of density.
What the earnings math says
Here is the fact that frames the whole trade. Wall Street expects roughly 22% year-over-year EPS growth for the S&P 500 this season, and by one estimate AI infrastructure stocks are expected to contribute nearly 60% of that growth — with Micron and Nvidia alone accounting for more than 40% of it. Read that twice. A majority of the index's earnings growth is expected to come from the AI buildout, and a huge slice of that from just the compute-and-memory layer.
That is the opportunity and the risk in one sentence. If the buildout keeps compounding, these names carry the index. If it stumbles, there is very little else holding the earnings math up.
My take
I don't chase the most obvious AI name on its best day. I'd rather own the constraint. The chip-design trade is well understood and priced for it; the bottleneck names — power, memory, cooling — are where the marginal demand is now hitting a physical wall, and physical walls are where pricing power lives. The discipline is to distinguish a genuine structural bottleneck (power, which can't be built quickly) from a cyclical one (memory, which the industry has historically over- and under-supplied in brutal cycles). Own the former through the cycle; trade the latter with your eyes open.
The concentration cuts both ways: when ~60% of the index's earnings growth rides on one theme, you are not diversified just because you own ten tickers. Size accordingly.
Bottom line: the AI trade has moved past the chip and into the plumbing — power, memory and cooling are the constraints now, and that's where the next leg of demand is hitting a wall you can invest in. Just remember how much of the market's entire earnings story is riding on it.
Not investment advice.