The 7 Bottlenecks of the AI Supply Chain: Where Pricing Power Actually Forms

Not every layer creates equal pressure at the same time. Pricing power shows up where things get tight. Right now there are exactly seven points where that tension is forming.

Memorizing 12 layers is the starting point, not the finish line. The harder question is where the pressure is actually building. Demand outpacing supply, capacity backing up, no substitutes available — that's where margin and pricing power live. By my read, there are seven such points right now.

Bottleneck 1 — Compute

Nvidia, AMD, and Broadcom all sit here. Demand is still outpacing supply on the leading edge. This layer has had leverage and it's holding it. It's the most visible layer, so it has run the most — but it's also where pricing power stays clearest.

Bottleneck 2 — Memory

SK Hynix, Micron, Samsung. Only a very small group can actually make HBM at scale. The point is simple: a GPU is only as fast as the memory feeding it. Even the most powerful processor underperforms if it can't be fed data fast enough. That constraint is absolutely real, and the once-boring corner of semis now sits at the center of the cycle.

Bottleneck 3 — Advanced packaging

TSMC's CoWoS capacity is one of the most constrained points in the AI supply chain right now. Amkor and ASE are in this space too. The logic is brutally simple: you can't ship a chip you can't package. Finish every chip and the whole line still stops if the final assembly step is jammed.

Bottleneck 4 — Networking

Nvidia's InfiniBand and Spectrum-X, Arista's data center switching, Cisco's infrastructure. When you scale to tens of thousands of GPUs and make them function as one machine, that's one of the hardest engineering problems in the buildout. Speed and latency at this scale are brutally difficult.

Bottleneck 5 — Power (maybe the biggest opportunity)

Constellation Energy, Vistra, and NextEra are being pulled directly into this story. You can have the chips, the building, the customer, and the signed contract. If you can't get electricity to that data center, it does not run. I think this might be the biggest bottleneck and the biggest opportunity of all. AI doesn't run on hype; it runs on electricity.

Bottleneck 6 — Cooling

Vertiv (VRT), Schneider Electric, and Eaton are doing real infrastructure work here. Yet most investors haven't started pricing it correctly. The heat problem scales with every new data center that comes online. More compute means more heat, and the leverage of the companies solving it grows with it.

Bottleneck 7 — Security

CrowdStrike, Palo Alto Networks, and Datadog sit at the intersection of AI infrastructure and protection. Every new workload that goes live is a new attack surface. At the enterprise level, security is not optional. The more critical AI systems become, the bigger the target on them gets.

The lag is the opportunity

Here's the point: the market does not price these seven bottlenecks simultaneously.

That lag is the opportunity. The obvious winner runs first, then the supplier, then the bottleneck, then the infrastructure layer. Compute and memory are already largely priced, but layers tied to the physical world — power and cooling — still have stretches the market hasn't valued correctly.

There's a bear case, of course. If grid buildout can't keep pace with data center demand, the entire physical layer slows down no matter how strong the demand signal is. But that's exactly why power may be the biggest opportunity — the constraint is the pricing power. Track the seven bottlenecks layer by layer and you can see where things are already tight, and where they're just starting to tighten, before the crowd does.