Anyone who deals with AI shares almost automatically ends up with Nvidia. That's understandable, GPUs (Graphic Processing Units) are the starting point. But that's exactly where the shift begins for me shift.
The limiting factor is no longer primarily the computing power, but the memory, that makes this power usable in the first place. There are several bottlenecks in the current area. One of these is HBM (High Bandwidth Memory).
These chips are located directly next to the GPU and determine how much data can be processed simultaneously, Without sufficient HBM, no AI system can scale. At the same time, this is not a standard product, production is complex, yields are limited and integration is challenging, while demand is increasing massively.
The bottleneck no longer lies with the compute, but with the manufacturers. For me, therefore, the current $HY9H (+5,1%) (SK hynix) and $MU (+4,02%) (Micron Technology) are closer to the bottleneck than the usual AI poster boys.
This is not a static picture, a year ago the focus was clearly on GPUs, today on HBM, tomorrow probably on other parts of the infrastructure. This movement is exactly the core of my approach: Bottlenecks migrate.
Next, I'm looking at Power & Cooling, Advanced Packaging and Energy. Areas that are just starting to become limiting factors themselves.
