I've been doing research about what kind of bottlenecks that actually last. The story now is not much about GPU any more. It shifts to CPU now as AMD and Intel explode. I do think RAM (HBM) is even more important as agentic AI requires tons of Ram. But eventually ram supply will catch up with demand. Could be years. But the reality is hyperscalers won't keep spending on Capex forever. Same for networking component suppliers like Lite, Cohr, Anet, etc. Eventually they won't build any more data centers.
But what will last forever like a forever subscription is the power cost. Power is constantly needed to feed those hungry data centers. I know power is not new when you discuss about AI bottleneck. But I think it's the safest play for long term.
Currently I'm super bullish on RAM supercycles. I'm looking to adding more DRAM etf. I do own MU as well. But I plan to build a substantial position in AIPO etf as well by trimming my NVDA stakes. What do you think?
What's more important than power in the AI era?
byu/coopermug ininvesting
Posted by coopermug
2 Comments
I’ve had the same view since 2015 when cloud was all the hype and every year I’ve been disproved. Every year the amount of data created grows by 20% CAGR so the infra required to support processing and serving that to humans will only increase. It’s difficult to try and game what part of the supply chain will outperform over different periods of time and future type of workload the end user needs (GPU inference vs Agentic CPU/RAM). As long as technology plays a role in advancing humanity, semis/memory/photonics/etc will always be needed.
Hinge companies. Data centres need good doors. Hinge companies are sure bet.