Key insights
- The author believes power consumption is the most enduring bottleneck in the AI era, surpassing GPU and CPU constraints. While bullish on a RAM supercycle and DRAM ETFs like AIPO, the author suggests power infrastructure investments offer a safer long-term play. This perspective implies potential shifts in capital allocation away from traditional hardware and towards energy solutions supporting AI infrastructure.
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. Most importantly, RAM doesn't care whose chips will be used. NVDA/ intel CPU, amd, Google TPU all need 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?