Recently, I've been studying the Rubin cost breakdown and found that the current bottleneck for AI isn't simply a lack of GPUs, but rather:

1. Insufficient context length: So HBM/DRAM/SOCAMM/SSD are still crucial, corresponding to $XL2CSOPHYNIX(07709.HK)$Micron Tech(MU.US)$Roundhill Memory ETF(DRAM.US)

2. Data movement bottlenecks: So optical interconnects, CXL, and memory pooling will become increasingly important, corresponding to $Marvell Tech(MRVL.US)$Broadcom(AVGO.US)

3. Inference is too expensive: If memory capacity expansion or architectural optimization drives costs down, the ultimate beneficiaries might be companies like $Alphabet - C(GOOG.US), OpenAI, and Anthropic, which are truly monetizing AI.

Memory is still the main theme, but it's unclear how long the rally will last. So, here are some corresponding hedging ideas:

1. Worried about memory being disrupted by optical interconnects/tiered memory: MRVL, AVGO

2. Worried about the price increase logic disappearing after memory capacity expansion: Google, OpenAI (future IPO), Anthropic