$Tradr 2X Long WDC Daily ETF(WDCX.US)

$Micron Tech(MU.US)$Sandisk(SNDK.US)$Western Digital(WDC.US)

Here's a simple discussion post. It's about cache and cache hit rate.

Everyone knows the mechanism of Kv cache, and everyone has experienced Fable 5. My personal view is: in the long run, achieving a high cache hit rate is a must for all AI companies, and DeepSeek's hard drive cache mechanism is very important.

The hard drives used for cache can't possibly be HDDs, as the most important random read/write performance is far superior in SSDs compared to HDDs. I want this cache, my HBM, right now, are you giving it to me or not?

Why has WDC fallen so much? Because the data used for training is stored in HDDs, which is normal. But now, what's actually being done is post-training, and the hard drives that needed to be bought have already been bought. The relatively easy-to-obtain part of human data is really already exhausted for training.

Fable 5 has excellent product intuition, acting as a leader among models, directing other models. Future models will also see significant differentiation. Personally, I think a 1000 GB NV Switch or something similar will become a significant threshold. Most small and medium-sized models will be around this scale, completing most programming tasks. For tasks like image generation or article writing, the scale is even smaller, and Gemma 4 already performs brilliantly.

So, I suggest waiting a bit longer after SK hynix lists in the US. After all, HBM4 reaching a reasonable price is crucial for Fable-level models. But honestly, this demand isn't widespread. The remaining small and medium-sized models can really be adequately served by the existing stock of cards or those that will be phased out in the future. The entire storage market is very likely to face a major turning point.