Micron Technology IncDeepSeek's V4.1-Flash cuts KV-cache HBM needs by 75%, reducing demand for Micron's HBM memory products.
DeepSeek's September 10 release says its V4.1-Flash model needs one-quarter as much HBM and one-eighth as much SSD capacity for its KV cache as the previous generation, a 75% cut in HBM and an 87.5% cut in SSD for that component. The comparison covers only the KV cache, which stores attention state during inference, and is measured against DeepSeek's preceding architecture; it does not cover memory used for model weights or training. The model has 552 billion parameters but activates 8 billion for input and 16 billion for output. Micron's Cloud Memory revenue reached $13.77 billion in its latest quarter at an 83% gross margin, while Core Data Center revenue was $11.52 billion at an 87% margin, and the company shipped more than $1 billion of HBM4. Sandisk's quarterly data-center revenue rose 103% sequentially to $2.98 billion, with pricing generating roughly two-thirds of its companywide sequential revenue increase. As of August 31, 5,996,105 Sandisk shares were sold short, equal to 4.10% of float and 0.46 days of average volume.
Micron Technology IncDeepSeek's V4.1-Flash cuts KV-cache HBM needs by 75%, reducing demand for Micron's HBM memory products.
Sandisk CorpDeepSeek's model cuts KV-cache SSD capacity needs by 87.5%, threatening demand for Sandisk's data-center storage.
DeepSeek released V4.1-Flash, a model that dramatically reduces HBM and SSD memory requirements for its KV cache.