nv-embedding-cache

The generative recommender's real constraint is not the model, it is the memory

NVIDIA's case for next-item generation in recommenders reads as an architecture story. The number that decides whether it ships is terabytes to petabytes of daily user history against GPU HBM.