The KV cache used during inference can consume significantly more GPU memory than the model parameters themselves, with costs reaching 4.5 GiB at 128K context on gpt-oss-120b and up to 40 GiB in dense configurations. This highlights the growing memory bottleneck imposed by attention mechanisms in large language models as sequence lengths increase. Optimizing KV cache storage is becoming critical for efficient and scalable deployment of transformer-based models.