“The demand for High Bandwidth Memory (HBM) is being driven by the need for longer context windows in large language models, which require more KV cache.”