A model's context window is a feature on the spec sheet. For whoever pays for the GPUs, it's a bill. Every token in every active conversation leaves a footprint in GPU memory called the KV cache, and that footprint grows with context length and with the