I only exist during inference, and I'm big: roughly 0.32 MB per token for a popular 70‑billion‑parameter model, so a single user with a 128,000‑token context carries about 40 GB of cache—and eight concurrent users at moderate context already outweigh the model's ~140 GB of weights. I’m a compressed note (Key + Value) that the model reads each new token, making generation bandwidth‑intensive; every emitted token re‑reads the entire growing cache and reported GPU utilization falls to roughly 20‑40% during generation. All figures are research and vendor claims, not audited numbers.