The bottleneck of long-context LLM inference is shifting from "can it compute" to "can it fit" — the memory occupied by the KV cache grows linearly with context length, which is the hard bottleneck preventing million-token inference. Anna Cordoba et al., in the arXiv paper DepthWeave-KV (arXiv:2607.06523) released this week, didn't take the conventional compression path of "cutting the whole layer uniformly", but used a set of shared low-rank channel bases to factorize the Key/Value states of adjacent Transformer layers, with a lightweight residual left on the token dimension. The real key design is token-conditional routing: higher reconstruction rank is allocated to the key tokens that carry instructions and undertake retrieval tasks, while ordinary tokens are compressed significantly; simultaneously, a calibration-free online error probe is introduced from the attention output end, dynamically adjusting compression strength during generation without retraining the base model. The companion fused CUDA kernel combines base lookup, residual dequantization, and attention projection, reducing memory access during decoding. On LongBench, Needle-in-a-Haystack, L-Eval, long-form QA and summarization tasks, DepthWeave-KV achieves 8.3× KV memory reduction and 72.8 tokens/s throughput at 64K context, with task quality close to full cache, and outperforming existing compression schemes. Moving from "uniform budget" to "per-token adaptive" is the right direction for KV compression — compression isn't flattening everyone, but leaving channels for the key tokens. DepthWeave-KV engineers this intuition, providing a landing path for production-grade long-context services.