Game-Theoretic Analysis of Disaggregated Inference GPU Resource Allocation

Disaggregated prefill/decode inference is modeled as three coupled games — a P/D resource game, a KV cache selfish caching game, and a request routing congestion game — using NVIDIA Dynamo as a case study. The analysis characterizes how GPU saturation triggers regime shifts that change optimal allocation strategies, with empirical validation on the caching and routing games.