An engineer measured how Kubernetes time‑slices a single GPU when two LLM agents run side‑by‑side. Using a five‑year‑old GTX 1080 and the stock NVIDIA device plugin, the test placed a latency‑sensitive FFT worker and a heavy matrix‑multiply worker in separate pods, each requesting nvidia.com/gpu:1. The scheduler reported both pods as Running, but the GPU handled both workloads through CUDA time‑slicing, mirroring cheap edge deployments today.

Metrics collected with CUDA events showed median latency unchanged, yet the p99 latency of the small agent rose from 3.68 ms to 6.10 ms – a 1.66× increase – and jitter jumped from 1.02 to 1.70. Throughput dropped only a few percent, so average‑oriented dashboards would miss the degradation that hurts deadline‑critical services, for real‑time pipelines under load.

The experiment proves that Kubernetes’ health checks hide resource contention; the tail‑heavy agent bears the cost while the system appears healthy. Operators deploying edge AI, such as 5G baseband processing alongside LLM inference, must monitor tail latency directly or isolate workloads on dedicated accelerators. The open‑source profiler in the accompanying GitHub repo quantifies this hidden penalty for production workloads across clusters.