Anyone else seeing GPU idle time spike after autoscaling?

We run a mixed inference fleet and our autoscaler keeps over-provisioning. During traffic dips, GPUs sit at 15-20% utilization but we're still paying for reserved capacity because cold starts on the larger models take four minutes. Tried predictive scaling off request history, but the load pattern is too bursty. Curious how others handle the tradeoff between keeping warm replicas for latency and burning money on idle silicon. Also wondering whether anyone has moved to smaller draft models for routing to cut the baseline footprint. What's actually working in production for you?