Squeezing more from existing GPU fleets: quantization vs batching?

We are evaluating options to cut inference costs without trading away too much quality. Curious how others are balancing quantization (like int8 or fp8) with dynamic batching in production. What is actually giving you the best throughput-per-dollar? We are also struggling with autoscaling when traffic is bursty; cold starts hurt us and we end up paying for idle GPUs. Are there good patterns for managing scale-to-zero and keeping latency acceptable? Let's share what is working