Your team runs an image-classification inference workload on a fleet of n1-standard-8 Compute Engine VMs, each with an attached NVIDIA T4 GPU. Requests arrive in unpredictable bursts throughout the day, and between bursts the GPUs sit idle for long stretches while you still pay for the running VMs. Latency requirements are moderate (a few seconds is acceptable), and each request is fully stateless. You want to reduce cost during idle periods without rewriting the model-serving code, while still handling sudden bursts. Which change best optimizes cost and performance?