How to choose between 8-GPU, 4-GPU and 2-GPU AI servers?

Large-model training: prefer 8 GPUs. Model fine-tuning and mid-scale inference: 4 GPUs. Small-model inference, testing and AI development workstations: 2 GPUs. Single-card VRAM, cluster size and budget decide together.