“The DeepSeek model's inference implementation requires clusters of approximately 160 GPUs, representing over $10 million in hardware for a single replica.”