"DeepSeek production deployment, which is well over a year old now, they were running on 160 GPUs. That is what they serve production traffic on. They split the model across 160 GPUs."