“The DeepSeek model was trained on 15 trillion tokens of data.”