They target this model at generating synthetic data. Data is the lifeblood of LLM training; quality synthetic data means more training can occur which means more demand for GPUs.