In the fiercely competitive AI landscape, a million-dollar experiment is quietly revolutionizing large language model (LLM) training. Jieyue Xingchen's research team recently released groundbreaking findings. By utilizing nearly 1 million NVIDIA H800 GPU hours, they trained 3,700 models of varying sizes from scratch, accumulating a staggering 100 trillion tokens. This revealed a universal scaling law dubbed 'Step Law,' offering a novel guide for efficient LLM training.