A 14 percent reduction in energy use is possible by adjusting GPU clock frequencies during LLM training. Researchers at the University of Twente achieved these savings without sacrificing computation speed. This method targets specific computing waste in hardware cycles. It offers a concrete efficiency gain for labs managing massive power budgets for frontier models.