“The ideal allocation of compute resources for an AI lab is a 50/50 split between experimentation and final model training.”