“A 30% increase in compute cost would translate to a 15% reduction in AI research velocity, delaying key milestones by a few months.”