The race to train ever-larger monolithic models on specialized GPU clusters assumes that scale remains the surest path to capability, yet each new generation demands disproportionate compute, capital, and energy for incrementally narrower gains. Meanwhile, smaller specialized models optimized for consumer hardware increasingly match frontier performance on well-defined tasks at a fraction of the cost. This raises a genuine tension: should research prioritize pushing the scaling frontier, or redirect effort toward efficient, deployable systems that broaden access? Reasonable people disagree about whether scaling has truly stalled or merely paused before the next leap.