The compute limit is dead on arrival, because models are becoming more capable with less training anyways. (See DeepSeek, Phi-4)