Constraints on AI previously thought to be hard (compute and energy) have been liberated by models like DeepSeek that can self-learn and accelerate faster than expected with the same compute and energy, and we don't have a lower bound on the energy required per inference or per job completed