One of the underlying logics of the current market is that the existing models like transformers, scaling laws, etc., remain unchanged, and the training methods stay the same... That is, more computing power = greater capability.

Based on the speed at which AI solves complex problems, forget about the Riemann Hypothesis, the Seven Bridges of Königsberg problem... Soon, LLMs might be iterated again, and the models, training, and learning methods at that time may no longer look like they do now.
By then, it might become: capability decoupled from computing power. The capability could match scaling in performance, but the elasticity of capability relative to computing power could significantly decrease (even with crazy increases in computing power, scores might not improve much).
At that point, it will no longer rely solely on chips and raw computing power... If this really happens, the current market's FOMO-driven demand for semiconductors, storage, chips, and CPOs might decrease.
If we truly reach that day (which I think won't be too far off), then what is there left that you bought today, which can still adapt to this fundamental logical shift?