The architectural efficiency risk you outlined is the most underappreciated threat to current AI valuations because a sudden 10x or 100x improvement would strand billions in infrastructure investment overnight. The Giotto.ai example is particulary compelling since 200M parameters competing on ARC-AGI leaderboard suggests we might be grossly overbuilding for the eventual cognitive core requirements. Your point about not being able to Jevons your way out of overcapacity is brilliant because even if demand doubles or triples from lower costs, you still can't absorb 99% redundant capacity fast enough. The comparison to railroad and dotcom bubbles is apt, but the key difference you identified is that in those cases demand materialized slowly, whereas here the risk is that demand stays strong but the substrate underneath completely shifts.
I mostly agree but there is a counter argument I would like to make. The trillions being invested in GPUs are capex with an amortization time of 3 to 5 years. Where is the income justifying this investment going to come from? Because current services di not generate anywhere near these amounts…
The architectural efficiency risk you outlined is the most underappreciated threat to current AI valuations because a sudden 10x or 100x improvement would strand billions in infrastructure investment overnight. The Giotto.ai example is particulary compelling since 200M parameters competing on ARC-AGI leaderboard suggests we might be grossly overbuilding for the eventual cognitive core requirements. Your point about not being able to Jevons your way out of overcapacity is brilliant because even if demand doubles or triples from lower costs, you still can't absorb 99% redundant capacity fast enough. The comparison to railroad and dotcom bubbles is apt, but the key difference you identified is that in those cases demand materialized slowly, whereas here the risk is that demand stays strong but the substrate underneath completely shifts.
I mostly agree but there is a counter argument I would like to make. The trillions being invested in GPUs are capex with an amortization time of 3 to 5 years. Where is the income justifying this investment going to come from? Because current services di not generate anywhere near these amounts…
Good point - not all of it though.