i mean my line is always "when." people have been saying ai is a bubble and will crash since the beginning. the question has always been "when." i think the best framing of the ai crash is not in terms of software, but in terms of real estate and infrastructure. the real cost is hidden in construction contracts and people buying up land (and chips, transformers, generators, etc), not in technology. this will be the 2008 real estate crash, not the 2000 dot com bubble. (altho this point is not my most well researched point, i tbh need to look more into hard numbers of where the most risk is distributed, ai software startups or data center contracts. but i also am in grad school i have papers to publish and no time.) im also gonna push back on the "diseconomy of scale" point. while its true that the best models rn use the most resources, we are building better smaller models as well. i recently set one up on my 12 gb vram and while its not nearly as good as claude, it works. i make it write really simple code for me. and we are still improving small models. i think theres some level of hope to economy of scale. obviously there are going to be physical limits, but im sure there will be a way to get economies of scale to work.