1. training costs are going to continue to escalate much faster then the models improve.
2. the current paradigm of trained ahead of time token completion turns out not to be a pathway to "AGI" or even anything much more then we have now. As this becomes clear investment dries up.
3. without investor support user's are forced to pay full price for LLM queries and a lot of things that are super useful when this shit is subsidized up the wazoo are not worth it at full price and when people are given the choice of stopping using LLMs or ponying up more cash people realize they can get by without them.
4. Everyone who was hyping LLMs moves on to the next shiny thing and all money for "AI", even for useful stuff like old fashioned machine learning, dries up and we end up somewhat backwards from where we started except GPUs are cheap again assuming the crash doesn't take down NVIDIA.