Unless we are really going to hit the Singularity and create Godlike superintelligence that improves beyond all limits, sooner or later the models will stop improving, no matter how much compute we throw at them. After that happens, the best models will be etched into model-specific ASICs with weights fixed in mask ROM (like Taalas is already doing), which will result in massive, order of magnitude inference efficiency increase. Couple this with the training datacenters suddenly becoming available for inference, and the cost of compute will plummet.
Yes, it might take a few years, but the hardware will be cheap again. And if the result of the whole process is human-level AI, then it will be worth it in the end.
This is empirically false when it comes to technology.
A few years of technological inflation (with a very specific cause) do not invalidate a hundred years of technological deflation.