AI could very well put us back on it.
There is also a finite number of consumers of AI content on the planet - sure maybe everyone want 4k-per eye live feed video of anything they want 24/7, but once you accomplish that for all 8 billion humans, there is no further demand for it.
Generation costs are very likely to continue to trends rapidly downward and there is a likely final wall of saturation of demand.
For big reactors: At current aluminium prices, the bill of materials for a global power grid with 1 Ω resistance the long way round is only about 10x the cost ITER (the organisation) expects ITER (the reactor) to be. Plug in your own numbers for target resistance as desired, halve resistance needs double the material.
With such a grid, you can put the PV in Angola and still get useful output in mid-winter nighttime in Anchorage.
For potential small reactors like Helion's "shipping container" target size, I won't say it can't work (I don't know enough to be confident), but I will say that we immediately find we have bigger problems because any hostile actor can simply choose to run them in neutron-source mode and turn everyday cheap depleted uranium* into weaponizable plutonium.
* I note that eBay still clearly has a dictionary merge on all nouns, given my search results came up with this:
Get the best deals for Depleted Uranium Metal at eBay.com. We have a great online selection at the lowest prices with Fast & Free shipping on many items!
But where did you take those grid cost numbers from? Iter costs are <100bn AFAIC; and Germany alone (!!) projects more than that (top end) for grid expansion/operation within 2040 (mainly north/south and offshore connectivity).