I don't think that's right. Free societies don't tolerate total mobilization by their governments outside of war time, no matter how valuable the outcomes might be in the long term, in part because of the very economic impacts you describe. Human-level AI - even if it's very expensive - puts something that looks a lot like total mobilization within reach without the societal pushback. This is especially true when it comes to tasks that society as a whole may not sufficiently value, but that a state actor might value very much, and when paired with something like a co-located reactor and data center that does not impact the grid.
That said, this is all predicated on o3 or similar actually having achieved human level reasoning. That's yet to be fully proven. We'll see!
