Detailed tl;dr for those who can't/won't read the thread I linked.
It seems expansion bases are the problem above and beyond any unit caps or any other factors. Essentially, for each expansion base, the AI creates a secondary AI called a "Build Manager". Build Managers are not removed from memory until every single engineer and factory associated with the expansion is stamped out, so if a damaged expansion is rebuilt, an additional Build Manager will be created on top of the previous one. Because Build Managers independently evaluate battlefield intel and adjust their production behavior accordingly, it shouldn't be hard to see why SupCom begins to rapidly slow past a certain point.
Solutions? First off, the makers of Sorian AI have taken great steps to quell a lot of memory waste associated with the traditional GPG AIs, but their optimizations still have limits and cannot address the Build Manager issues. Using Core Maximizer can potentially help, but I don't have solid intel regarding its effectiveness on Win7. Avoiding water-heavy maps seems to make the splintering Build Managers less wasteful. And the final, and probably most effective, solution is to clamp the number of expansion bases the AI can build. I hypothesize that limiting water expansions would have more effect than clamping land expansions.