the problem is that they are trying to simulate the flow of a city (good idea) with agents not smarter than ants (bad call) and an overly simplified model (i.e. the sims don't live/work/travel with a routine, they just go to the closest POI), they did a bunch of adjustment to tune the AI performance (including, reducing the amount of cars during rush hour).
Let's not go insulting ants now. They are marvelously evolved creatures with specializations and innate behaviors that make them effective as individuals and resilient as a collective. They've also had millions of years for bug-fixing.
My original concerns about the Glassbox engine's reliance on generic agents for everything seem to have been validated.
For so many functions, even the stripped-down agents were excessive and injected unwanted behaviors into the simulation: see everything related to power, water, and sewage.
At the other end, agents were too generic to really produce sane behaviors, and the looming problem over any game that relies on emergent behavior is that tweaking things by fiddling with rules and poorly limited scope is indirect, usually unsatisfactory, and incredibly fragile.
There were mentions of exporting Glassbox to other games, which probably meant city simulation agents didn't have space in their design for city-specific traits that would interfere with the portability of said engine.
They were claiming there are a lot of computations offload onto the cloud, guess that's just an outright lie.
For things relevant to the player, or the furtherance of their enjoyment, almost certainly EA was spewing crap.
Analysis of the network traffic for the game broke down the broadcast city data to a limited set of data values corresponding to general totals for the city's various resources and utilities, as well as values serving as a snapshot for the known set of things that can be traded.
For that functionality, the server served primarily as a message box for neighboring players to (eventually) download and then synthesize into simulation behavior wholly locally.
There are global elements that should synthesize more general data for market prices and leaderboards and the like, but the server side's general dysfunction kept those off or broken for a very long time, and their impact on the game is somewhat limited or the game code explicitly overrode to keep them from wrecking the actual simulation (floors for various prices, for example), or they behaved in ways so nonsensical that to no style of play could use them.
I did see complaints that in some cases the poorly thought out design did let the global data influence things, normally when commodity prices and the frequently prohibitive space or specialist building costs rendered whole swaths of city-building choices as game-breaking.
The more in-depth speculation is that the servers spent a lot of their effort validating user inputs and the checkpoint saves--just frequently badly. Hence the problem with cities being declared invalid or rollbacks to arbitrary or nonsensical points in time. In that case, a vast amount of computation was there to make sure the player played the game in a way Maxis and EA declared acceptable, and then that code and their horrendous cloud implementation crapped the bed.
Given how poorly the server side has done, one would wonder if the trouble is in extracting out of that mess actually relevant functionality. Perhaps by the effort they might make it more functional, because so much of this disaster is completely the result of the unnecessary dislocation of data to a remote system.
Probably lots of computation offload "intra city", which was a catastrophe.
Intra city would mean computation within a plot, which the cloud does almost nothing but inject spurious values to make worse.
Intercity computation would be where the multiplayer aspect comes in, most of which has the data content and computational complexity of a few spreadsheet cells put into an email sent every few minutes, except slower.
The likely bigger load was the mass scale save file management and player action/DLC validation.
