It is a long post dear fellows, not surprising to those who knows me. Starting with some other games who got/have similar problems, then checking on CoE's specificities.
Single shard worlds and activity concentrations
If we have a quick look at EvE Online, the safe (no-combat zone) tribal capitals are a server nightmare, because of two things: there is a lot, and i mean a lot, of people in close vicinity on the same local map and which are very active transaction-wise (transactions that each impact the whole world and its economy).
It regularly lead to waiting queues to simply enter the place.
Second is when there are those famous battles (guiness book records by the way) with thousands concurrent online players, but it is easier to handle with a predictive allocation of server resources... and they don't impact the whole world, just the local space-ography.
Optimize later?
Then you have similar titles who tackled the problem too late, like Star Citizen: despite their budget, they still have a hard time about just that, with a goal of 200 players per solar system, though I didn't follow their recent conference.
And almost every MMO out there, except maybe Dual Universe: since its core feature is a technology to handle a single shard server with any number of clients/players in close vicinity.
Culling limits, many
Culling was used as a last line of defense by Guild Wars 2 for both the RvR and event maps, trying to pass the 200~250 participant limit.
And they also had a really hard time trying to optimize that for a couple years after launch... to the dismay of their PvP crowd (among other things not thought when they should have).
Soulbond partnerships
So yeah, the OP question is a very valid one, one that was resolved early by SbS with their partnership with Improbable (massive network engine)... except it didn't work for reasons we can only speculate, so back to square one.
Unreal Engine is a very potent and experienced graphical engine (actually more like a full game engine), but doesn't do much for network, and definitely not for online requirements above the dozen.
On screen no-limit
In a sense, don't worry on how many "characters" a game may have simultaneously on screen, it is a concern of the past: graphic processors and memory can handle that just well, just like any other "destructible" or "interactable" asset.
Total War series can vouch for me as an example and repeated proof of concept.
Data exchange is the issue
The real problem is about how much information has to be exchanged among clients and between clients and servers: since in current and classical implementations, it grows exponentially as clients (both players and scripts) find themselves close to each others and their data/interaction bubbles cross paths.
And you can't rely on peer-to-peer, though it's "cheap" and done well can alleviate the issue, it leaves the door open for hackers and cheaters (as any PvP player would tell you, it has ruined many games that didn't have a dedicated server infrastructure).
Redeeming factors in favor of Elyria
An announced maximum of around 100k players per server (a bit less for OCE) in total.
In Elyria, local numbers are surprisingly not that big, density-wise... around 60 players per county in average from what I read on the forums (within an average 16km² if I recall correctly), and ten times that in NPCs (at most).
An elyrian city at max-level only needs a minimum of 250 elyrians in population: roughly an average of 3 people per parcel, while a single perfect parcel could hold 64 cottages side-by-side (on just the ground level).
Note that there is always the possibilty to join a server through a NTC code, but as player characters never disconnect, the total elyrian (PC, OPC, NTC, NPC) population to handle by a "server" stays relatively constant: that's smart.
A measure of "lots of people" in Elyria
I don't know, players will try to push the enveloppe.
But I can only speculate that it won't be that much... or to be exact, it will be highly difficult to gather a thousand characters in the same place (for a memorable concert or battle), and live for another day: resource depletion, body collision, elyrian time to do the trip... heck, they will probably not hear or see what happens on stage!
Business model to the rescue
Actually like any major website or online app, a "game server" is a myriad of hosted servers working seamlessly together... and generally linked with fiber optics (with next to no ping at all). If one fails for any reason (or is suddenly overworked), another kicks in and pick up (or share) the workload.
The Life Spark we use to spring life into one character for a RL year is our paying subscription to the online service and support. Rhetoric question, who pays for the NPCs? ;)
That's why the business model includes a ratio of NPCs per Soulborn (player character, online or not)... and I imagine that ratio will evolve during the development phase to match server performances.
In a nutshell
If they manage to solve the conendrum of data exchange between clients, servers, the game will be smooth.
If they don't, there are still the (solo/team) prologue and Discord-mode to enjoy. :p
Organic design
That's why an organic design approach is on the safe side: you get what you can "sense".
"Sense", not "see", that's a mistake that shouldn't be made: you can look at the horizon (40 kms away on earth I think) and perceive a sail or a landmark... that's things the engine will draw for you if you look that way.
Doesn't mean the character (and player) need to get updates from everything happening within that range, actually it shouldn't... graphical engines learned to do it the hard way, internet too, gaming network not so much.
What matters is on whatever someone focus their attention and their peripheral "vision" (ie. perception, including sounds and maybe smells if they are implemented), there are the only things that need constant data updates outside their own status.
The rest can be leisurely updated (at slower paces, like weather changes, etc.) For instance, people passing down the streets, you only need their "silhouettes", not their entire character sheet.
In that example, if you give a particular group a "second look", then the network/game-master engine will respond and bring you more details... and provide them with a possible hint that they are observed (skill checks, etc).
It is the same over-detailled data in the background, but a character (played or not) get only simplified views, just what was needed/determined in such situation/case.
It also make things easier to script behaviors reliably... something few online titles took time to look at, generally engrossed in theme parks and level designs instead.
And that is, my two cents on the topic, you can keep the change. :)