Hi Bobby,
My organisation is using EA to do exactly what you are intending to do.
We implemented the same high level structure you have (being a "corporate" model and "projects" model)on SQL Server 2000.
Here are some stats on our database size to date:
# of packages = 1,503
# of elements / objects = 5,352
# of connections = 2,517
# of diagrams = 343
# of tagged values (as we use custom stereotypes) = 37,161
SQL Server DB size = 1.1Gb (without optimisation).
Here are my observations (all local LAN access / high speed):
Performance:
=-=-=-=-=-=-
1. EA takes approx 12 seconds to open the project
2. The more elements that are on a diagram - the slower it will open. It's not that bad though - I have an "overall" diagram of our applications and with 100+ it only takes about 5 seconds
3. Adding and deleting elements has slowed somewhat. It now takes about 1 - 2 seconds per element. This is now my biggest concern as this is a frequent operation.
Security
=-=-=-=-
Remember in EA every user has read access to everything when given access. There is no way to restrict this. Why is this an issue? Confidentiality. Consider you might have vendors / external consultants working on a project... do you really want them to be able to browse things not relevant to their project?
We have implemented a "domain" level group - who can read/write the corporate model, then a "project" group who can only read the corporate model but update the project model.
Governance / maintenance
=-=-=-=-=-=-=-=-=--=--=-=
You need to make sure you have worked through how project created elements get moved into the corporate model on project completion. This allows you to slowly "grow" the corporate structure
Standards
=-=-=-=--=-
Agree on all stereotypes to be used and additional metadata. Key elements I would recommend you stereotype are "application", "database" and "node". Lock this down so project people can't create tagged values and stereotypes otherwise you'll get a real mess.
Keeping data current
=-=-=-=-=-=-=--=-=-=
The more you put in, the more useful it is, but easier to get out of date which erodes user trust. I suggest you automate feeds directly into the back end DB to keep things current . I am doing this now for our infrastructure - feeding off a Computer Associates AMO database that maintains stats on our nodes
I did evaluate Metis (from Troux) and whilst a powerful tool - was expensive. I have also found Sparx support extremely good and responsive. I was the one who reported the slow WAN open performance problem - and Sparx fixed it their next build which was released only 10 days after I logged the bug! I don't know any other vendor with that level of responsiveness!
Single database and "newbies"
=-=-=-=
The newbie can trash a corporate database easily. Undo in EA is limited. Make sure you mitigate this with "playpens" for newbies, adequate training, locking down the security, implementing a version control system.
There's a few other small gotchas but nothing major. I think you are heading down the right path (or at least your ideas concur with mine

)
Hope the above helps.
David.