Some thoughts
A high capacity, reliable, general purpose proxy, like Squid for example, could bridge from your camera network to your client network. It can proxy media streams, so you could proxy from unicast on your camera network to multi-cast on your client network.
You won't have much in the way of licence fees if you create an open source solution, and you can engage people with lots of experience in high performance computing platforms. Open source though, is not going to vendor sell itself to you, even if it is the best solution. You have to choose it and chase it down.
100Gbit line cards are becoming common on mid range cisco, so that scale of bandwidth is doable.
You're going to need something like puppet or ansible to push out config updates to your camera fleet.
Your dhcp, dns, snmp & ntp servers will need to be carefully configured, you'll need real tools, not toys. If you get them all to work together nicely you'll have better uptime statistics.
You'll need traffic probes at strategic points in your network, so you can see what's going on with dodgy camera configs, misconfigured network elements, unintended consequences.
Clients will need to limit the camera feeds they choose to their available incoming bandwith. Even if you go multicast, the last mile to the client carries the same amount of traffic as unicast.
You might end up going UDP, simply because the ack & retrans request traffic will kill your network the moment you have congestion. It will make a bad situation a lot worse, rather than just gracefully dropping a few frames as you approach network limits.
If your clients love your project, they will buy dark fibre from their data centre to yours and light it up with 10Gbit or 100Gbit gear, since they are likely in the same city. Then you'll need more intermediation servers, as they consume a greater percentage of your available feeds.
Half your clients will be public entities of some sort. If your regional and national government is smart about how it spends your tax dollars, it will direct them all to collaborate, then they'll ask you to install the VMS, and they will then just consume finished and ready to use video walls & history searches.
I've been deploying NX Witness at micro scale. It seems very efficient, and their people innovative. Because it runs well on Linux, it could probably scale to a substantially larger cluster than any Windoz platform VMS. It also virtualises well on Proxmox or OpenStack, so you could scale out with automation. Could be a project NX would adapt their software for. It would certainly need some more meta tools.
Good luck, exciting project.