Bandwidth to where? Do they really have 1gbit per customer peering with say level3? And with cogent, and telia? And a full non blocking internal network? And enough packet buffers to ensure that microbursts don’t saturate any link?
Depending on the SLA, or lack thereof, they could have enough peering for a sustained +(x = 3?)σ demand spike without having a full dedicated peering for each customer, and still reasonably guarantee throughput availability.
Even ignoring peering, I’m still trying to picture the non blocking network. An Arista 7368X4 isn’t cheap, but will cope with 12,000 customers with appropriate switches downstream (128x100G ports with each port breaking out to 96 customers. Not sure how you’d scale beyond that without blocking.
Your peering will only scale to your (combined) interface speed regardless of your SLA. Our 2x 40G peer with level 3 serve far more than 40 1G devices (I personally have 600 in one building alone, and that sets aside the rest of the users), but it’s rare it’s more than 30% utilised.
Clearly that’s not going to be a domestic isp architecture, so a reasonable question is what does uncontended actually mean.
I very much doubt an isp with 10,000 customers has 10TB of peering physically available. Linx public peering is less than half that for the entire UK, and while private peering will likely increase that, the suggest is it’s welll under tenfold, so the 50 million plus internet users in the UK only use at peak times 1M each, and that ignores all the non-domestic use.
Even a 100:1 contention ratio seems enough at a core level, so any isp spending money on improving on 10:1 ratios seems that they are just burning cash.
Clearly as you go the the edge contention ratio needs to drop - but 40G, or maybe even 20G uplink for 48x1G users would be reasonable to me.
