While isolating my AI model cluster on its own network segment, I could not reach a host on another network from my lab machine. I was suspecting a faulty NIC card, a faulty cable, or a wrong connection on the switch in the lab.
AI came in handy to detect it. I asked it to trace why I was not able to reach that host. The machine had two NIC cards, one on my subnet and the other used as a bridge to another subnet. It said the new IP being assigned was from the same subnet as my own. Two networks with overlapping addresses.
So I gave the bridged network its own range.
If you are using a VLAN, ensure that it is isolated and only one of the nodes has the bridge enabled. It could be wired or wireless.
To check quickly: ip -br addr shows each card with its address and prefix. If one range sits inside the other, you have the same problem.
With private AI this matters more. Model servers should sit on their own segment, with one way in and out. I wrote more on why: https://yantrakar.com/self-replication-network-boundary.html
Have you lost time on overlapping subnets?
Written with help from Claude. The experience is mine. Addresses in the pictures are deliberately obfuscated to avoid leaking internals.