If you saw how digital data traverses, it’s mind boggling.
Almost all tier 1 backbones do what is called closest exit routing, colloquially referred to as, "hot potato routing". If the route originates in AS7018 (AT&T) and the next BGP hop is AS701 (Verizon), the 7018 network does not give a hoot about the packet destination, 7018 only cares about dumping the traffic to AS701 as quickly as possible (closest exit).
If the packet originates in Los Angeles on 7018 and the destination is on 701 in Los Angeles, but the closes 7018-701 peer is in Atlanta, then the packet is going to go to the Atlanta peering point, exit 7018 for 701, then be routed back to Los Angeles over the 701 network.
Most tier 1 backbones are well peered, meaning there are "close" exits. In my example above 7018 and 701 are peered in LA, so there is very little latency going from 7018 to 701. However, there are smaller providers that are not well peered and my example traversing the country to get to the closest exit is fairly common. LA to ATL is about 50 milliseconds (ms) round trip delay (RTD), so the traffic forced to take that route has another 50 ms delay added to every packet. It doesn't seem like a big deal, but for interactive traffic like a terminal session it can be highly noticeable in application performance and responsiveness vs a local connections. Then if you add a few packet drops and TCP needs to re transmit and sync, IT is going to get a call from an end user saying, "why is application X so slow today?".
Latency is bad for application performance and packet loss is devastating. Keep the packet path as short as possible on high quality well peered backbones.
