How mobile internet traffic travels across networks
Tap a link on a phone, and the request rarely takes the short route. It crosses a radio interface, lands in a carrier's core network, gets rewritten, and often travels hundreds of kilometres before reaching the server that answers it.
Mobile networks moved 210 exabytes of data per month in the first quarter of 2026, up 22% year over year. That volume runs through an architecture that behaves nothing like home broadband, and the differences show up in latency, in geolocation accuracy, and in how websites decide whether traffic looks trustworthy.
The first hop is a negotiation, not a connection
A phone doesn't simply join a network. It authenticates using SIM credentials, gets attached to a specific cell, and receives scheduling slots on a radio interface shared with every other device in that sector.
Signal strength, tower load, and frequency band decide what happens next. A handset on a congested 5G sector inside a train station can perform worse than one sitting on a quiet 4G band two kilometres away, which is why speed tests taken outside peak hours tend to flatter the network.
But the radio hop is short. Real routing work begins once packets leave the base station.
Carrier NAT scrambles the picture
Operators don't hand every subscriber a public IPv4 address. There aren't enough addresses left, so carriers run large-scale NAT and pack hundreds (sometimes thousands) of customers behind a single public IP.
That quirk sits underneath what is mobile proxy routing: traffic leaving a cellular IP inherits the reputation of everyone else sharing it. Blocking one address means blocking an entire neighbourhood of paying subscribers, so most platforms won't do it lightly.
The IETF reserved a dedicated address block, 100.64.0.0/10, for exactly this purpose in RFC 6598. A phone's internal address usually looks nothing like the one a website records, and that mapping can change mid-session when a device moves between gateways.
Tunnels do the actual carrying
Between the tower and the internet, packets don't travel as plain IP traffic. They're wrapped inside carrier tunnels using GTP, a protocol built so a session survives when a subscriber drives from one cell to the next.
Those tunnels terminate at a gateway, and gateways are centralised. A carrier serving a whole country might route everything through two or three breakout points, meaning a phone in Porto and a phone in Lisbon can both appear online from Madrid.
Handover makes the design necessary. A subscriber on a motorway can pass through a dozen cells in ten minutes, and without tunnelling, every one of those transitions would drop the connection and force applications to reconnect.
This is why IP-based geolocation gets mobile traffic wrong so often. City-level targeting works reasonably well for fixed broadband; on cellular it's frequently off by several hundred kilometres, and databases like MaxMind carry that error forward into every service that trusts them.
Latency is mostly geography
Radio access adds maybe 10 to 30 milliseconds on a healthy 5G connection. The higher cost comes from the detour through the core network, then from wherever the destination server actually sits.
Content delivery networks fix part of this by caching near the breakout point, which is why a Netflix stream feels instant while a query to a small regional API crawls. Video already accounts for roughly 75% of all mobile data traffic, and Ericsson's quarterly traffic data shows uplink volumes now growing faster than downlink for most operators.
Consistency matters more than peak speed for most workloads. A connection averaging 40 Mbps with stable latency will beat one that spikes to 300 Mbps and then stalls whenever the sector fills up.
Uplink growth changes the engineering problem. Networks tuned for a decade around downloading video are being asked to handle photo backups, video calls, and cloud sync, all of which push data the other way.
What comes next
Edge computing and 5G standalone deployments are slowly pulling breakout points closer to users, which will trim the geographic penalty. Network slicing goes further by giving specific traffic classes their own guaranteed path, though commercial offerings are still thin outside a handful of markets.
None of that removes the shared-address reality. As long as IPv4 scarcity forces carriers to pool subscribers behind common gateways, mobile traffic will keep carrying a collective identity, and anyone building tools that depend on how the internet sees a connection should plan around it rather than fight it.
