Researchers are pointing out an awkward mismatch: the mobile networks carrying data to and from phones were not built with AI training in mind, and that's starting to show.
The issue is federated learning, where a large language model gets fine-tuned using data scattered across many phones instead of one central server. A new arXiv paper argues that when this happens over mobile networks, the updates each phone sends back arrive at wildly different times because of wireless conditions, device movement, and hardware differences. Even though all those updates belong to the same training round and are headed to the same destination, the transport network treats them as unrelated traffic. The researchers propose that cell towers should aggregate the updates in-network, turning scattered device-level transfers into fewer, predictable bulk transfers that networks can actually plan around.
Why it matters: optical transport networks work best with predictable, schedulable demand, not a trickle of unpredictable device traffic. If this approach holds up, carriers could reserve high-capacity optical paths only during these aggregate transfer windows and release them the rest of the time, rather than over-provisioning capacity that mostly sits idle. That's a meaningful efficiency argument as more AI training work moves to the network edge instead of centralized data centers.
It's also a reminder that the industry's AI infrastructure story usually stops at GPUs and data centers, when the plumbing connecting edge devices to anything useful is just as unsolved. This is a proposed architecture, not a deployed one, and the usual caveats about asynchronous, flaky mobile connections apply.