Do we have something, here, more coordinated than just a burst of AI enthusiasm from AT&T – and all the rest of the telco pack, over the last several months? MWC has just been, and so too now has GTC, and just maybe a playbook is emerging from all of the noise about how telcos might reposition themselves in the AI value chain. AT&T’s latest work with Cisco and Nvidia is a signal about embedding AI inference into the network edge, and a shift from connectivity as transport to connectivity as execution – where networks don’t just move data, but act on it. 

Parallel alignments with AWS in the metro, Ericsson in the RAN, and Microsoft at the edge suggest something more deliberate: a disaggregated, multi-partner AI platform spanning the full network stack. At the same time, Nvidia’s push into AI-RAN with Nokia and T-Mobile extends the logic into the radio layer itself. The cell site, long treated as a cost centre, might become a distributed compute node – the logic goes; part of a wider “AI grid” designed to run inference closer to the action. Jury’s still out on AI-RAN, but it’s also not going away. 

You might remember RCR caught up with Verizon (Business) at MWC, which had lots to say about all of this. It brings clarity, and some constraint. Where the rest are sketching a vision of a distributed AI compute fabric – in press notes, anyway, until RCR speaks with them – Verizon talks about a three-tier system combining centralised training, metro/long-haul orchestration, edge inference. This is probably the cleanest articulation of the architecture. Its emphasis on programmable latency tiers (edge, metro, cloud) hints at an engineered hierarchy with trade-offs.

The message aligns with AT&T’s macro-edge strategy and Nvidia’s “AI grid” concept: telcos are the middle layer, linking training and inference, placed according to workload requirements, bandwidth permissions, and stuff like security and sovereignty. A deeper economic logic is emerging, too. Jensen Huang’s framing of AI around tokens as a unit of value – produced in data centres, consumed by applications; discussed here yesterday – implicitly introduces the notion of latency as a pricing variable. Not all tokens are equal: some are for real-time decisions.

So programmable latency tiers – as articulated directly by Verizon, and indirectly by AT&T – moves beyond engineering into economics, where each layer (attached to a relevant compute engine) maps to a different class of token. So, the network also shapes how and where tokens are delivered – and therefore how they are monetised. The architecture looks like a regular supply chain: data centers as token factories, edge environments as token dispensaries, and the network – particularly the metro layer – as the quality-of-service engine in between. 

Something like that, anyway. But the question is whether telcos can make a business out of it, by translating a functional role into a commercial one – by turning latency and determinism into billable traits, or whether they just carry increasingly valuable token flows without capturing much value. This is going to be the story from this point; same as it has been to this point. It is up to them to change the narrative. Credit to the US telcos, though; they are closer to it, but they also have a lead on it.