It’s Friday, after a long week, and I should resist the urge to put pop references into this unwieldy tale; let’s try and hold it together a while longer. Yesterday, in this blurb, we sketched a telco architecture for the future AI economy, geared for some kind of tiered cloud-to-edge inference proposition. Today’s news flow, as RCR has it, asks a simpler question: is this even a market, or just a build-out? I mean, of course it’s a build-out, and the ideas are speculative. But they’re what we like. The bricks and mortar, cables and ducting – I mean, that’s just marketing numbers. 

There is a bit of that today. Nvidia’s ‘AI grid’ – is it even Nvidia’s concept, or is its voice just the loudest in the room (planet)? – is being filled in piece by piece, from cloud to edge, wireline and wireless, via hyperscale interconnects, global longhaul systems, metro transport networks, regional RAN deployments, and private enterprise setups. The whole telecoms sector is turning on a dime; March 2026, when it all came out – an industry-wide move to converge a single AI-optimised infrastructure layer. Nvidia says jump, and the telco industry digs in its pockets, says how much? 

For mobile operators, this feels like a unique inter-generational infrastructure overhaul, which will straddle the 5G and 6G eras, and reach way behind-the-RAN into their backhaul and longhaul fiber systems. Nokia’s announcements about application-specific optical networking reinforce this: the transport network is being re-engineered for AI traffic patterns, not human ones. Sulagna’s write-up from a strategy briefing at OFC is worth reading. Ciena, now as much of a rival for the Finnish firm as Ericsson (in the AI networking narrative), is also working on the back-end.

It has just set a speed record with Meta (800 Gbps) across 16,608 kilometers of trans-pacific submarine cable, heralding massive new capacity to shuttle model outputs between data centers and edge domains. So yes, it is a building project. But the Telefónica stuff shifts the conversation from placement of compute to extraction of value – with automation, exposure, slicing, orchestration as monetization levers. Which aligns with yesterday’s papers (can’t help myself), per the discussion about AT&T and Verizon, and telcos as a “middle layer” in the AI supply chain.

Their role is not just to host inference workloads, but to operationalise them – on programmable telco networks, in programmable telco products. The risk for telcos is the infrastructure is built faster than the business models; and the risk for everyone is that the use cases don’t show up. (There is a-whole-other narrative about how clever/stupid AI really is, at least in its present form; at least enough to build a new global economy on, in the medium-term.) Build it and they will come (another pop cliche). Are you in, or are you out? Carpe diem, says Jefferson. 

The Siemens story, about private 5G, provides a pragmatic enterprise counterpoint to this hurried telco‑centric AI grid narrative. The tech isn’t the hero of this story, or any story; it needs to go to work, and work is difficult, idiosyncratic, self-absorbed, conservative. 5G is an enabler; so too is AI – despite the clamour. Use cases count, and physical AI is already alive and kicking. Another thing: on‑prem control (and on‑prem LLMs) might just knock the tokenomics off its stride. Siemens had an MWC demo with just such a setup; the on-prem directive is clear in the industrial economy. 

Security, determinism, reliability; all of that. Which is a different value prop, versus programmable latency tiers in the macro RAN. While telcos are busy building infrastructure, serious industrial adopters want self-contained systems, on their terms – to keep data close and secure, and cheap (versus public toll roads). Of course, most large-scale AI workloads – and parts of all workloads – will rely on telco edge nodes, metro orchestration, longhaul fiber optics (maybe even AI-RAN) for training, scale, collaboration. But it is a reality check – if only about how varied the work is.