Why telcos won't see AI in their bandwidth

The aha moment of the conference came in a single sentence. 

A senior telco executive told a packed FutureNet World 2026 audience that his team doesn't see AI in their network. Their proof: bandwidth consumption hasn't moved. The increase, he said, is “almost irrelevant.” 

That comment captures the most important misconception in today's telco AI conversation. Not because the speaker was wrong about the data - bandwidth genuinely hasn't surged - but because the metric itself is the wronginstrument for the question being asked. 

If you measure AI by the bandwidth it consumes, you will conclude that AI isn't happening in your network. And you will be wrong. 

The provocative truth is that AI is, structurally, less hungry for bandwidth than what came before it. 

A modern large language model with around 200 billion parameters fits on roughly 200–400 GB of disk - small enough that essentially the entire body of human textual knowledge sits comfortably on a single drive. NVIDIA'sneural rendering, introduced at CES 2025, achieves an 8× compression factor for visual scenes. Text-to-image models routinely encode photographs in a few bytes of prompt - compression ratios on the order of 20,000:1. Generate a 3D world from a single still image and the ratio gets steeper still. 

The era when bandwidth was the asset to sell is, in structural terms, over. AI doesn't show up as more pipe. It shows up as more intelligence at each end of the pipe. 

If bandwidth is the wrong meter, what is the right one? 

I'd argue it's this: how much intelligence does your network carry as a primary asset - not as a side-effect? 

I use the word “intelligence” deliberately. AI is a category, not a thing - it covers thousands of distinct technologies, and lumping them together obscures more than it clarifies. “Intelligence” is what the customer actually buys: theability to interpret, decide, and act inside a flow that the network already moves. 

Voice is the most underrated example. The telco industry talks about voice as a declining modality, a legacy revenue line. That framing is exactly backwards. Voice is the richest interface humans use - it carries intonation, intent, urgency, hesitation. The entire AI ecosystem outside telco is moving toward voice: developers are talking to their tools instead of typing at them; agentic systems are converging on voice as the universal control surface. In fiveyears, the keyboard will be a niche peripheral. 

And the asset that decides whether voice AI is a $1 commodity or a strategic capability is location. If voice AI sits in an app on the phone, anyone can build it. If voice AI sits inside the SIP layer of the network, only the operator can. 

Which brings the strategic question into focus. 

Will telcos become wholesalers of intelligence - embedding voice intelligence, video segmentation, spatial reasoning, action control directly into the substrate of the network - or will they remain the pipe? An ever-better, moresecure, more privacy-aware pipe, but in the end, just a pipe. 

The honest answer is that becoming a wholesaler of intelligence isn't a technology decision. It's a trust decision. 

This is where my FutureNet panel closing word came in. When the moderator asked each of us to give one word that would decide AI in telcos in 2026, mine was Trust. 

Because here is what I see again and again across enterprise pilots: the technology works in the demo. It moves intelligence through the network. It does the thing. And then procurement asks three questions: 

• Can the agent's decision be reconstructed when a regulator asks? 
• Will the agent honour an SLA, with a contract-grade failure mode? 
• Will the same input produce the same decision, every time, so we can defend it under the EU AI Act? 

That triangle - auditable, deterministic, SLA-bound - is what trust actually means inside a telco PO. And it is exactly the gate where most agentic deployments fail today. 

This is the gap we've been building Tallence's deterministic-core agent orchestration system to close. LLMs propose; a rules engine decides; every decision is logged, reproducible, and auditable by default. Telco-grade governanceon day one - not retrofitted after pilot. Voice intelligence in the SIP layer, alarm-correlation intelligence in the OSS, billing-dispute intelligence in the BSS - each one a deterministic module that procurement can sign for. 

If you're a telco leader thinking about whether the next decade of revenue comes from a bigger pipe or from intelligence inside the pipe, the answer isn't in your bandwidth dashboard. It's in a different question altogether: how much of our network's value is now packaged as intelligence - and is it ours to sell, or someone else's? 

That's the question I would have liked to spend the whole panel on. Happy to spend it with you instead. 

DM me, or visit tallence.com. Let's talk about what intelligence in your network would actually look like.