Agentic AI set to transform network design and economics

A new report by Cisco on the impact of AI on wide area networks estimates that by 2035 AI inference will constitute one quarter of all network traffic.

Nine years is a very long time in the AI era, and the report extrapolates its projection from today’s low (and unspecified) AI inference traffic rates. Nonetheless, the question of how to design networks today so that they (profitably) accommodate customers’ future AI usage is both critical and unresolved, as speakers at Telecom TV’s recent DSP Leaders World Forum event made clear.

Cisco’s report states that “agentic AI fundamentally amplifies network usage”. Describing AI agents as “network power users”, it calculates they generate up to 450% more total traffic per task compared to being carried out by humans, with AI inference making up approximately 70% of that traffic.

Through a glass darkly

Telco’s dense fiber and mobile networks in urban and business districts promise them a critical role in transporting data traffic generated by AI computing.

“We're well positioned as telcos for inferencing,” said Gabriela Styf Sjöman, Managing Director Research and Networks Strategy, BT Group, speaking at the Telecom TV event. "We're deeply distributed... from a metro perspective; we have our POPs, we have an architecture that's ... prepared us.”

But telcos have little visibilty of how agents will function at scale on their networks, as Andy Linham, Principal Strategy Manager, Vodafone Business, pointed out.

“There's a lot of talk around agentic services and agentic commerce. But how those agents are going to discover each other, how they're going to identify, authenticate, authorize what each one of them is going to do, that's a massive unknown that does change how my network needs to operate,” said Linham.

A major challenge telcos face is transitioning from building and sweating physical network assets over a 10 to 20-year period to implementing rapid, end-to-end network lifecycle upgrades, according to Styf Sjöman.

“The biggest challenge we have is that AI is fundamentally changing the economics and architecture of the network, and in particular the lifecycle of how we upgrade our networks today ... to make it end-to-end ready and to have no bottlenecks,” she said.

Managing for now

As things stand, today's networks can cope with customers' AI demands, said Linham.

“We've got the capacity, we've got the engineering, we've got the access networks now. So, I don't think today we have a problem. I think the problem comes in the future when we have to work with customers that aren't human beings that will sit there spinning GPU cycles that are very expensive, and that's the latency, that's when the performance really kicks in. I think we're going to have to work out how we deal with that in a really, really rapid way, much faster than we've evolved previously.”

AI training and inference, for example, present very different use cases for networks.

“I need to have super highways, but I also need ultra-low latency, deterministic smart transport, and I'll probably need to ...[ be] congestion aware,” said Styf Sjöman.

The extent to which customers use huge general-purpose language models and very small, domain-specific language models will also shape network usage and invetsment.

“If it's a small model, I can probably run it on my device, or ... on a ... server on the customer site, or on the CPU. If it's a bigger model, I want to run it at the edge," said Vodafone’s Linham. "Or do I have to put it in the cloud because I need that sort of capacity? These are things that I'm less sure about today."

Linham continued: “Flexibility becomes the most important thing, because I need to be able to change how my architecture flows, the traffic flows .... really quickly."

As for whether telcos' investment in optimizing networks for AI traffic will generate new sources of revenues, much depends on how intelligently they can use data.

“We're going to the coordination era, away from the information era. The network is just a super sensor, that's the number one thing ... generating lots and lots of data,” she said. “You need that intelligent layer on top of that, and then you need the applications."

It means that future network "infrastructure needs to be intelligent,” said Styf Sjöman. We have a lot of telemetry in the network, we have lots of data already today. We're not doing a lot with it, especially on end to end, but that's the opportunity."