Consensus is steadily growing when it comes to networks: it’s the next big thing in AI infrastructure. Having spent the last few years oogling at shiny graphics processing units (GPUs) and the recent reinvention of the central processing unit (CPU) as an agentic orchestration powerhouse, the thing keeping these connected has often been left to the wayside.

But no longer, it seems. Nvidia, through Spectrum-6, has been putting networks front and center of its next-generation Vera Rubin platform. And AMD’s Helios rack-scale platform boasts bandwidth levels that boggle the mind.

But that’s inside the data center.

Now the telecom players are getting in on the action, as without their fiber infrastructure, all the raw performance metrics in the world would matter for naught.

“Data centers by themselves, without an interconnected network, are just buildings with a lot of expensive equipment in them,” Verizon Business Chief Product Officer Scott Lawrence said.

In an interview with SDxCentral, the CPO warned that rethinking connectivity, “the network becomes a bottleneck,” even as data center capacity is set to double in the U.S. over the next five years.

Lawrence cited Omdia forecasts that AI network demand will grow at a 120% compound annual growth rate (CAGR) between now and 2030, with a flood of AI traffic set to hit Verizon and other service providers like a freight train.

And that traffic isn’t just human either. The rise of agentic AI – systems capable of performing tasks with little to no human intervention – represents the means to push traffic even higher. Since those agents rely on inference workloads processing vast amounts of data in real-time, demands on the network look nothing like the traffic patterns carriers have spent decades optimizing for.

Lawrence pointed to a broader shift already underway inside data centers, with intra-cluster traffic tied to AI training more than doubled every six months for the past two years.

“We're starting to see that shift toward AI inferencing, and even [Nvidia CEO] Jensen Huang at GTC in March commented that the inferencing inflection point has arrived," Lawrence said. "With that, you're going to see a more distributed compute model because you're going to need that distributed compute not only to serve specific use cases and industries but also to support agentic sprawl.”

That shift changes what carriers like Verizon actually need to build for. Lawrence suggested a single inference prompt alone can require between one and 10 Mb/s per user. So multiply that by the roughly one billion weekly active ChatGPT users, and suddenly the scale of what's hitting the network becomes stark.

The Verizon Business exec argued that new metrics are needed for key units of value to optimize for when it comes to networks beyond solely bandwidth.

Among these “network currencies” are uplink capacity, which Laurence said is becoming more important as agents push data back into the network rather than just pulling it down. Another potential focus pertains to network slicing, a playbook straight out of the carrier playbook that lets providers carve out dedicated capacity for specific workloads. And another is latency-sensitive service level agreements (SLAs) like time-to-first-token – or how quickly a network can return the first piece of a model's response.

But the network currency that players like Verizon are well placed to take advantage of is distributed compute. Not a new concept by any means, but given power constraints, the idea is gaining traction in the data center space beyond solely campus-to-campus connections. Suddenly, having disparate facilities separate by miles but working as one consolidated compute stack seems possible, and only so through effective fiber buildout.

Again, this edge play is nothing new. Around a decade ago, multi-access edge computing (MEC) was the next big thing: placing compute and data storage directly at cellular base stations and local network edges. In Lawrence’s own words, Verizon was “the OG of MEC,” but the concept quickly went the way of the dinosaurs.

Lawrence told SDxCentral that the carrier “learned a lot” from its work on MEC. That work has helped feed into Verizon's AI Connect strategy, which integrates network transport, edge compute, and GPU infrastructure into a kind of AI-fit architecture. And while it’s been learning, a lot of the use cases hyped over a decade ago are now beginning to come to the fore.

“A lot of the use cases that we saw at MEC are showcasing themselves now with data sovereignty or regulation … [like] gaming or gambling that needs to stay in a particular geography and requires that hyper-local compute," Lawrence said.

But the resurgence of all things distributed comes with a caveat in the exec’s view in that the edge shouldn’t be seen as a single place anymore, and treating it as one risks missing where the real opportunity sits.

“I think it's important to define the edge right, because the edge can be many different surfaces,” Lawrence said. “The edge could be on device, it could be on a customer prem, it could be a service provider edge, it could be in the service provider core, in our radio access network (RAN) environment, or it could be in a cloud environment.”

When it comes to AI, that multiplicity is tied directly to a bet on where the device ecosystem is headed.

“If we really believe that this smartphone is going to be the same device we're going to be using in five, 10, 50 years from now, I think that's a very unrealistic expectation,” Lawrence said. “In fact, many people are already calling that for the end of the smartphone, or the end of the era of the smartphone, because AI is going to bring new devices to the market that are inherently AI native and increasingly agentic: acting, deciding, and engaging autonomously.”

Betting on device unpredictability in Lawrence's telling is exactly why the network ends up as the connective tissue holding the next phase of AI together and not the data center or the shiny chips they house.

“It's not just about faster networks,” Lawrence concluded. “It's about intelligent proximity and monetizing relevance instead of just bandwidth.”