- AI inference is creating a new role for distributed carrier infrastructure
- Connectivity gets operators into the game, but compute, security and control could move them up the stack
- AT&T, Cisco and NVIDIA agreed carriers have an opportunity – but need to move now
For the best part of 30 years, telecom operators have been trying to solve essentially the same existential problem: How do you move up the stack before somebody else captures all the value?
Cloud didn't solve it. 5G hasn't solved it (yet). And AI could simply become the latest enormously valuable technology that carriers connect while somebody else makes the real money.
So on Tuesday, FNTV put the question to Masum Mir, SVP and general manager of Cisco (Nasdaq: CSCO) Provider Mobility; Shawn Hakl, SVP and head of product at AT&T Business (NYSE: T); and Chris Penrose, global head of business development for telco at NVIDIA (Nasdaq: NVDA).
That's the room you want if you want a real answer to our industry's defining question. These are not AI washers—they are the AI vanguard standing on the front line of the carrier battle for relevance. Cisco provided the architecture and platform perspective. AT&T brought the operator and commercialization perspective. NVIDIA sat at the heart of the AI compute and developer ecosystem.
And while they approached the problem from different directions, a remarkably coherent answer emerged.
AI may give carriers a better opportunity to move up the stack than previous technology cycles – provided they move now.
Inference changes the infrastructure equation
The starting point to our conversation was a recognition that there has been a fundamental change in AI itself.
The first phase of the AI boom was dominated by enormous centralized AI factories, initially driven by the need to train increasingly large foundation models. But inference – putting those models to work – is now becoming the dominant workload.
That changes where compute needs to live.
"We really tipped over this year with inferencing now becoming the primary workload and putting these models into actual use," said NVIDIA's Penrose.
The question now, he said, was which workloads were best served centrally and which made more sense in distributed environments.
Service providers could potentially deploy smaller compute footprints using existing land, power and physical infrastructure and then orchestrate workloads across multiple locations.
That could be driven by economics, but also by requirements including sovereignty and latency.
"We want to make sure we can land on the right infrastructure at the right performance at the right cost," said Penrose.
AT&T's Hakl saw the same shift from the customer side. AT&T surveyed more than 1,000 customers and found nearly three quarters already operated multiple AI systems, while more than 60% were working on multi-step agent problems.
More than 49% were undertaking some form of real-time AI project involving content such as video or audio.
In other words, AI was already escaping the centralized data center.
Connectivity gets you in. It doesn't get you paid twice
That distributed architecture inevitably increases the importance of the network.
Cisco's Mir was unequivocal: "AI will drive network demand. It is happening."
But simply carrying more AI traffic doesn't solve the industry's fundamental problem. It could leave carriers doing precisely what they have done during previous technology cycles: investing heavily in infrastructure while value accrues somewhere above them.
The opportunity discussed Tuesday was to combine connectivity with compute, inference, security, orchestration and enterprise services.
Mir argued that carriers could provide secure, trusted connectivity for distributed AI workflows and then add curated inference applications on top, allowing them to participate directly in enterprise and public-sector AI spending.
Hakl's position was more pragmatic.
"Connectivity pays the bills and it's not a bad thing," he said.
But providing ubiquitous connectivity also meant operators had "earned the right to come up the stack," adding capabilities around AI security, optimization, intelligence and control.
And there may already be measurable value available.
Hakl said AT&T was seeing potential savings above 45% from intelligently placing inference requests, while time to first token could improve by between 20% and 80%.
The network has to become consumable
There was, however, a familiar telecom trap waiting for operators.
Carriers can't build another closed ecosystem and expect the AI developer community to come to them.
Hakl argued that network capabilities needed to become programmatically consumable through developer ecosystems that already exist rather than forcing developers into carrier-specific environments.
"The developer ecosystem is going to land where they already exist," he said.
That requires carriers to expose network capabilities consistently and at scale.
Mir added another problem: speed.
Telecom operators have traditionally worked through five-to-seven-year technology cycles. AI isn't going to wait that long.
His prescription was to identify a handful of enterprise verticals, start building services now and work backward from genuine customer requirements.
"Act now," said Mir. "Focus on targeted few enterprise verticals, start to build this service and start to walk backward from the customer need."
But he also cautioned against reverting to bespoke telecom engineering. Those services needed to be built on platforms that could subsequently scale.
Who controls it?
Moving up the stack also raised a more fundamental question: If compute, connectivity, cloud platforms, enterprise infrastructure and AI agents are increasingly interconnected, who actually controls the resulting infrastructure?
The answer wasn't NVIDIA, Cisco or AT&T.
"The customer needs to be in control," said Hakl. Refreshing!
What that meant varied by customer. Governments and large enterprises might require direct control over individual components, while smaller organizations could consume much more of the infrastructure as a service.
But the underlying requirement remained the same: enterprises needed visibility and control without requiring one supplier to own the entire environment.
Mir similarly argued that enterprises needed sufficient control over services even when capabilities were built, curated and operated by a service provider.
Penrose positioned NVIDIA differently again. It would provide developers and partners with the tools and capabilities to build and orchestrate AI infrastructure rather than attempt to own "a big control plane."
This question of control is becoming increasingly important as telecom, cloud, AI, energy and industrial infrastructure converge. I've described the resulting architecture as the Unified Infrastructure Stack, with operational sovereignty – who ultimately controls the infrastructure – becoming as important as who owns or hosts it.
Then physical AI changes everything
The strongest argument for distributed infrastructure may ultimately come as AI moves out of the chatbot and into the physical world, an argument I've made repeatedly as the AI economy has developed.
Factories, vehicles, robots, cameras and industrial machines create a fundamentally different infrastructure problem from asking an LLM to generate a PowerPoint.
AI controlling a physical object may have to process sensor, audio and visual information and make decisions in real time. Sending everything to a distant AI factory isn't always practical.
Mir said that raised the requirement for network trust and security "one notch up."
"Zero trust is going to become a mandate across anything that we connect," he said.
Hakl said AT&T was already seeing the beginnings of this world. The carrier has worked with Cisco and NVIDIA on an oil and gas deployment using cameras for applications including asset tracking, worker safety and security.
And AT&T's enormous connected-car business already provided another example of highly distributed infrastructure.
This is also why AI-RAN could give telecom a second chance to own part of the AI economy. The really big prize isn't simply making the existing network more efficient. It is extending distributed intelligence beyond the traditional telco edge into factories, transport, energy, robots and the rest of the physical economy.
None of this means carriers automatically win.
They still need products enterprises want, platforms developers can consume and services customers will pay for.
But Tuesday's conversation suggested something important has changed.
Carriers have connectivity. They have distributed infrastructure. They have enterprise relationships, security expertise and experience operating critical systems. Increasingly, AI needs precisely those things.
The existential question is therefore changing.
It's no longer whether carriers have a role in the AI economy.
It's whether they can move fast enough to turn that role into money.
Further reading
AI-RAN could give telecom a second chance to own the AI economy – Why distributed AI infrastructure could turn telecom assets from cost centers into revenue platforms.
The FNTV 2026 Global AI Economy Index – A broader framework for measuring who can turn AI into productive economic capability rather than simply accumulate GPUs and models.
Introducing: The Unified Infrastructure Stack – How telecom, cloud, AI, energy, cybersecurity and industrial infrastructure are converging into a single operational fabric.
Why AI speed – not scale – will define the next global digital economy – Why inference latency becomes increasingly important as AI moves from centralized model building into real-world applications.
Sovereignty is not a place – Why control, rather than geography or infrastructure ownership alone, is becoming the defining test of digital sovereignty.
Stephen M. Saunders MBE is a communications analyst and USPTO-registered inventor examining how digital infrastructure — 5G, cloud and AI — is reshaping industry, power and society, as well as underpinning the emerging, ubiquitous global digital economy. As anchor of FNTV and a longtime industry insider, he focuses less on growth narratives and more on execution, risk and how hyperscale technology is distorting markets, governance and society at scale.
Opinion pieces from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.