Lumen’s Rob Ward: 'The AI factory has no walls'

Rob Ward, Lumen, at Broadband Nation Expo
In a keynote at Broadband Nation Expo about connectivity’s new job description, Lumen's network engineering chief said that moving data, not just computing it, is now AI’s challenge. (Photo via Fierce Network)
  • Silicon is going to continue to be a constraint, but Lumen's Rob Ward thinks power will be a growing constraint
  • How AI sovereignty will be implemented is not settled, he said
  • Finally, AI means that the way networks are sold must change

BROADBAND NATION EXPO, NEW ORLEANS – The defining challenge of the AI era is no longer compute alone, Lumen Technologies’ Rob Ward told the keynote audience at Broadband Nation Expo, but how data moves between data centers, clouds and the edge.

Ward, who is SVP of global network engineering at Lumen, said the familiar image of AI as rows of racks and accelerators, “maybe even the eerie green light that gives a little sci-fi feel,” captures only part of the story. “It’s a critical component of AI within the data center community, but it’s not the whole picture,” he said.

The constraints are shifting, too. “Silicon is going to continue to be a constraint, but I think a growing constraint will be power,” Ward said. “How much power can we get into this facility to support those rows of rows, and where does that exist?”

Looking farther out from a single facility, he said, and models, data, users and inference all live in different places. “That AI factory that sometimes we think about, it really doesn’t have walls. It’s going to be very distributed.”

He backed that up with survey data. A Cisco and Omdia study of roughly 1,200 enterprises found that 96% run infrastructure across a mix of cloud, edge and on-prem, and 97% plan edge inference within about a year. Ward noted the group surveyed skewed toward heavy AI adopters, “so they’re sort of leading the curve.”

Meanwhile, a Deloitte survey of about 500 large enterprises found 70% expect to scale their AI factory out to the edge, he said.

Lumen's Rob Ward said, "That AI factory that sometimes we think about, it really doesn’t have walls. It’s going to be very distributed," as he walked attendees through the data on the above slide. (Photo via Fierce Network)

Different workloads, different demands

Interactive inference, Ward said, needs to sit “reasonably close to that user,” while synchronized training cares less about absolute latency than about consistency, and “you want big pipes.” The ideal network would offer “the biggest pipes, the lowest latency, lowest loss,” he said, but cost and geography force trade-offs.

Sovereignty may be the first requirement to settle, he added, and the industry is not there yet. “I don’t think we have a good agreed-upon definition for what that really means. Is it a country boundary? Is it a region?”

Ward said the way networks are sold must change. Customers have long bought a circuit between two endpoints, with humans in the background designing and implementing it. “AI really changes that model,” he said. Instead, a customer, or a customer’s agent, will define the desired outcome, “and it’s up to the programmable network to create that outcome for them.”

That outcome must also be continuously assured. Locations and workloads shift as models retrain and data moves. “Those are dynamic, and those are moving at machine speed,” he said. “We need a network at machine speed.”

Physical innovation under pressure

Beneath the software layer, demand is driving a rethink of the physical network. “For as long as I’ve been in the business, I haven’t seen this rapid acceleration,” Ward said. “It’s that scale that is driving really the need for a new physical architecture.”

He pointed to work on bringing optics closer to silicon to raise density and make equipment “more consumable within a data center footprint,” and to what he called a “resurgence or rebirth of optical innovation.”

Five years ago, Lumen was lighting one fiber at about 20 terabits roughly every 18 to 24 months. Two years ago it was lighting two fiber pairs at 50 terabits each every six to nine months. “It’s really hard to keep up with for infrastructure that we deployed and planned 20 years ago,” he said.

Newer equipment has delivered “a 12x space improvement,” Ward said, “taking 12 racks down to one in these very small locations,” while cutting power in half. Multi-core fiber, which puts four optical cores in one strand, and hollow-core fiber, which could mean significantly lower attenuation, are both being explored. “Really some interesting things playing out here that help us deal with scale,” he said.

'Machine speed, human judgment'

Ward described a cycle of turning intent into possible solutions, enforcing them, then verifying and adapting. Energy-aware orchestration is one example, shifting workloads based on power availability and cost, with a network that can adapt in near real time.

On autonomy, he drew a line. “We want to be able to remove the human latency from that equation, but not the human judgment,” he said. Using the TM Forum’s zero-to-five autonomous network maturity scale, he said, “We don’t have aspirations for Level 5. We have aspirations for Level 4.” The goal, he said, is “keeping that expertise in the loop.”

Lumen’s agents are meant to multiply existing know-how, not replace it. “Our view is not that it allows an uninformed person to execute on something; it’s that we can multiply the experience and the network know-how that we have,” Ward said.

Security, he said, deserves its own session, but he offered a framework for agents: “Do we know who you are and who is responsible for you? Do we know what you’re capable and approved to do? Do we have traceability and audit? And then, can we stop you?”

Ward ended his keynote by tying the threads together: “The AI factory has no walls, but it still needs that foundation,” he said. “We believe that foundation is the programmable network.”

Read more about Lumen on Fierce Network

Lumen expects $1B to $1.5B from copper in decommissioned networks

Lumen wants to use the network to help AI workloads chase power

AI is breaking bandwidth forecasts. Lumen wants enterprises to scale internet on demand