Lumen sees the biggest AI networking opportunity between data centers and AI factories, but admits inference will also spread to the edge, leaving the future balance between east-west and north-south traffic unresolved.
In sum – what to know
East-west bet – Cloud-to-cloud traffic is growing 10-13% annually, notes Lumen – versus 1-2% for traditional prem-to-cloud, as AI workloads spread across data centers and clusters, and get staffed with agents.
Home offices – North-south traffic is well-served already, and AI demand is switching to home offices; Lumen’s billion-dollar PCF deals monetize its fiber infrastructure, while Alkira provides the programmable layer.
Edge wildcard – Industry 4.0 could push inference into factories and edge locations, but Lumen says many will use existing networks; it also acknowledges how physical AI exists between the cloud and edge is “the unknown”.
Continuing from yesterday, then (and rushing a little before holiday) – when we discussed highlights of Lumen’s business turnaround (‘strategic revenue’ rising 14%; its NaaS base doubling to 3,000-plus), and its twin strategy to build-out physical fiber optic infrastructure (40-odd million new strand miles; 400G for 90 percent of cloud on-ramps) and pull it into a new global control plane (via its $475 million Alkira acquisition) to serve big enterprises in the buzzy AI economy. Per the sign-off yesterday, Jim Fowler has more to say, and most interesting, perhaps, about where AI workloads will live, and how they will be served. But this will be shorter than yesterday’s post.
Part of the dynamic is this north-south / east-west divide, of course. Lumen is focused on the latter. Traditional enterprise connectivity – going ‘north-south’, premise-to-cloud – is hardly slowing, but it is well served by metro fiber already, and not growing very much either. A couple of times, Fowler, chief technology officer at Lumen, quotes IDC stats that north-south traffic is growing at one-to-two percent per year – on a commoditized service, to an extent. “Some of that’s going to be new connectivity, and some’s going to be capacity,” he says. “There’s just not much expected. Think about it. Companies aren’t building new buildings and locations. Most people work from home.”
He goes on: “I mean, you’re seeing home growth, north-south – which is not actually a part of that number. I said six months ago, the office space of the future is in a data center. It is not in a physical location downtown; it’s in these data centers – not in the facilities that used to house all the staff. Enterprise prem-to-cloud – there’s just not that much doing. The growth is going to come from where the agents sit.” There’s the kicker, and the question ultimately. To an extent, most of the answer is mapped out. The new ‘east-west’ AI superhighways – mostly between data centers and new AI clusters – is where the action is for fiber providers like Lumen. “Cloud-to-cloud is growing 10-to-13 percent,” says Fowler, quoting IDC again. “We’re focused where the growth is, and where the demand is.”
Pipe and platform
There is a useful distinction here between Lumen’s two plays. The $13 billion of so-called ‘private connectivity fabric’ (PCF) deals are mostly a way to monetize its physical network – dedicated, high-capacity connectivity for hyperscalers and other AI-infrastructure customers, using assets Lumen has already built. But they are contracts, and not booked revenue in the quarter or even the year; PCF deals generated about $91 million in revenue for Lumen last quarter. The other piece is the digital layer, now based on Alkira, where the economics are more interesting: digital revenue was $39 million in Q2, still small, but growing as Lumen adds customers and services around its global programmability pitch. So it is trying to monetize both the pipe and the platform on top.
A quick return to the discussion yesterday; Fowler positions these hardware and software pieces in this spiralling east-west mix. “Sometimes we need to own the fiber, so we will put more in the ground, or else do over-builds to provide the capability,” he says. “But with Alkira, we don’t always need to own the fiber anymore. Customers can bring the network, and we can provide the mesh capability to help them connect all of it together globally – to better manage routing and security, and to think about how inference and training work for their businesses. So there will be no less a focus from us on hyperscaler-to-hyperscaler; what you will see is more focus on helping enterprises plug into it.” Which sounds like a north-south exercise, but is less a retreat from the east-west thesis, as such.
It might involve a north-south connection at the edge, but the value proposition is broader – about how enterprises, once plugged-in, orchestrate east-west traffic between distributed global data centers, training clusters, and inference locations. The north-south gateway has to be open, and available, but the complexity Lumen wants to monetize is in the background, in the software layer behind, cutting across increasingly distributed AI infrastructure. “The control plane allows enterprises to get the most use of the infrastructure while optimizing and modernizing how their wide area networks run globally.” Fowler provided examples yesterday of how enterprises have rationalized their cloud footprints for faster routes and better performance, slashing redesigns from a year to a month.
Criss-cross comms
As well, he mentioned before that 53 percent of public internet traffic is “between machines, transacting back and forth”. These new east-west networks are being built, by all fiber firms, to connect distributed cloud sites that are increasingly staffed by agents. He says: “AI will grow at 100 percent CAGR through 2030, and that’s agent-to-agent comms and inference. That’s where most of the growth is going to happen – between data centers as more and more of the work is done by agents.” But inference, Jim: does that not get distributed down to the edge, and drive demand for capacity on north-south networks, too? “It’s both,” he responds. “It’s in the cloud, and it’s between.”
And then he hits upon Industry 4.0, to explain – and the conversation crosses into lots that RCR has been writing about with regards to private networks, fiber-to-the-factory, edge compute, domain models, and all the fun of IoT. Fowler goes on: “If you’re a manufacturing company, the edge component is going to sit in a manufacturing plant that’s generating data from an IoT perspective which most likely feeds back into multiple clouds that use the data to perform different business functions. And then it’s cloud-to-cloud as those functions talk to each other. The company recognizes inventory levels are going down based on its IoT data, and transacts to a supplier to order inventory. That’s where the growth is going to come from over the next five years.”
So don’t those industrial edge inference workloads, private and sovereign and secure, feed into some kind of north-south capacity crunch, driving growth higher than two percent? I mean, if all these humanoids are about to disembark, then you surely need new fiber backhaul from the 5G network in the factory? “That’s part of that one-to-two percent,” he responds. “That is in there.” He has an anecdote about a friend with an agent-based ordering service for drive-throughs. “A small version of the model goes in the store to do the processing, but you don’t need a new network line for that; they just use what’s there. Versus other companies, going the other way, which want to feed data out of their factories into a central model for real-time planning.”
He adds: “I don’t think we know exactly how that’s going to work itself out yet. That’s the unknown in that part of the design.”
