Why the edge finally has its killer use case

Connect (X) industry panel covers the transition to AI inferencing and infrastructure scaling

The edge has been a conference talking point for the better part of a decade, but a panel at Connect (X) made a pretty convincing case that the conversation has finally moved past theory. Moderated by Iain Gillott, WIA’s VP of Innovation and Technology, the session pulled together Sean Farney (JLL), Bhupesh Agarwal (Intel), Joe Constantine (Ericsson), Dr. Ozge Koymen (Qualcomm), and Jim Poole (American Tower) — a lineup that covered just about every layer of the stack, from silicon to towers to the real estate underneath the data centers.

WIA used the run-up to the show to announce its Edge AI Infrastructure Initiative (EAII), which is meant to bring member companies together to identify ecosystem opportunities and, in Gillott’s word, make edge deployment as “frictionless” as possible.

Rethinking the edge and finding the killer use case

If you’ve been to enough of these panels, you’ve heard the edge defined a dozen different ways. This one was no exception, but the panelists were perhaps a bit more honest about why earlier attempts didn’t stick. Poole said MEC was “a nail looking for a hammer, right? Like, the use cases just didn’t materialize even though we envisioned — and I think you’re right — the infrastructure correctly. We need this distributed compute layer. We just didn’t need it yet.” Koymen called MEC ahead of its time. Agarwal pointed out that the industry didn’t help itself by spending years arguing over far edge versus near edge versus network edge versus on-prem edge while enterprises tuned out.

What’s different now is demand, and the demand is AI inferencing. Koymen noted that user behavior is shifting away from downlink-heavy video consumption toward uplink-centric, AI-generated traffic, with agentic data on track to overtake human-generated data within a couple of years. Constantine cited Ericsson Mobility Report figures projecting global data traffic will triple between 2023 and 2029, with uplink growing 10x by 2035.

Rather than landing on a single definition, the panel converged on the edge as a continuum. Constantine described it as “a flexible and programmable execution environment spanning from central cloud, regional and local edges, and many times also enterprise edges,” with workloads placed wherever latency, privacy, and cost dictate. Poole’s framing was the one that stuck, though: edge is “where physics collides with economics.” Data gravity decides where compute has to live, and the math is increasingly pushing it outward.

Farney was blunter: “Edge AI inferencing is bringing sexy back to the infrastructure world.” After 20 years of chasing it, he said, the industry finally has the killer use case — AI inference, dense enough and latency-sensitive enough to force compute outward.

Overhauling physical infrastructures and power

So what does an edge site actually look like? For the last 25 years, Poole said, roughly 95% of data centers were built around 5 to 10 kW per rack. AI systems coming out now are running 150 to 200 kW per rack, and Google has shown a 1 MW configuration. That’s not a delta you solve with better airflow.

Two consequences fall out of that. First, self-hosting is largely off the table. As Poole put it: “If you were in the data center industry for the last 20 years, your biggest competition was in-house. You are no longer in a situation where in-house building of data centers is even an option. You can’t do that anymore.” Second, liquid cooling is becoming table stakes, and on-site power generation is moving from a nice-to-have to a regulatory expectation in some states.

The geographic distribution is just as dramatic. Today, most North American compute is concentrated in roughly 15 metros. Poole’s prediction is that the industry will expand into 30 to 50 Tier-2 and Tier-3 markets in a fraction of the 25 years it took to build out the original 15. The open question is the trade-off between aggregation and dissipation: “Do we end up with 300 10-megawatt facilities scattered around the country, or do we end up with 2,000 60-kilowatt cabinets deployed at every cell tower? We’d love that, by the way.”

Qualcomm’s pitch, predictably, was that not all of this needs to run on GPUs. Koymen argued that GPUs are great for training but expensive and power-hungry for inference, and that inference-optimized NPUs at the device and far-edge level are a better fit for the lighter end of the continuum.

Industry predictions for 2028 and 2029

Koymen tied his prediction to Qualcomm’s 6G roadmap — including pre-commercial devices in 2028, commercial launches in 2029 lining up with global operator deployments, and edge infrastructure underpinning use cases like AI recall, “see what I see” AR glasses, and distributed compute for robotics. Constantine predicted that 75% of global data traffic will run on 5G by 2029, that the industry will stop arguing about what the edge is and start arguing about SLAs and TM Forum Level 4/5 automation, and that sustainability pressure on data centers will become a first-order design constraint.

Agarwal’s bet was that by 2028 the panel itself will look different — populated by retailers, miners, and port operators talking about deployed ROI rather than vendors talking about architecture. His warning was that the industry has to avoid repeating the private wireless pattern, where networks got deployed but the success stories never really surfaced. Farney went the furthest out, predicting humanoid robots will start showing up in data center operations to help close the labor gap, alongside a much denser distributed network of edge sites.

Power, labor, and data challenges

The closing round was about “what keeps you up at night.”

Farney flagged workforce. JLL alone has roughly 1,000 open data center roles right now, and the digital infrastructure talent pipeline isn’t keeping up with what’s being built. Poole’s concern was power — the US utility industry has grown at roughly 6% a year for 150 years and, in his words, “cannot get out of its own way.” The grid wasn’t designed to absorb the localized load profiles that edge AI is about to demand.

Constantine’s answer was data. The companies that win, he argued, won’t be the ones with the best AI engineers or models — they’ll be the ones with the best data fidelity and structure. That’s a useful corrective to the model-centric framing that dominates most AI coverage.

Taken together, the panel’s through-line was that the edge has stopped being a marketing problem and started being an execution one. The use case is real, the demand is on the way, and the money is there. What’s left is the unglamorous work of actually building it all out.