Edge AI depends on the network, not just the compute

by Christian de Looper
Edge AI infra

Uplink, device visibility and workforce needs shape industrial edge AI

In sum – what to know:

  • Spread, not moved – Ericsson’s Julien Forgeat said AI is spreading to the edge rather than leaving the cloud, so each workload needs pragmatic placement.
  • Uplink as foundation – Industrial AI sends video and sensor data upstream, making uplink capacity and reliability the key network requirement.
  • Hidden devices – A SIM identifies a subscription, not the machine behind the router, so OneLayer’s Avishag Daniely urges a full device inventory.

There’s a lot of talk of AI moving from the cloud to the edge — but the truth is a little more complex. In reality, AI is spreading across both the edge and the cloud, and the harder question is where each workload should run. That was the practical thread running through “Edge AI in action,” a panel at the Industrial Wireless Forum, hosted by ABI Research research director Leo Gergs with OneLayer’s Avishag Daniely, Cargill’s Robert Greiner, Ericsson’s Julien Forgeat, and China Telecom’s Dr. Yu Zeng, who is also vice chair of ETSI TCNet.

Forgeat corrected the premise of the opening question. “AI is not really moving away from the cloud; it’s spreading to the edge,” he said. He framed the decision around four factors. Latency matters when a round trip to a centralized cloud is too slow. Data gravity, the pull created by large volumes of data being generated in one place, can make it cheaper to process information near its source rather than pay to move it. Privacy and sovereignty can require tighter control, while the full cost calculation has to include data transfer, network peering, and energy. Peering is the exchange of traffic between networks, and those charges can change the economics of sending everything upstream. Centralized cloud still offers scale that a plant server or phone can’t match. For industrial video and large device fleets, though, those other constraints can favor processing regionally, locally, or on the device itself.

That means there isn’t one edge, and “edge-first” can be as unhelpful as “cloud-first.” Forgeat argued for placing each workload pragmatically across devices, local compute, carrier infrastructure, and cloud rather than automatically assigning inference to one and training to the other. “You have to be very pragmatic,” he said. “Don’t bring any ideology.” Private cellular 5G, a dedicated mobile network operated for a particular site or organization, is becoming an important connective layer for autonomous mobile robots, or AMRs, automated guided vehicles, or AGVs, remote operations, and sensor data. At Cargill, Greiner described a “build it and they will come” approach borrowed from Iowa’s Field of Dreams. The company began with tablets and relatively small plant-connectivity projects. Now the discussion has moved to robots and mobile control, and Greiner said it has “finally taken off.”

Uplink and reliability become the new network requirement

The network requirement that stood out most was uplink, the traffic traveling from devices back into the network. Consumer cellular design has generally prioritized downlink, because phones and other devices have mostly downloaded content. An industrial AI deployment can reverse that emphasis. Cameras, lidar sensors, and other equipment may continuously send video, three-dimensional spatial data, and control information to nearby compute for processing. “The biggest change that we have to implement is to increase uplink capacity,” Forgeat said.

Capacity alone won’t settle it. Forgeat said synchronous applications, where connected processes have to respond in step with one another, also need low jitter and dependable service levels. Jitter is the variation in how long packets take to arrive, and an uplink service-level agreement, or SLA, sets expectations for the connection’s performance. A high peak throughput figure doesn’t tell you whether a robot’s control loop will behave consistently throughout a shift.

Asked to choose among latency, reliability, mobility, and uplink performance, Forgeat called his own answer boring and chose uplink again. His reasoning was that additional capacity and reliability can improve the others mechanically. Applications are less likely to stall waiting for bandwidth, latency becomes easier to manage, and moving devices can reconnect to a new cell with useful uplink performance sooner. “If you improve the uplink, you improve everything,” he said. That’s an argument for treating uplink as a foundation, rather than a separate box to check after the compute has been purchased.

Zeng approached the constraint from the operator and standards side. Uploading everything isn’t sustainable when cameras, Internet of Things devices, and video applications all generate traffic continuously. Cloud processing works efficiently when bandwidth and peering are cheap enough, but dense video and constant telemetry put pressure on that assumption. Local processing can reduce what has to travel, provided the local hardware is actually up to the job.

That qualification matters. Zeng described edge hardware maturity as “the most easily overlooked point”, pushing back on the idea that the edge should simply mean cheaper, less capable compute. A useful service still needs sufficient processing resources. Looking ahead, Ericsson’s research focus on 6G points toward giving uplink, mobility, and deterministic performance, meaning consistent and predictable behavior, more attention than downlink-heavy cellular design has historically given them.

Device visibility is the missing security layer for physical AI

Security discussions around physical AI often begin with protecting the model or the edge node running it. Daniely said the larger blind spot is the collection of devices supplying those models with data and acting on their decisions. Without governance over those devices, she said, the model is “just as good as the data feeding it.” A secured model doesn’t help much if the information reaching it comes from equipment the organization can’t properly identify or govern.

Cellular authentication can create a false sense of completeness here. A SIM card identifies a subscription to the network, but it doesn’t necessarily identify the machine behind the connection. Many devices across factories, electrical grids, ports, mines, and airports don’t have built-in SIMs or cellular modems. An AGV bought before a factory adopted private cellular, for example, may connect over Ethernet to a router or gateway mounted on it. The router carries the SIM.

From the cellular network’s perspective, that can mean visibility stops at the adapter. The network sees the SIM and router, while the AGV sending sensor data and receiving instructions remains hidden behind them. The same setup can cover moving assets across a large site, but authenticating the router isn’t the same as knowing what’s connected to it.

Daniely’s recommended starting point was an inventory that maps SIMs, routers, protocols, and the devices behind them before additional policy enforcement or security layers are added. “Know what you have connected,” she said. “You need visibility.” That inventory has operational value beyond security. It can help diagnose connectivity problems, support troubleshooting, and establish a more controlled path for low-latency data transport. The blind spot isn’t specific to any one vendor’s product. OneLayer sells security for private cellular networks, so its emphasis on device governance naturally aligns with its business.

Business cases grow from workforce automation and remote operations

At Cargill, the business case is closely tied to a changing workforce. Greiner said younger employees arriving as older workers retire expect immediate access to information on tablets and phones while walking the plant floor, rather than a job built around watching screens in a static control room. Over time, that could extend to starting and stopping equipment from an app while standing beside the machine. His comparison was the consumer market, where controls that once seemed bound to fixed switches and dedicated rooms have moved into phone apps. He said he tells skeptics, “You just have to look at where the consumer side is.”

The workforce is also shrinking, which is pushing Cargill to consider remote operator centers. Greiner raised the possibility of a center in Amsterdam running five European plants. That could make better use of scarce operators, but Greiner said it “puts a whole lot of strain on your whole ecosystem.” That strain lands on plant-edge connectivity, local control systems, and security. Someone may still need to stand in front of equipment with a mobile device, even when the main operator is hundreds of miles away.

Warehousing and packaging offer another path. Greiner described the shift from skilled forklift operators to fleets of AMRs loading trucks, and the difficult transition between those two operating models without rebuilding the entire process. Suppliers don’t necessarily make that transition obvious. Cargill has worked with MiR, a major AGV supplier, and Greiner said a regional salesperson or third-party integrator may initially default to Wi-Fi or say private cellular isn’t supported. Digging into the documentation can reveal a supported gateway path instead.

His experience with that approach was positive, though it remains one company’s account rather than a controlled comparison. He said people end up “wonderfully surprised” by how it works once installed, since private cellular avoided the recurring Wi-Fi dropouts that could stop equipment during operations. For industrial mobility, predictability can be the difference between a useful deployment and one that keeps interrupting work. It’s also why a generic preference for Wi-Fi, based on what an integrator has always installed, isn’t much of an architecture decision.

Zeng’s closing advice was to check for applicable standards before building new mechanisms, keep raw data where it’s generated, and transmit trends or inference outputs where regulation permits. Rules such as the EU AI Act, alongside country-specific requirements, can turn data sovereignty from a preference into a hard boundary on where information may travel and how it may be used. Cloud infrastructure still makes sense for broader analytics and training. Low-latency controls and inference may belong closer to the operation when sovereignty and transfer costs narrow the options.

The unresolved question is how deliberately organizations will design that placement. Forgeat warned against assuming edge AI is finished once the compute arrives, since it needs “very strong and very well controlled network resources,” while Greiner pointed to a remote Caterpillar loader manual that devoted roughly two pages out of 300 to networking. That imbalance captures the risk pretty well. Buying edge compute is straightforward compared with proving that the network, device inventory, and controls will hold together once a real operation depends on them.