AI-RAN pilots begin in shipyards and petrochemical plants
Samsung Electronics has signed contracts with both KT and SK Telecom to participate in South Korea’s government-backed Hyper AI Network program, a two-year initiative designed to test whether private 5G networks and edge AI can support robots doing real industrial work. Samsung says it will serve as the sole global vendor for KT’s project and a main vendor for SKT’s, with field pilots slated to begin in October 2026.
To be clear, this is a supplier-selection and deployment-contract step, not the launch of a commercial-scale AI-RAN service. South Korea’s Ministry of Science and ICT (MSIT) and the National Information Society Agency (NIA) selected the KT- and SKT-led consortia back in July, establishing budgets, industrial use cases and SKT’s multi-vendor approach. What the September announcements add is Samsung’s confirmed roles, the NIS and CognitiV NOS stack it plans to deploy for KT, and the October start date.
The more interesting question underneath all of this is whether these trials can move AI-RAN past optimization rhetoric and toward something industrial operators actually want to pay for.
Program scope and goals
The Hyper AI Network program carries a total budget of roughly KRW 17.2 billion (about $12.7 million) split between KRW 9.2 billion for KT’s consortium and KRW 8 billion for SKT’s. That’s a modest sum, and it signals demonstration and ecosystem-building rather than nationwide RAN replacement. The stated aim is to combine 5G Standalone (SA) private networks, AI-RAN and distributed edge compute to validate “physical AI” in industrial environments. Physical AI, as the program defines it, means AI systems embedded in machines like robots and autonomous vehicles that sense, decide and act in the real world.
Rather than peak downlink speeds, the targets are predictable uplink performance for video and sensor feeds, low and stable latency for control loops, high reliability, local processing and automated network operations. These are the demands of a shipyard or petrochemical plant, not a consumer mobile network. Authorities plan to expand the effort toward humanoid-robot demonstrations after 2027, and the program sits within a broader national push toward what the government calls “AI-native 6G.”
The actual pilots run on 5G SA — today’s technology, deployed in private-network configurations. “AI-native 6G” remains a forward-looking objective, and conflating it with what’s being tested on the ground would overstate where things stand.
AI-RAN itself isn’t a single standardized product. In this context, it refers broadly to two things. The first is applying AI to radio-access-network operations — automated optimization, orchestration, anomaly detection. The second is using RAN and edge-compute resources to serve AI workloads close to industrial devices. The program’s central question is whether this kind of integration delivers measurable safety, productivity and operating-cost improvements, or whether the added compute, power and lifecycle complexity simply isn’t worth it relative to conventional setups.
Inside KT’s trials
KT’s consortium includes Samsung and HD Hyundai Samho, and the work is centered on HD Hyundai Samho’s Yeongam shipyard. The planned use cases reflect the difficult, high-risk work that shipbuilding involves. They include a quadruped AI welding robot, AI-assisted painting robots, and autonomous robots used for telecom-facility checks. KT is also developing what it calls an “AI Core Orchestrator,” intended to link the 5G core network’s data analytics function (NWDAF, a 3GPP standard for analyzing network performance data and exposing insights to other 5G functions) with AI to detect problems and trigger automated responses in real time.
Samsung says it will supply two main pieces of technology to KT. The first is its Network in a Server (NIS), an on-premises edge platform that consolidates virtualized RAN (vRAN), an AI core and AI applications into a compact server. Samsung says NIS runs on a Supermicro edge server with AMD EPYC 8004 CPUs and Wind River Cloud Platform. The second is CognitiV Network Operations Suite (NOS), Samsung’s network-automation platform for managing and orchestrating the private 5G deployment. Together, they’re designed to handle the local processing and orchestration that control-loop workloads demand.
KT is setting up a separate multi-vendor testbed at its Umyeon-dong R&D center, where it plans to validate Samsung’s equipment alongside domestic suppliers. That testbed will also cover base-station energy-saving technologies and RedCap, a reduced-capability 5G device category aimed at lower-power devices. While Samsung is KT’s sole global vendor for the field trial, KT clearly isn’t limiting itself to a single supplier across the board.
SKT’s approach is deliberately different. Its initial industrial site is SK Incheon Petrochem, an energy and chemical complex serving the Seoul metropolitan area. The primary use case is an autonomous patrol robot navigating hazardous areas and streaming high-definition video in real time, with AI analyzing the feeds to identify possible hazards. SKT’s plans extend further — unmanned autonomous transport at a Pangyo physical-AI living lab, extended in the second year to a KG Mobility plant in Pyeongtaek, a digital-twin system that ingests LiDAR data from site infrastructure to enable remote driving commands, and a low-power mode for humanoid robots that offloads some AI compute to the network to reduce on-device power and battery demands.
Samsung provides NIS for SKT’s project as well, but the competitive dynamic is quite different from KT’s arrangement. SKT has explicitly said it will build and compare AI-RAN equipment from Samsung, HFR, Ericsson and Nokia, testing different CPU and GPU configurations and varying where AI servers and user-plane functions sit in the architecture. SKT’s consortium also includes HFR and Ericsson Korea as equipment partners, alongside KG Mobility, Seoul Robotics, Intellivix and Clevi. This is a genuine vendor and architecture bake-off, and Samsung is one participant among several.
Why this matters for industry and vendors
Field trials that put AI-RAN on an actual factory floor with actual robots represent a different kind of test than what’s come before. Samsung and KT ran an AI-RAN optimization trial in December 2025, involving roughly 18,000 users on KT’s commercial network. That work focused on conventional mobile traffic. The Hyper AI Network projects shift the focus to private industrial networks, where patrol robots, welding machines, cameras and LiDAR systems generate uplink-heavy, latency-sensitive data streams that stress network reliability in ways standard mobile use cases don’t.
For Samsung, the program doubles as a proving ground for its integrated stack — vRAN, private 5G core, edge AI and network automation, tested together in environments that will expose weaknesses pretty quickly. Winning roles in both operator-led projects gives Samsung a prominent domestic reference account while RAN vendors globally are competing to define what AI-RAN architectures actually look like ahead of 6G. Samsung and SKT had signed an AI-RAN-focused 6G memorandum of understanding in November 2025, so the September contracts represent something more concrete and government-backed.
The technical architecture across both trials favors local edge compute over centralized cloud. NIS consolidates vRAN, AI core and applications on premises, while automation layers — including SMO-like orchestration (service management and orchestration, a framework tied to Open RAN network management) and NWDAF analytics — handle network-level intelligence. That design reflects the latency and reliability requirements of closed-loop industrial control, where sending data to a distant cloud and waiting for a response doesn’t work.
That said, the contracts were announced September 23, with pilots due to begin in October over a two-year horizon. The KRW 17.2 billion budget makes the scope clear. This is validation, not commercialization. Whether the program leads anywhere larger depends entirely on whether the trials produce measurable industrial results — safety improvements, productivity gains, operating-cost reductions — that justify the added integration complexity.
Risks and caveats
The performance claims circulating around these projects come almost entirely from Samsung and program participants, and no independent results from the trials had been published as of late September. Terms like “ultra-low latency,” “high reliability” and autonomous robotic control describe objectives, not independently verified outcomes. The gap between a vendor announcement and a field result is often substantial.
Industrial robot safety isn’t a pure connectivity problem, either. Network quality is one component, but robot design, sensing accuracy, workflow integration, safety certification and human oversight all play critical roles. A perfectly reliable 5G link doesn’t automatically make a welding robot safe, and reducing the safety question to a network-performance story would be misleading.
SKT’s multi-vendor approach is arguably the most valuable aspect of the program from a broader industry perspective. Comparing Samsung, Ericsson, Nokia and HFR equipment with different hardware configurations and network-function placements should surface real architectural trade-offs — the kind of head-to-head data that’s rarely available in telecom. If SKT shares those findings, the results could shape how operators elsewhere evaluate AI-RAN deployments.
The bring-up now under way at the shipyard and petrochemical sites should produce early KPIs on latency variation, uplink throughput and robot task-completion rates. SKT’s comparative findings across its four vendors will be especially informative. Progress on KT’s AI Core Orchestrator and NWDAF training should reveal whether the network-intelligence layer adds real operational value. Beyond 2027, any movement toward humanoid-robot demonstrations and inputs to AI-native 6G requirements will signal whether the program has legs — or whether it remains a government-funded proof of concept.
Table of Contents
