SK Telecom and Nvidia have moved beyond a partnership announcement and into a standing operational structure: a formal joint committee, running on two tracks, to co-develop physical AI technology and bring it to semiconductor fabs, shipbuilding yards, and defense facilities. The committee launched following Nvidia's GTC San Jose 2026 keynote in March, where Jensen Huang first named SK Telecom as a key physical AI partner alongside Boston Dynamics and other global robotics firms. Three months later, at his GTC Taipei keynote on June 1, Huang did it again — the second consecutive major Nvidia event at which SK Telecom appeared as a named strategic partner in manufacturing physical AI, this time with an Omniverse-powered chip fab deployment on the keynote screen.

The committee divides into an executive-level track for strategic direction and a working-level track for hands-on co-development of an Omniverse-based digital-twin platform. Nvidia engineers responsible for Omniverse have already traveled to SKT Tower in Seoul for at least one technical workshop under the committee structure, according to industry sources.

What Agentic Digital Twin Modeling Does: OpenUSD, NVIDIA Agent Toolkit, and Fab-Scale 3D Data

At the center of the technical work is a system SK Telecom built using Nvidia's Agent Toolkit, which the company calls Agentic Digital Twin Modeling. The problem it solves is a genuine engineering bottleneck: a semiconductor fab contains thousands of pieces of equipment with complex spatial relationships, and building a working digital twin of that environment requires ingesting CAD models, sensor feeds, equipment layouts, and environmental data — then converting and optimizing all of it into a format a simulation platform can use. Before SKT developed this system, engineers performed that conversion manually, piece by piece.

Agentic Digital Twin Modeling automates the data-conversion and scene-optimization step using AI agents running on top of Nvidia Omniverse. OpenUSD — Universal Scene Description, a framework originally developed by Pixar Animation Studios — serves as the common scene language: it describes the geometry, materials, spatial relationships, and physical properties of every component in the fab as a single versionable 3D file that multiple teams can work on simultaneously. SKT's platform then integrates Nvidia Omniverse libraries to improve loading speed, GPU utilization, and memory efficiency when handling the enormous OpenUSD scene files that a chipmaking facility generates.

"Semiconductor fabs are among the most challenging manufacturing environments, combining massive amounts of 3D data, complex equipment structures, and the need for high-level optimization," said Mike Geyer, head of industrial digital twins at Nvidia. "SKT has demonstrated a high level of technical capability in applying and validating NVIDIA Omniverse libraries, as well as the NVIDIA Agent Toolkit in real-world industrial settings."

Cho Ik-hwan, head of physical AI at SK Telecom, said the collaboration validated that manufacturing digital twins could evolve beyond visualization into platforms capable of understanding and optimizing large-scale 3D manufacturing data — adding that SKT would expand its role as a physical AI technology partner with Nvidia across industrial sectors including semiconductors.

How SK hynix Uses Simulation Before Deploying Changes on Real Chip Lines

The first full deployment is at SK hynix, which manufactures a significant share of the high-bandwidth memory that powers Nvidia's own AI accelerators and is targeting full autonomous fab operations by 2030. SK Telecom and SK hynix completed a proof-of-concept together in 2025, confirming the digital twin platform could handle the complexity of a working semiconductor fab. Phased commercialization is now underway.

The operational purpose is straightforward: instead of testing a new equipment layout or process change on a live production line — where a mistake could halt wafer output and cost chipmakers millions per hour — engineers run the proposed change inside the digital twin first. The virtual fab replicates equipment placement, process flows, and logistics in a physics-accurate 3D model. If the simulation shows a bottleneck or conflict, engineers fix it before anything is touched on the real floor.

The robot-learning component adds a second function. SK Telecom runs training iterations in the simulation environment to fine-tune a robot foundation model for specific industrial tasks before any robotic system is deployed on the actual fab floor, according to the company. The general-purpose model is adapted into an industry-specific version through this simulation-heavy training loop.

To run these workloads, SK Telecom is building an industrial AI cloud with more than 2,000 Nvidia RTX PRO 6000 Blackwell GPUs. These GPUs are purpose-built for Omniverse workloads — their architecture supports the high-memory-bandwidth demands of large OpenUSD scene files and the parallel compute required for robot simulation at scale.

How Does a Semiconductor Fab Digital Twin Work at Scale?

The short answer: very differently from a static 3D model or a standard simulation. An Omniverse-based fab digital twin is a synchronized virtual replica of a physical facility that ingests live data from the real floor — sensor readings, throughput metrics, equipment states — and keeps the virtual model current. Changes proposed in the virtual model can then be tested against the live-data context before implementation.

At the architecture level, OpenUSD's composition system allows the twin to be assembled from modular assets — a single production cell can be defined once and referenced across multiple factory layouts without rebuilding it from scratch. This means an engineer can rearrange an entire fab section virtually in hours rather than days. Nvidia's Omniverse microservices handle the rendering, physics simulation, and collaborative access layers, while the RTX PRO 6000 Blackwell GPUs supply the compute headroom for the parallel simulation threads that make real-time response possible.

The scale demand at a semiconductor fab is qualitatively different from most digital twin deployments. Where a logistics warehouse digital twin might involve hundreds of OpenUSD objects, a chipmaking facility involves tens of thousands of precisely calibrated tools — each with its own spatial constraints, maintenance schedules, and process dependencies. That data volume is why SKT's automated ingestion layer is a prerequisite rather than a convenience.

SK Telecom Targets Sectors With Highest Production-Stoppage Costs

SK Telecom's stated rationale for sector selection is explicit: it is pursuing industries where process complexity is high and production stoppages are disproportionately expensive. Semiconductor fabs fit that description precisely — a single unplanned downtime event on an advanced-node line can cost chipmakers millions per hour. Shipbuilding and defense have equivalent characteristics: complex assembly sequences, expensive physical assets, and high cost of rework.

The company plans to expand beyond group affiliates once the SK hynix deployment is at commercial scale. The longer-term target is all industrial sites, including logistics and service operations, with SK Telecom positioning itself as an end-to-end physical AI provider — supplying the integration layer above Nvidia's GPU infrastructure and below individual industrial customers.

That positioning is not without precedent elsewhere in the industry. Samsung announced its own Omniverse fab digital twins inside a 50,000-GPU AI Megafactory in October 2025. TSMC disclosed its FabTwin environment — built on Omniverse libraries — for evaluating process-tool layouts at GTC Taipei on the same day as the SKT announcement. Both competitors are building on the same underlying platform; SKT's differentiation rests on its Agentic Digital Twin Modeling layer and its role as a third-party integrator serving a chip manufacturer rather than operating as one.

SK Telecom reorganized its internal structure to support this strategy, building a dedicated 60-person physical AI organization under its AI CIC in September 2025. The unit was expanded from the company's former metaverse team and is led by Cho Ik-hwan. Its mandate covers digital twin development, robot learning, and manufacturing-oriented AI data platforms.

The company's ambitions arrive against the backdrop of a significant institutional setback: in August 2025, South Korea's Personal Information Protection Commission fined SK Telecom ₩134.8 billion — approximately $97 million — following a data breach in April 2025 that compromised subscriber identity records for as many as 27 million users. Investigators found systemic failures in access control and encryption of authentication keys. SK Telecom filed a lawsuit in January 2026 challenging the penalty, pledging it had already spent approximately ₩1.2 trillion on user compensation and data protection reforms. The case is ongoing in Seoul's administrative courts. The episode underscores the governance demands of handling sensitive industrial data at the scale SK Telecom now targets.

Frequently Asked Questions

What is physical AI in manufacturing, and how does it differ from standard industrial automation?

Physical AI refers to AI systems that interact with and operate in the physical world rather than generating text or images — they perceive environments, make decisions, and direct physical systems such as robots or production equipment. In manufacturing, physical AI goes beyond traditional automation by using simulation-trained models that can adapt to new tasks without reprogramming every motion from scratch; a robot trained in a digital twin can be fine-tuned for a new industrial setting much faster than a conventionally programmed system.

What is SK hynix's Autonomous Fab 2030 goal?

SK hynix is working toward semiconductor fabrication plants that can monitor, adjust, and optimize their own processes with minimal human intervention by 2030. The Autonomous Fab 2030 program uses digital twins to simulate and verify process changes before deploying them on real production lines, and it relies on AI agents built on SK Telecom's A.X foundation model and Nvidia's NIM microservices to handle routine operational decisions automatically.

How does Nvidia Omniverse power digital twins for semiconductor manufacturing?

Nvidia Omniverse is a collection of libraries and microservices built on OpenUSD, a scene-description framework originally developed by Pixar that stores geometry, materials, and physical properties of every component in a unified, versionable 3D file. For semiconductor fabs, the platform handles the massive data volumes and parallel simulation threads needed to replicate tens of thousands of precisely calibrated tools in a physics-accurate virtual environment; the RTX PRO 6000 Blackwell GPUs provide the memory bandwidth and parallel compute headroom those workloads require.

What makes SK Telecom's role in this partnership unusual for a telecom company?

SK Telecom's entry into industrial AI integration stems from its transformation of a former metaverse team into a 60-person physical AI unit and its position as an Nvidia cloud partner building the GPU infrastructure SK hynix needs. Telecom operators historically have managed network infrastructure, not factory simulation software — SKT is betting that its data-handling scale and its existing relationship with SK Group affiliates gives it a structural advantage as the integration layer between Nvidia's compute platform and industrial customers.