With North America's largest industrial automation trade show opening in Chicago in three days, NVIDIA and South Korea's Doosan Group arrive with the most vertically integrated physical AI partnership either company has announced: a collaboration that now spans four Doosan subsidiaries and touches every layer of the emerging AI factory stack — from the copper laminate inside NVIDIA's AI accelerator boards to the gas turbines and small modular reactors that could power the data centers running them.

The expanded partnership, announced June 7 during Jensen Huang's Seoul visit and formalized in a Doosan press release June 8, pulls together Doosan Robotics, Doosan Bobcat, Doosan Enerbility, and Doosan Corporation Electro-Materials BG into a single strategic alignment with NVIDIA's full physical AI and AI factory platform. The scope sets it apart from NVIDIA's other industrial partnerships: where most deals attach one company to one part of the AI stack, Doosan is threading into all of them simultaneously.

How Industrial AI Is Trained Before It Touches Metal

The central engineering challenge in physical AI — and the one that has blocked industrial robotics from reaching the 99-plus percent reliability manufacturers require — is called the sim-to-real gap: the disconnect between how an AI policy performs in simulation and how it performs when confronted with the unpredictable friction, material variation, and contact dynamics of a real factory floor.

Doosan Robotics is integrating NVIDIA's entire pipeline for closing that gap into its new Agentic Robot OS platform. The architecture works in layers. Isaac Sim provides the high-fidelity physics and photorealistic rendering environment where robot policies are first trained; Isaac Lab is the GPU-parallel training framework that runs thousands of simultaneous simulated robot instances to generate policy data at a speed and scale impossible in the physical world. On top of both runs Cosmos, NVIDIA's open world foundation model, which generates synthetic video and sensor data of physical environments — teaching robots what the world looks like and how it behaves before they ever leave the lab. The Newton physics engine provides the contact dynamics and material simulation accuracy that makes synthetic training data physically believable. When a trained policy is finally ready for hardware, it runs on Jetson Thor, the on-device compute module that serves as the robot's inference brain.

Doosan and NVIDIA are targeting two initial industrial applications — depalletizing and sanding — along with new robot form factors including dual-arm and humanoid platforms. By 2028, Doosan Robotics has set a target for commercial deployment of industrial humanoid robots built on this pipeline.

The significance of that target is technical, not merely commercial. If Doosan's Agentic Robot OS successfully demonstrates sim-to-real transfer at production reliability thresholds across multiple task classes, it will establish a repeatable deployment architecture — Cosmos-trained, Isaac-validated, Jetson-deployed — that every other industrial conglomerate competing in the physical AI space will need to match or license. Competitors including ABB Robotics are building on the same NVIDIA stack, which makes Doosan's rate of industrial application — not the tools themselves — the competitive variable.

Doosan Bobcat Takes Physical AI Off the Factory Floor

The collaboration extends well beyond traditional manufacturing. Doosan Bobcat — the company's construction, landscaping, agriculture, and material-handling equipment line — is integrating NVIDIA physical AI to develop specialized world models for outdoor and job-site applications. Where factory robots operate in controlled environments, Bobcat equipment encounters unstructured terrain, variable weather, and irregular objects: a harder sim-to-real problem.

The companies say they aim to help establish an industry-standard ecosystem for compact autonomous equipment. The choice of language — "industry-standard ecosystem" — suggests an intent to produce open frameworks that other compact machinery manufacturers could adopt, similar to what NVIDIA has done with its open-source Isaac Lab framework for general robotics.

What Powers an AI Factory: Why Doosan Enerbility's Reactors Matter

The power bottleneck is now the primary constraint on AI factory buildout. AI training clusters running NVIDIA Blackwell systems require hundreds of megawatts of continuously available electricity — a load profile that solar and wind generation cannot reliably provide. That structural mismatch is why Doosan Enerbility's generation portfolio is strategically relevant to NVIDIA's DSX AI factory platform.

Doosan Enerbility is exploring how its large-scale power infrastructure — gas turbines, steam turbines, and small modular reactors, along with Doosan Fuel Cell's hydrogen systems — could serve as the energy backbone for NVIDIA DSX deployments. Small modular reactors, defined as nuclear fission reactors under 300 megawatts that are factory-built and transported to installation sites, are particularly well-matched to AI data center requirements: they provide always-on, carbon-free baseload power without grid dependency, and their modular design allows operators to start with a single unit and scale incrementally as compute demand grows.

NVIDIA's own DSX documentation states that its MaxLPS technology maximizes tokens per megawatt by combining 45-degree Celsius liquid cooling with in-rack power optimization, enabling operators to run up to 40 percent more GPUs at their most energy-efficient operating point. Pairing that efficiency layer with Doosan Enerbility SMR generation would give AI factory operators both sides of the power equation: more compute per watt on the infrastructure side, and carbon-free always-on power on the supply side.

Future collaboration could include AI factory power supply design tailored to NVIDIA DSX configurations, optimization of Doosan Enerbility's generation equipment using AI tools, and formal evaluation of SMR as a dedicated AI factory power source.

The Layer No One Talks About: PCB Substrate and AI Hardware

Perhaps the least visible but most concrete dimension of the partnership is Doosan Corporation Electro-Materials BG's role as a copper clad laminate supplier. CCL — a copper foil bonded to an insulating substrate — is the foundational material for printed circuit boards used in NVIDIA's AI accelerators, networking gear, and server motherboards. In high-performance AI systems, CCL must minimize signal loss and resist deformation at the elevated temperatures produced during sustained inference and training workloads. Standard-grade CCL used in consumer electronics cannot meet these requirements.

Doosan Electro-Materials is positioning its CCL production to support NVIDIA's MGX modular server architecture, which helps system manufacturers build rack-scale AI factory infrastructure. Analyst coverage from Korea Investment and Securities has noted that Taiwan's Elite Material, the world's largest CCL producer, has faced supply difficulties with NVIDIA, creating an opening for Doosan's quality-competitive product.

Doosan's commitment to the AI hardware supply chain extends beyond the partnership announcement. In April 2026, the company announced a 180 billion won (approximately $130 million) investment to build a new CCL production facility in Thailand's Araya Industrial Estate, targeting mass production in the second half of 2028. The Thai plant will be Doosan Electro-Materials' third overseas production base, following existing facilities in China and Vietnam, and will specialize exclusively in high-performance CCL for AI infrastructure and networking equipment.

What Has Changed Since October

The NVIDIA-Doosan relationship is not new. A narrower partnership was first announced in October 2025, covering physical AI applications in construction and power equipment. The June 2026 expansion pulls in Doosan Robotics and Doosan Electro-Materials as full participants, extends the technology stack to include Cosmos, Newton, and Jetson Thor, and adds the CCL materials supply chain dimension. The result is a partnership that now reaches from AI accelerator substrate to robot operating system to AI factory power generation — all within one Korean industrial conglomerate.

The expansion coincided with NVIDIA CEO Jensen Huang's five-day Seoul visit in early June, during which he met with chairs of every major South Korean technology conglomerate. Huang threw out the ceremonial first pitch at a Doosan Bears home game at Jamsil Baseball Stadium on June 7, with Doosan Group Chairman Park Jeong-won serving as the ceremonial batter. Park wore jersey number 96, marking Doosan's 1896 founding year; Huang wore number 93 for NVIDIA's 1993 founding.

The market reaction was concrete. Doosan Robotics shares climbed approximately 4 percent on the news; NVIDIA gained approximately 2 percent amid a broader wave of South Korean AI deals that also included agreements with SK Group, LG Group, Hyundai Motor Group, and NAVER.

What the Automate 2026 Stage Means for the Partnership

Automate 2026 — North America's largest robotics and automation trade show, running June 22–25 at McCormick Place in Chicago — is the venue where these announcements will face their first large-scale industry audience. The show floor will include competing approaches from ABB Robotics, FANUC, KUKA, and Universal Robots, all of which are also building on NVIDIA's physical AI stack. Automate will be the first opportunity to compare Doosan's Agentic Robot OS implementation against those of peers integrating the same NVIDIA tools.

For engineers and automation executives attending, the practical question is not which company is adopting NVIDIA's platform — nearly all of them are — but which company has advanced furthest from integration to validated industrial application. Doosan arrives with the claim of a full-stack integration spanning perception, reasoning, simulation, and inference, backed by a 2028 commercial humanoid deployment target.

Frequently Asked Questions

What is the sim-to-real gap, and why does it matter for industrial robotics?

The sim-to-real gap is the engineering problem at the center of physical AI: AI policies trained in simulation frequently fail when deployed on real hardware because simulated physics — friction, contact forces, material deformation, sensor noise — only approximates the real world. The discrepancy means robots that perform well in virtual training can fail on actual factory tasks. NVIDIA's physical AI stack, which Doosan is integrating, addresses this through domain randomization (training across thousands of varied simulated conditions), the Newton physics engine (high-accuracy contact dynamics), and Cosmos world foundation models (generating physically accurate synthetic training data). Whether this pipeline narrows the gap enough for the 99-plus percent reliability manufacturing requires is the central open question in industrial physical AI as of mid-2026.

What is NVIDIA's AI factory platform, and what does power have to do with it?

NVIDIA's DSX AI factory platform is a reference architecture for large-scale AI data centers that coordinates compute, networking, storage, and power infrastructure as a unified system. At full scale, an AI factory running NVIDIA Blackwell systems requires hundreds of megawatts of continuously available electricity. Grid power, solar, and wind cannot reliably provide that load profile, which is why NVIDIA is working with energy companies including Doosan Enerbility to evaluate small modular reactors as dedicated AI factory power sources. SMRs — compact, factory-built nuclear reactors under 300 megawatts — provide always-on, carbon-free baseload power that does not fluctuate with weather or time of day, making them the most architecturally compatible power source for always-on AI training workloads.

Why does NVIDIA need copper clad laminate from Doosan, and what is it?

Copper clad laminate is the foundational material for printed circuit boards: an insulating substrate bonded with copper foil on both sides. AI accelerators require higher-grade CCL than conventional electronics because they process enormous amounts of data at ultra-high speeds, generating significant heat and demanding signal integrity at bandwidth densities that standard CCL cannot maintain. Doosan Electro-Materials produces high-performance CCL aligned to NVIDIA's MGX modular server architecture. Analyst coverage has noted that the world's largest CCL supplier, Taiwan's Elite Material, has encountered supply difficulties with NVIDIA — creating a meaningful opportunity for Doosan's product.

What does the 2028 Doosan humanoid robot target actually require to succeed?

Commercializing industrial humanoid robots by 2028 requires that Doosan's Agentic Robot OS — built on NVIDIA's Isaac Sim, Isaac Lab, Cosmos, Newton, and Jetson Thor — demonstrate reliable sim-to-real transfer for dexterous manipulation and locomotion tasks in real manufacturing environments. Industry experts note that demo-quality performance at 70 percent reliability is not sufficient; production deployment requires 99 percent or better. The 2028 target is achievable if the sim-to-real pipeline produces policies that generalize across the unstructured variation of real factory conditions — which remains the central engineering challenge the entire physical AI industry is working to solve.