NVIDIA and Dassault Systèmes unveil industrial AI platform with physics-grounded world models
A long-term partnership between NVIDIA and Dassault Systèmes aims to make physics-grounded “world models” and virtual twins a mission-critical system of record for engineering, manufacturing, and the sciences.
From digital twins to physics-grounded world models
This collaboration moves beyond today’s project-level twin pilots toward industry-scale models that capture both geometry and behavior, validated against real physics and trusted industrial knowledge. Dassault Systèmes brings its 3DEXPERIENCE platform, Virtual Twin portfolio, and model-based systems engineering expertise; NVIDIA contributes accelerated computing, Omniverse simulation technologies, and open AI model stacks. The goal: use virtual environments not just to visualize, but to design, verify, and operate products and factories before steel is cut or code is deployed.
Platform architecture: 3DEXPERIENCE, Omniverse, and accelerated AI stacks
The companies outlined a shared architecture spanning design, simulation, and operations. On the science side, BIOVIA world models pair with the NVIDIA BioNeMo stack to speed molecule and materials discovery. For engineering, SIMULIA leverages CUDA-X and physics-informed AI libraries to predict behavior in near real time. On the shop floor, DELMIA integrates with Omniverse physical AI libraries to enable autonomous, software-defined production systems. An agentic layer on 3DEXPERIENCE combines NVIDIA AI technologies and Nemotron models with Dassault’s industry world models to create “virtual companions” that augment engineers throughout the lifecycle. NVIDIA will also apply Dassault’s model-based systems engineering to its own AI factory designs, starting with the Rubin platform and the Omniverse DSX Blueprint for large-scale deployments.
Enterprise scale and sovereign cloud for regulated industries
To run these workloads at enterprise scale, Dassault Systèmes’ OUTSCALE brand is deploying NVIDIA-powered AI factories across three continents. That sovereign cloud design targets data residency, IP protection, and regulated-industry needs—key for aerospace, automotive, life sciences, and public sector customers. Early customer voices from Bel Group, OMRON, Lucid, and NIAR signal demand for validated virtual twins to accelerate product cycles, support compliance, and optimize manufacturing.
Why it matters: engineering, manufacturing, and telecom convergence
As factories and supply chains turn into software-defined systems, the stack that connects design, simulation, and operations increasingly depends on low-latency compute and reliable networks from edge to cloud.
Real-time loops: faster iteration and audit-ready certification
Physics-informed models shrink iteration time by orders of magnitude, turning days of simulation into interactive workflows. That changes program economics: more options explored earlier, fewer bad decisions downstream, and faster handoffs from design to manufacturing. In safety-critical sectors, tying world models to model-based requirements and compliance artifacts can cut certification effort by aligning simulation outputs with audit-ready evidence. The near-term payoff is cycle-time compression; the longer-term impact is the ability to create products that were previously too complex or costly to validate.
Software-defined factories require deterministic, converged networks
Virtual twins only deliver operational value if the physical systems they control are connected and predictable. As production becomes autonomous and reconfigurable, enterprises will need deterministic networking, edge inference close to machines, and consistent security policy across plants and regions. This favors converged architectures that blend private 5G, industrial Ethernet with time-sensitive networking, and Wi‑Fi 7 where appropriate, stitched to GPU-rich edge nodes and regional clouds. For telecom operators and integrators, the opportunity is to host simulation, inference, and orchestration on multi-access edge sites; expose network quality APIs to the twin; and co-design network digital twins alongside factory twins to test changes safely before rollout.
Biology and materials R&D becomes a first-class enterprise workload
Combining BioNeMo with BIOVIA world models formalizes life sciences and materials R&D as a first-class enterprise workload. That shifts infrastructure requirements toward mixed precision HPC, curated scientific datasets, and rigorous governance of model lineage and real-world validation. Expect demand for specialized accelerators at research campuses, sovereign cloud zones for sensitive data, and high-throughput links between lab equipment, data lakes, and simulation pipelines.
Execution challenges: V&V, governance, and cost control
The upside is significant, but realizing it requires disciplined systems engineering, governance, and cost control.
Verification, validation, and audit-ready traceability
“Trustworthy by design” means maintaining traceability from requirements to models to test results, and continuously checking simulated behavior against physical measurements. Organizations will need policies for physics-informed model training, versioning and rollback, synthetic data quality, and acceptance criteria by domain. For regulated industries, integrate these controls with established certification processes in aerospace, automotive, medical devices, and process industries.
Data sovereignty and IP protection at scale
Virtual companions and world models will tap CAD, BOM, MES, sensor streams, and supplier data—often proprietary and export-controlled. Data residency, contractual controls, and encryption across edge, sovereign cloud, and public cloud must be non-negotiable. Establish role-based access, purge policies, and red-teaming for model leakage risks. Ensure that partners consuming or enriching models via APIs adhere to the same governance to prevent downstream exposure.
Cost, energy, and utilization optimization
AI factories can be capex- and energy-intensive. Build a robust TCO model that spans GPUs, interconnects, facilities, software licensing, and engineering time, and track utilization and queueing to avoid stranded capacity. Tie energy use to sustainability targets; exploit workload-aware scheduling across on-prem, telco edge, and sovereign cloud to balance latency, cost, and carbon.
Next steps for leaders: roadmap, pilots, and architecture
Use this announcement as a trigger to align product, operations, IT/OT, and network teams around an industrial AI roadmap.
Launch lighthouse programs with measurable KPIs
Pick two to three use cases where physics-informed models and virtual companions can move the needle—e.g., shortening a certification loop, increasing first-pass yield, or reducing ramp-to-rate time. Define baselines and target KPIs, and instrument processes so improvements are auditable. Engage domain experts early to encode validated rules into the models.
Design the edge-to-cloud fabric for AI and simulation
Map where compute must live (machine cell, plant edge, regional edge, sovereign cloud) and what network guarantees each step requires. Evaluate NVIDIA-accelerated nodes for simulation and inference, leverage Omniverse and the DSX Blueprint where digital-physical synchronization is critical, and ensure 3DEXPERIENCE integration with PLM, MES, ERP, and data platforms via APIs. For multi-site operations, plan for private 5G and TSN interoperability and consider telco MEC for latency-sensitive workloads.
Engage the ecosystem and avoid vendor lock-in
Work with Dassault Systèmes and NVIDIA solution partners to accelerate onboarding, but insist on open interfaces and exportable artifacts. Align with relevant standards in networking, data exchange, and model management, and document exit strategies for key components. Where appropriate, pilot on OUTSCALE sovereign zones to align with data residency requirements.
The bottom line: blueprint for system-of-record industrial AI
This partnership is a credible blueprint for making industrial AI a system of record—grounded in physics, scaled by accelerated computing, and operationalized through sovereign cloud and edge.
Industrial AI moves from promise to platform at scale
Enterprises that pair validated world models with software-defined factories and robust network fabrics will compress cycles and de-risk decisions faster than peers. Start small, validate hard, scale what works, and connect the virtual and physical worlds with the right compute and connectivity in the loop.
