AI-Native Digital Twins: NVIDIA + Dassault Systèmes Are Turning Simulation Into an AI Engineering System
by 
Digital twins are no longer just virtual replicas of factories, machines and products. The next generation is becoming an AI-native engineering environment where physics, simulation, real-world data and autonomous agents work together. NVIDIA and Dassault Systèmes are building toward that future, and the implications for engineering teams could be far bigger than faster simulation.
For decades, engineers have used simulation to answer a deceptively simple question: what happens if we change something in the physical world? The problem is that answering that question at high fidelity can consume significant computing time and engineering effort. A design may need to be simulated, analyzed, modified and simulated again before a team can confidently move forward. That creates a bottleneck at the exact point where modern engineering needs more experimentation, not less.
The NVIDIA and Dassault Systèmes partnership announced in February 2026 points toward a different model. The companies are combining Dassault Systèmes’ Virtual Twin technologies with NVIDIA’s AI infrastructure, accelerated software libraries and computing platform to develop science-validated industrial world models across engineering, manufacturing, materials science and other complex domains.
But the more important development came in July. NVIDIA added PhysicsNeMo and CUDA-X libraries to its Agent Toolkit as agent-ready tools and skills, specifically positioning them for autonomous AI engineering workflows. Just days earlier, Dassault Systèmes announced that its Virtual Companions, including LEO for complex engineering and simulation, were available on its AI-native agentic platform.
That changes the story.
This is no longer simply about using GPUs to accelerate a digital twin. It is about creating an AI digital twin that can become part of an engineering agent’s reasoning and execution loop.
Digital Twin Technology Has Already Become an Industrial Priority
The business case for digital twin technology is no longer theoretical. McKinsey reports that 70% of C-suite technology executives at large enterprises are exploring or investing in digital twins, while its research indicates that companies in advanced industries are increasingly deploying twins at meaningful levels of complexity. McKinsey also estimates that digital twins can reduce the time required to deploy new AI-driven capabilities by up to 60%, while relevant implementations can reduce capital and operating expenditure by up to 15% and improve commercial efficiency by around 10%.
Those numbers explain why digital twins have moved beyond visualization projects. A digital twin can become a decision environment where companies test production layouts, optimize equipment, predict asset behavior and evaluate design alternatives before making expensive physical changes.
The limitation is scale. The more complex the physical system becomes, the more simulation scenarios engineers want to explore. That means computational cost becomes part of the engineering problem. This is where an AI digital twin becomes interesting.
The Real Problem NVIDIA Is Trying to Solve Is Time
CUDA-X is not itself a digital twin platform. It is a collection of GPU-accelerated libraries within NVIDIA’s CUDA ecosystem. Its significance in this partnership is the ability to accelerate computational workloads that sit underneath advanced engineering simulation.
Dassault Systèmes is integrating NVIDIA CUDA-X and AI physics capabilities into SIMULIA’s AI-driven Virtual Twin of Physics Behavior. The objective is to use AI models trained on physics-based data to predict engineering outcomes much faster while maintaining a connection to validated physical behavior.
That creates a different engineering loop. Instead of repeatedly running an expensive high-fidelity simulation for every possible design, engineers can use AI physics models to rapidly explore a much larger design space and then apply higher-fidelity simulation where it matters most.
The distinction is important. AI is not necessarily replacing the physics solver. It is becoming a faster way to explore what the solver tells us. That is the foundation of modern industrial AI simulation.
Foxconn Shows What Near-Real-Time Simulation Can Mean
The strongest evidence that this approach is moving beyond a laboratory concept comes from Foxconn. NVIDIA reports that Foxconn is using PhysicsNeMo AI models with Omniverse for factory digital twins and has achieved up to 150x faster computational fluid dynamics simulations for thermal analysis, reducing some workloads from hours to minutes.
That difference matters because simulation speed determines how many questions an engineering team can ask. If a thermal simulation takes hours, engineers naturally limit the number of scenarios they investigate. If a useful prediction can be produced in minutes, teams can explore substantially more possibilities before committing to physical changes.
Foxconn is also using digital twins to simulate robot operations and automated guided vehicle paths before deploying them in physical facilities.
The value is therefore not simply faster computation. It is the ability to experiment digitally before spending physical capital.
PepsiCo Provides Another Real-World Data Point
PepsiCo is taking a similar approach to manufacturing and logistics. In its collaboration with Siemens and NVIDIA, the company is using digital twins to recreate machines, conveyors, pallet routes and operator paths with the goal of testing changes virtually before modifying physical facilities.
PepsiCo reports that an initial deployment produced a 20% increase in throughput, identified up to 90% of potential issues before physical modifications, achieved nearly 100% design validation, and contributed to 10% to 15% reductions in capital expenditure.
These are company-reported results from a specific deployment, not universal benchmarks for digital twins. That distinction matters when evaluating the technology.
But the example demonstrates the economic logic extremely well: the more accurately a company can test a physical decision before making it, the more expensive mistakes it can potentially avoid. AI makes that proposition more powerful because it increases the number of scenarios that can be evaluated.
From Digital Twin to AI Digital Twin
This is the conceptual shift engineering leaders need to understand. A conventional digital twin can represent a physical asset, connect to data, visualize its state and support simulation. An AI digital twin adds predictive intelligence and AI-driven reasoning to that environment.
Instead of only asking:
“What is happening to this machine?”
an AI digital twin can increasingly support questions such as:
“What is likely to happen next?”
“What happens if we change this parameter?”
“Which configuration produces the best outcome?”
“Which failure scenario should we investigate?”
“Which design should move to high-fidelity validation?”
That turns the twin from a passive representation into an active engineering environment.
And that is precisely where the NVIDIA and Dassault Systèmes partnership becomes strategically interesting.
The July 2026 Development May Be More Important Than the Original Announcement
The February partnership established the architecture. July demonstrated where that architecture is going.
NVIDIA’s expanded Agent Toolkit now makes PhysicsNeMo and CUDA-X available as tools that AI agents can use for engineering tasks. NVIDIA describes this as enabling developers to build autonomous AI engineers with capabilities including AI physics, accelerated solvers and scientific computing.
Dassault Systèmes is developing the complementary software layer. Its July announcement made LEO available as a Virtual Companion focused on complex engineering and simulation, alongside AURA for program management and MARIE for deep science. Dassault Systèmes says these Virtual Companions can execute work grounded in industry knowledge and mission-critical expertise.
Put those developments together and the emerging workflow looks very different from a traditional engineering assistant.
An engineer could eventually define an objective in natural language. An AI agent could determine which models and simulations are appropriate, configure the required workflow, execute simulations, analyze the outputs, explore alternative configurations and return a recommendation for engineering review.
That is not simply a chatbot inside CAD software.
It is an agentic engineering workflow grounded in digital twin technology and physics.
What This Means for Engineering Teams
The immediate impact is likely to be a change in how engineers spend their time. Today, engineering teams can spend significant effort preparing models, configuring simulations, waiting for results, comparing scenarios and repeating the process. AI can increasingly automate portions of that workflow, allowing engineers to spend more time evaluating tradeoffs, validating assumptions and making system-level decisions.
The objective should not be to remove engineers from the loop. It should be to increase the number of meaningful engineering decisions they can make.
That distinction is particularly important for safety-critical industries. A fast AI prediction cannot automatically be treated as engineering truth. NVIDIA itself has highlighted the need for rigorous, repeatable evaluation of AI physics models, including transparent benchmarks and testing against engineering baselines.
For that reason, the most credible architecture is not AI replacing physics. It is AI accelerating exploration while validated physics and engineering judgment remain the foundation for critical decisions.
Add Your Heading Text The New Engineering Stack Here
The emerging architecture can be understood as five connected layers.
At the bottom is the physical system, including machines, products, factories, vehicles and sensors. Above that sits the digital twin, which represents the physical system using engineering models, data and operational context. The next layer is physics and AI physics, where traditional simulation and machine-learning-based surrogate models work together. Above that comes the AI agent, which can plan and execute engineering workflows. At the top remains the human engineer, responsible for validation, tradeoffs, creativity and consequential decisions.
That structure is important because it explains why the industry is not simply moving from simulation to AI. It is moving toward a system where simulation becomes a tool that AI can use.
What Engineering Leaders Should Measure
Companies evaluating an AI digital twin should resist measuring success through the number of virtual assets created or the amount of GPU capacity consumed.
The meaningful metrics are engineering outcomes.
How much faster can a team reach a validated design decision? How many more scenarios can engineers evaluate? How accurately does an AI physics model reproduce validated simulation results? How many physical prototypes can be avoided? How much capital can be saved? How much energy can be reduced? How much earlier can a failure be identified?
These measurements connect industrial AI simulation to business value rather than treating it as an innovation experiment. They also make it easier to determine whether an AI digital twin deserves to move from pilot to production.
The Biggest Risk Is Trusting Speed More Than Science
There is a natural temptation to celebrate a 150x improvement in simulation speed. But engineering organizations should ask a harder question:
150x faster than what, under which conditions, and with what accuracy?
AI surrogate models can be extremely powerful, but they can also fail outside the conditions represented in their training data. Rare events, unusual operating conditions and poorly represented edge cases can expose weaknesses that are invisible in a conventional demonstration.
That is why validation will become one of the defining issues in the next generation of industrial AI simulation. The future engineering workflow should therefore look more like:
AI prediction → physics constraints → high-fidelity simulation → experimental validation → engineering approval
rather than:
AI prediction → production
The winners will be the companies that combine AI speed with engineering discipline.
Why This Matters for the Digital Twin Industry
The conversation at every serious Digital Twin Summit is moving beyond 3D visualization. The important questions now concern AI physics, real-time simulation, agentic engineering, physical AI, virtual factories, autonomous optimization and science-validated world models.
NVIDIA and Dassault Systèmes are building directly into that transition. NVIDIA brings accelerated computing, CUDA-X, PhysicsNeMo and Omniverse.
Dassault Systèmes brings Virtual Twins, simulation, engineering workflows and industry knowledge. Their partnership combines those capabilities into a platform where AI can increasingly interact with physical systems through scientifically grounded digital representations.
That is why the February announcement should not be viewed in isolation. The July agentic developments make the strategic direction much clearer.
Conclusion
The next generation of digital twins will not simply tell engineers what is happening in a physical system. They will increasingly help engineers predict what could happen, test what might work, compare alternatives and decide what to do next. That is the real significance of the NVIDIA and Dassault Systèmes partnership.
CUDA-X matters because engineering simulation needs computational speed. PhysicsNeMo matters because AI needs to understand physical behavior. Virtual Twins matter because AI needs a structured representation of the real world. AI agents matter because someone or something needs to orchestrate the increasingly complex engineering workflow.
Put those pieces together and the AI digital twin becomes much more than a digital replica. It becomes an intelligent engineering environment.
The industrial winners will not necessarily be the companies with the biggest digital twins or the most AI models.
They will be the companies that create the shortest reliable path between engineering question and validated physical decision. That is where digital twin technology is heading in 2026.
And that is why the NVIDIA and Dassault Systèmes partnership may ultimately be remembered not as a faster simulation story, but as an early blueprint for AI-native engineering.
FAQ’s
An AI digital twin is a digital representation of a physical product, machine, factory, or system enhanced with artificial intelligence, real-world data, and physics-based simulation. Unlike a basic digital model, an AI digital twin can help predict behavior, evaluate scenarios, identify anomalies, optimize performance, and support engineering decisions.
NVIDIA and Dassault Systèmes announced a strategic partnership in February 2026 to combine Dassault Systèmes' Virtual Twin technologies with NVIDIA's accelerated computing, AI infrastructure, CUDA-X libraries, and Omniverse technologies. The companies are working toward science-validated industrial world models for engineering, manufacturing, materials science, and other complex industries.
NVIDIA CUDA-X provides GPU-accelerated libraries that can speed up computationally intensive engineering workloads. Combined with AI physics models such as PhysicsNeMo, CUDA-X can help engineering teams explore simulations and physics-based scenarios faster, potentially allowing more design iterations before high-fidelity validation or physical testing.
PhysicsNeMo is NVIDIA's framework for building and deploying AI models that learn from physics and scientific simulation data. It can be used to create AI surrogate models capable of predicting physical behavior much faster than repeatedly running computationally expensive simulations. NVIDIA has also made PhysicsNeMo available as an agent-ready tool for AI engineering workflows in 2026.
AI agents can potentially use digital twins, physics models, simulation tools, and engineering data as part of an autonomous workflow. Instead of an engineer manually configuring every simulation, an AI agent could eventually interpret an engineering objective, select appropriate models, run simulations, analyze results, test alternatives, and present recommendations for human validation.
A traditional digital twin represents a physical asset or system and can connect to operational data and simulation models. An AI digital twin adds machine learning, AI physics, predictive analytics, and potentially AI agents, allowing the twin to move from simply representing and monitoring a system toward predicting, optimizing, and recommending actions.
AI digital twins can potentially help engineering teams reduce simulation time, evaluate more design alternatives, identify problems earlier, reduce physical prototyping, optimize energy and material usage, and accelerate design decisions. McKinsey estimates that digital twins can reduce the time required to deploy new AI capabilities by up to 60% and reduce relevant capital and operating expenditures by up to 15%.
No. The more realistic direction is for AI digital twins and industrial AI simulation to complement traditional physics-based simulation. AI models can accelerate exploration and identify promising scenarios, while high-fidelity physics simulation and physical testing remain important for validating critical engineering decisions.
AI digital twins can be valuable in manufacturing, automotive, aerospace, electronics, energy, robotics, industrial equipment, transportation, logistics, construction, and other engineering-intensive industries. They are particularly useful where physical testing is expensive, simulation workloads are computationally intensive, or operational optimization can produce significant financial benefits.
An AI digital twin can allow engineers to evaluate many more design scenarios before committing to physical prototypes. AI models can rapidly identify promising configurations, while high-fidelity simulation can validate selected designs. This can shorten the design iteration cycle and help engineers explore larger design spaces.
There are already several 2026 industrial examples. NVIDIA reports that Foxconn achieved up to 150x faster computational fluid dynamics simulations for thermal analysis using PhysicsNeMo. Siemens reports that PepsiCo's digital-twin deployment produced 20% higher throughput, identified up to 90% of potential issues before physical modifications, and supported 10% to 15% lower CapEx. These are vendor-reported results from specific deployments, not universal benchmarks.
The main risks include inaccurate predictions, model drift, insufficient training data, failures outside the model's validated operating range, cybersecurity vulnerabilities, and overreliance on AI predictions. For safety-critical engineering, AI outputs should be validated against physics-based models, experiments, engineering standards, and human expertise before consequential decisions are made.