Nico Peng
← Back to the Physical AI index

Physical AI · Enabling market

Simulation & digital twins

The software and test systems that let builders examine a design or behaviour before relying on a real machine.

AS OF 2026-09-13DATED · NON-EXHAUSTIVE

Research reviewed 2026-09-13. Available means documented access or production, not assured spare capacity for every specification. Tight refers to the scope identified in the node, not every robot component. Supply unclear means the reviewed evidence does not settle availability at the selected boundary. Expansion describes how additional usable supply can be delivered, not a promised lead time. Pricing power describes our assessment of the ability to defend price or terms; it is not inferred from technical importance alone.

A useful rehearsal of reality

Different questions need different models

A joint designer asks whether a motor will overheat. A policy developer asks whether a grasp succeeds. A factory engineer asks whether a robot can reach a fixture without stopping the line. A vehicle team asks what its controller will do when another road user behaves unexpectedly. All use simulation, but they do not buy the same model, software or evidence.

The useful unit is a toolchain that answers one of those questions. Geometry preparation, physics, scenario authoring and results analysis are often parts of that purchase. Making each a mandatory supply node would overstate the number of independent markets. Conversely, merging all simulation into one node would conceal the difference between a downloadable physics engine and an installed hardware test system.

A digital twin is valuable when it preserves the properties needed for a decision. It need not reproduce everything about the machine. A visually convincing scene can still give the wrong contact forces; a good mechanical model can omit the sensor conditions that cause a perception failure.

More experiments are not automatically more revenue

Physical AI can increase the number of designs, behaviours and operating conditions worth testing. Faster simulation can also reduce the cost of each experiment. Supplier revenue depends on the purchased unit: developer seats, execution capacity, enterprise support, an integrated test rig or engineering services.

Open simulation changes where the opportunity lies. NVIDIA's Isaac Sim documentation distinguishes its open-source software from licensing for redistribution of the underlying Omniverse Kit. A free starting point can expand a hardware or enterprise ecosystem without creating a toll on every simulated episode. The economic question is which difficult, recurring work remains with the supplier.

Supply-chain map

Designing the machine and learning its behaviour

Engineering simulation and robot-learning environments travel across machine families. They answer different questions and can be bought independently.

Offering / activityEnables / feedsConditional route◇ Downstream boundary
Select a node to read its explanation
Designing the machine and learning its behaviourEngineering simulation and robot-learning environments travel across machine families. They answer different questions and can be bought independently.

The diamonds are downstream uses, not additional markets. A model may pass parameters between these workflows, but that does not make one software purchase a prerequisite for the other.

Engineering simulation

Available
Role
Tests physical design choices
Expansion
Software access scales readily; trusted models require engineering
Pricing power
Model and workflow switching costs; not a proven robotics premium

Why it exists

A component has to survive its complete duty, not just reach a peak specification. Engineering simulation connects electromagnetic, thermal, structural and other effects before a builder commits to hardware. The purchase is the analysis environment and its supported capabilities, not a separate market for every equation it solves.

Relevant businesses

  • Ansys, a Synopsys business: multiphysics tools for robot design, including electromagnetic, thermal and mechanical analysis.
  • COMSOL (private): a multiphysics environment; its actuator example couples force calculation and motion rather than treating the actuator as a static drawing.
  • Dassault Systèmes: SIMULIA tools for electrical-machine design and system-level operating analysis, distinct from its DELMIA workcell-planning offering below.

Worth understanding

Ansys's motor workflow links concept work to detailed physical analysis. This is distinct from a robot policy simulator: a designer can investigate motor heating without training a policy, and a policy team can use simplified motor behaviour without buying a detailed motor-design package.

The software is an established offering. That is the scope of the availability label, not a claim that expert modelling labour is unlimited. More machine designs can increase seats, solver usage and support work, while better automation can reduce analysis time per design.

Our inference is that validated models, accumulated methods and links to an engineering workflow can make replacement costly. They may support pricing even when another solver is available. The reviewed product evidence does not isolate a Physical AI price premium or segment margin; the case should be tested through retention, realized prices and how much customer work is genuinely reusable.

↑ Back to the map

Robot-learning environments

Available
Role
Generates experiments and training experience
Expansion
Open software; execution needs compute and credible task models
Pricing power
Weak direct licence toll; complements and support can be sold

Why it exists

A robot developer needs repeated interactions with a task: observations, actions and consequences. A simulation environment can generate these interactions without wearing out hardware or resetting every trial by hand. Physics, sensors, environments and training interfaces belong together when they are purchased or adopted as the working environment.

Relevant businesses

  • NVIDIA: Isaac Sim provides robot simulation and synthetic-data workflows, with Isaac Lab supporting robot learning.
  • Google DeepMind / Alphabet: MuJoCo is a freely available, open-source physics engine used for robotics and related research.

Worth understanding

The model's usefulness depends on the question. Contact, friction and sensor differences can invalidate a learned behaviour even if the simulated motion looks plausible. More generated episodes are not equivalent to more transferable learning. This is why environment calibration and real-machine evaluation remain part of the developer's work.

The public software releases support an availability judgment for the tools. They do not establish spare GPU capacity or a ready-made model for every task. Copying an environment can be inexpensive; maintaining a useful collection of assets and running many experiments still consumes engineering and compute.

Our economic inference is that open engines constrain the idea of a compulsory per-episode licence. A supplier may instead benefit through compute, enterprise tooling or support. Those revenues must be assigned to the actual paid offering, not counted again as an independent simulation fee merely because the engine helped create demand.

↑ Back to the map

Supply-chain map

Testing the installation or the vehicle

The environment, interfaces and acceptance questions create distinct validation toolchains. Neither route stands for every Physical AI application.

Offering / activityEnables / feedsConditional route◇ Downstream boundary
Select a node to read its explanation
Testing the installation or the vehicleThe environment, interfaces and acceptance questions create distinct validation toolchains. Neither route stands for every Physical AI application.

Factory and vehicle platforms can share simulation techniques without sharing a validated scenario library or a customer programme. Simulation contributes evidence; it does not by itself establish that a deployed system is safe.

Workcell simulation

Available
Role
Tests robot programmes in a production context
Expansion
Software deployment plus robot models, connectors and commissioning work
Pricing power
Installed workflow and integration can support retention

Why it exists

The workcell includes fixtures, other machines, people and production logic. A valid robot movement is not yet a workable production sequence. Workcell simulation combines layout, reach, collision, timing and control interactions so that engineers can test the installation before occupying a real line.

Relevant businesses

  • Siemens: Process Simulate X supports robot simulation and commissioning workflows; the Stäubli RCS integration supplies robot-specific controller behaviour.
  • Dassault Systèmes: DELMIA's virtual-factory tools cover production layout and simulation, including virtual commissioning.

Worth understanding

Siemens's February 2025 description of its Stäubli integration illustrates what the customer obtains beyond generic 3D animation: robot-specific behaviour within a production engineering environment. Libraries and connectors can be valuable precisely because the installed equipment is heterogeneous.

A more adaptable robot does not remove this work. It changes which behaviours can be tested and which assumptions must be verified. The demand driver is new or changed installations, not simply the number of model-training runs. A long-lived factory may continue using its engineering environment through several robot generations.

Our inference is that accumulated layouts, process models and controller interfaces can favour the incumbent platform. The competing forces are interoperable formats and the engineering effort required to maintain each installation. Software supply is established; the capacity and margin of customer-specific commissioning are separate questions, examined in Production & deployment.

↑ Back to the map

Vehicle validation platforms

Supply unclear
Role
Exercises vehicle software and hardware against scenarios
Expansion
Software can replicate; real-time test rigs need hardware and configuration
Pricing power
Programme integration and trusted test assets may matter more than raw run count

Why it exists

A vehicle team must examine the interaction of sensors, decisions and vehicle response across operating conditions. A validation platform supplies scenario execution and test workflows; hardware-in-the-loop adds interfaces to the actual controller. These are related purchased toolchains, not a market for generic simulated miles.

Relevant businesses

  • Applied Intuition (private): simulation, real-world data and validation tools for vehicle software, with integration support.
  • dSPACE (private): SCALEXIO-based hardware-in-the-loop systems test autonomous-driving computers in open- and closed-loop simulation.

Worth understanding

Applied Intuition's Neural Sim release describes sensor rendering that responds to the vehicle's changing position. This matters because replaying a fixed recording and evaluating the consequences of a different action are not the same experiment.

Supply is unclear at this combined purchase boundary. Released software and a configured real-time test rig do not expand on the same clock. The reviewed evidence establishes offerings, but not an industry queue for installed rigs; different delivery paths alone do not establish mixed supply conditions. Splitting every software module from its interfaces would obscure the system the validation team actually buys.

Our inference is that reusable scenarios, integration with a programme's data and test infrastructure, and confidence in results can create switching costs. However, a customer can retain parts of its own testing stack. Neither a large scenario count nor a platform's presence in a vehicle programme demonstrates that it captures the economics of the eventual autonomous service.

↑ Back to the map

Value comes from decisions the model improves

The common opportunity is to reduce expensive physical iteration. The commercial mechanisms differ: established engineering licences, open learning engines with paid complements, production workflows and integrated vehicle test platforms. Their revenues should not be forecast from one total for simulated hours.

The most useful follow-up is to examine which models and interfaces remain valuable across customer projects, how much new engineering each deployment needs, and whether customers keep paying as execution becomes cheaper. Those questions distinguish a reusable product from project work wearing a software label.

Related reading

Learning & data · Compute & memory · Motion · Production & deployment · Road autonomy

This dated map selects economically useful boundaries. Company participation does not establish a trade relationship between the named suppliers, and a product announcement is not evidence of accepted machine output.