The New Simulation Layer for Physical Product Development

Engineering simulation is the step where teams virtually test how a physical product will behave before building it. Traditional workflows are often slow, expensive, and expert-heavy. In this landscape, I looked at the new generation of AI-native simulation startups trying to change that: faster solvers, physics AI models, automated setup, real-time simulation, and more continuous design exploration. Below are the main things I learned about how traditional simulation works, what these new products bring, and why the category is becoming interesting from an investment and acquisition perspective.

How does engineering simulation traditionally work?

• At a basic level, simulation means testing a product virtually before testing it physically. Instead of building a physical prototype immediately, engineers create a digital version of the product and test how it behaves under real world constraints: heat, pressure, airflow, vibration, stress, fluid movement, electromagnetic effects, etc..

• The starting point is usually a CAD model, but a CAD model alone is not enough to run a simulation. A 3D design first needs to be prepared for simulation. Engineers often need to clean the geometry, remove unnecessary details, define the materials, set the boundary conditions, and describe what is happening around the object. For example: where the air enters, where the heat comes from, where the part is fixed, or what load is applied.

• One of the most important steps is meshing. Because the simulation software cannot usually calculate physics on the entire object, it breaks the geometry into many small pieces called a mesh. You can think of it as turning a physical object into a grid of tiny elements where the equations can be solved locally. The more detailed the mesh, the more accurate the simulation can be, but also the slower and more expensive it becomes.

• The solver is the engine that actually computes the physics. This was also the part I had to clarify for myself. In simple terms, the solver takes the mesh, the material properties, the constraints, and the equations of physics, and then calculates how the system behaves. Different solvers exist for different types of problems: fluid dynamics, structural stress, heat transfer, electromagnetics, multiphysics, etc.

• After the solver runs, engineers need to interpret the results. The output is not just a clean yes/no answer. It can be stress maps, pressure fields, temperature distributions, airflow patterns, deformation curves, or vibration modes. This means simulation is not only a compute problem. It is also an expertise problem, because engineers need to understand whether the result makes sense and what design decision should follow from it.

• The traditional simulation loop is powerful, but it is often slow. A high fidelity simulation can take hours, sometimes days, especially for complex products or multiphysics problems. This matters because simulation is part of the iteration loop. If every design variation takes a long time to prepare and run, engineers cannot explore hundreds of alternatives. They test a few options, not the full design space.

• A big bottleneck is that setup work is still very manual and expert driven. Before the actual computation starts, a lot of time can be spent on geometry cleanup, meshing, choosing the right solver settings, defining constraints, and checking that the simulation is physically meaningful. This makes simulation less accessible to non specialists and creates dependency on experienced CAE engineers.

• The consequence is that simulation often validates designs more than it explores them. In theory, simulation should help engineers discover the best possible design. In practice, because each run is expensive and setup heavy, it is often used to test whether a design is good enough. This is why AI native simulation is interesting: the promise is not only to make simulations faster, but to turn simulation into a more continuous, exploratory, and interactive part of product development.

What new generation engineering simulation products bring?

• The general direction is that simulation is becoming faster, more continuous, and more embedded in the engineering workflow. Traditional simulation is often a heavy step that happens after a design is already reasonably defined. What I find interesting with this new generation of products is that they try to move simulation closer to the design process itself. Instead of “design first, simulate later,” the direction is more “design while continuously simulating.”

• The first obvious promise is speed. Many of these companies are trying to turn simulation from something that takes hours or days into something that can happen in seconds or minutes. This is the core idea behind companies like Luminary, Emmi AI, PhysicsX, BeyondMath, NP Company, Vinci, or Neural Concept: using AI models, surrogate models, GPU native solvers, or physics foundation models to make simulation fast enough to support real iteration.

• This changes the role of simulation from validation to exploration. If a simulation is slow and expensive, you use it to check a few designs. If it becomes fast and cheap, you can use it to explore many more options. This is where the workflow change becomes very concrete. NablaFlow shared a case study where KASK moved aerodynamic testing into AeroCloud and increased simulations from 1 to 2 per month to 20 to 60. This is not just a runtime improvement. It changes how many ideas an engineering team can actually test.

• Some startups are packaging simulation around specific engineering use cases instead of generic simulation tools. Luminary is a good example with SHIFT-SUV for automotive aerodynamics, SHIFT-Wing for transonic wing design, and SHIFT-Crash for full vehicle crash prediction. Emmi AI does something similar with NeuralMould for injection moulding. I find this interesting because the product is not “here is a faster solver, good luck.” It is closer to “here is an AI model for this specific engineering problem.”

• Another important change is that simulation is becoming more accessible to non specialists. Traditional CAE (Computer Aided Engineering) workflows require a lot of expert setup. Some of the new products try to hide part of that complexity behind easier interfaces, automated setup, or more guided workflows. NablaFlow’s AeroCloud, for example, is positioned around cloud based aerodynamic simulation that teams can use without deep CFD expertise. This matters because simulation capacity is often limited not only by compute, but also by the number of experts who can run it.

• AI is also attacking the painful preparation layer before simulation. Simulation is not only slow because the solver is slow. It is also slow because preparing the model is tedious. This is where companies like deepmath are interesting, with AI-powered mesh generation for CAE, CFD, and multiphysics workflows. Luminary also launched with automated meshing and mesh adaptation. This part is less flashy than “physics foundation model,” but probably very important in practice.

• Some companies are building reusable physics AI models, not just running isolated simulations. PhysicsX is probably one of the clearest examples here. Their product direction seems to be around Large Physics Models, Large Geometry Models, and a Simulation Workbench that helps automate workflows, manage simulation data, and prepare datasets for AI model training. In other words, simulation becomes a data and model flywheel, not only a one-off calculation.

• Integrate rather than disrupt. Most of these products fit into existing CAD, CAE, PLM, cloud, and security environments. 

• Some companies are taking a more radical “foundation model for physics” angle. NP Co explains that it is building a transformer based industrial physics platform pre-trained on industrial physics. BeyondMath talks about generative physics. Vinci describes a foundation-model-based platform for hardware simulation. This is probably the most ambitious version of the category: not only accelerating a solver, but making simulation feel more like inference from a learned model.

• Beyond pure “tech improvement”, the biggest product shift is that simulation could become an “always on” layer in product development. Today, simulation is often a checkpoint. The companies on the landscape want to make it a continuous feedback loop: as the design changes, the physics feedback changes with it. If this works, the main value is not only faster simulation. It is faster engineering decisions, more design exploration, fewer physical prototypes, and a much tighter loop between design, simulation, testing, and manufacturing.

Investment & Acquisition Environment

• The funding environment shows that AI native simulation is a mature software category. PhysicsX raised a $135M Series B in 2025 and then a $300M Series C in 2026. Neural Concept raised a $100M Series C. Luminary Cloud raised a $72M Series B. So investors are clearly thinking that some of these startups can become large companies.

• The investor mix is broader than classic Software VC. Of course, you see well-known VC funds like Atomico, General Catalyst, Khosla, Eclipse, Speedinvest and others. But you also see strategic and industrial names like Siemens, Applied Materials or NVIDIA.

• A capital intensive market.  There are many large rounds, which means that the category is capital intensive. Building physics AI models, enterprise integrations, infrastructure, and industrial grade workflows probably requires much more than a classic SaaS product team.

• Obviously, most of these companies are founded by experts and industry insiders. The founding teams are often people who spent years inside very specific technical worlds: F1 aerodynamics, Stanford aerospace, CFD research, computational hydrodynamics, quantum simulation, mesh generation, or EPFL computer vision.

• Some companies are funded around vertical wedges, while others are funded around horizontal platforms. NablaFlow starts from aerodynamic and wind simulation. Vinci is very focused on semiconductor and hardware simulation. Emmi AI launched NeuralMould for injection moulding. Luminary has model families for automotive, aerospace, defense, and crash prediction. PhysicsX and Neural Concept feel more horizontal, trying to become broader AI engineering platforms. So you can find both vertical and horizontal plays here.

• The Emmi AI acquisition by Mistral AI is probably the most interesting acquisition signal in the landscape. What is surprising is not that Emmi was acquired, but that it was acquired by a large language model provider, not by a traditional engineering software company like Siemens, Autodesk, Dassault, or Ansys. My interpretation is that Mistral did not buy Emmi only to own a simulation tool. They probably bought a team and technology stack that can help them move from language models toward scientific and industrial AI models (diversification).

• There’ll probably be many more acquisitions in the next months/years. Traditional engineering software companies (Siemens, Dassault, Autodesk, Ansys, Cadence…) are used to buying simulation assets to complete their platforms. So I would expect more “AI-native simulation” M\&As in the coming years.