- No buyer publishes a price for a simulation archive on its own.
- Simulation is bought today as expert work, not archives: engineers are paid by the hour to build tasks.
- Runs that are validated against test data, with meshes, inputs and solver settings kept together, are the usable ones.
- Simulation-AI companies mostly train on a customer's own runs inside the customer's systems; they are not archive buyers.
- Simulations made for clients usually belong to the client under the contract.
Who buys CFD and FEA simulation data from engineering firms for AI training?
Nobody publishes a programme for simulation archives alone. Simulation shows up in four places: as a nice-to-have inside hardware archives, as paid expert work, inside liquidation lots, and as the input to software that a firm runs on its own data.
| Buyer or source | Stated range or rule | Records asked for | Type of figure | Date checked |
|---|---|---|---|---|
| Mercor, hardware engineering listing | $45K to $1.35M per company | FEA and CFD simulation files (nice to have), with native CAD assemblies required | Published range for the whole archive | 19 Nov 2026 |
| Mercor, Hephaestus project | $70 to $90 an hour (posted rate) | New CAD, FEA and CFD tasks with reproducible solver runs, built by engineers | Posted hourly rate for expert task work | 19 Nov 2026 |
| A buyer posting on Cad Crowd | No figure | Project bundles where simulations and physical test results are valuable | None | 19 Nov 2026 |
| Hilco Global, Project Beacon | No price; NDA sale process | 373 magnetics simulation models, 287 with solved results, in a wider archive | Liquidation lot | 19 Nov 2026 |
| DimensionLab and PhysicsX | Not buyers of archives | A firm's own simulation runs, used to train models for that firm | Software and services | 19 Nov 2026 |
| DrivAerNet++ (public dataset) | Free | 8,000 car designs with high-fidelity CFD | Public data that competes with private archives | 19 Nov 2026 |
No buyer has published a price for a CFD or FEA archive on its own. The Hephaestus work "focuses on native CAD and analysis artifacts, physically sound models, reproducible solver runs, and measurable design outcomes", paid at "$70-$90 per hour". In other words: simulation is bought today as expert work, not archives. Mercor's hardware listing also asks companies "Do you keep simulation or test results?" when it scopes an archive.
How much is an archive of simulation runs, meshes and results worth to AI companies?
There is no public price for a simulation archive, so any number is set in negotiation. What the public record does show is how the value is judged.
- Inside a hardware archive. Simulation files raise the value of a native CAD archive in the $45K to $1.35M band. They do not qualify a company on their own.
- Against free data. A public dataset already offers "8,000 diverse car designs modeled with high-fidelity computational fluid dynamics (CFD) simulations", "With more than 39 TB of publicly available engineering data" (DrivAerNet++). A private archive has to offer what public sets do not: real products, real loads and test results.
- Against labour. One project posts an hourly rate of $70 to $90 for an engineer to build a checked task. An archive is worth more than that only if many runs can be used without rebuilding them.
- With results attached. In Project Beacon, Hilco counts "373 FEMM magnetics / coil simulation models, including 287 with solved-result files" (Hilco). Solved and checked runs are the unit that gets counted.
What makes a CFD or FEA archive usable for AI training?
Validation, first. NASA's CFD tutorial quotes the AIAA definition of validation, "The process of determining the degree to which a model is an accurate representation of the real world from the perspective of the intended uses of the model", and adds that "One can only validate the code for a specific range of applications for which there is experimental data" (NASA Glenn). A run that was checked against a test is evidence. A run that was never checked is a guess.
Second, completeness: the geometry, the mesh, the boundary conditions and loads, the material data, the solver and its settings, the results, and the decision the run supported, kept together. Mercor's Hephaestus brief asks engineers to "assess mesh quality, convergence, sensitivity, and physical consistency", which shows what a buyer checks. A results file without its inputs cannot be rerun or trusted.
- Keep each run with its CAD revision, mesh, inputs, solver version and settings, and results.
- Mark which runs were validated against physical tests, and keep the test data.
- Keep the design decision or change order each study led to.
- Record the solver and license used for each run.
Can an engineering consultancy license its Ansys and Abaqus simulation archives to AI companies?
Yes, for the runs it owns, but most consultancy runs were made for clients. Three checks decide what can go in. First, the client contract: who owns deliverables, and whether the models hold the client's confidential geometry. Second, the solver license: your Ansys or Abaqus license covers the software; read it for any limit on how results or exported files are used, and keep the solver installation itself out of the deal. Third, export control: a study of a defense article is ITAR technical data.
Internal research, method development and studies on your own products are usually yours. Those are also the runs most likely to have been validated, because they were done to build your methods.
On the practical side, an Ansys Workbench project or an Abaqus job is only useful if it opens and reruns. Archive each study as a whole: the geometry, the mesh, the input deck or project file, any user subroutines or scripts, the material cards, the solver version and the result files. Note which studies used custom material models or proprietary client data, because those are the parts a buyer will ask about first.
Do not hand over license files, license server details or solver installers. A buyer licenses your engineering work, not your software seats, and it will run any reruns on its own licensed copies.
If our FEA and CFD simulations were done for clients, can we license them for AI training?
Usually not without the client's written consent. Under US copyright law your staff's work belongs to your firm as a work made for hire, "A work prepared by an employee within the scope of his or her employment" (U.S. Copyright Office). But most engineering services contracts assign the deliverables to the client and treat the client's geometry and loads as confidential. Mercor's hardware listing says the same from the buyer's side: designs owned by your customers are not a fit.
What you can often keep: your methods, templates, material libraries, scripts and internal benchmark studies. The design consultancy ownership guide covers the contract terms in detail.
Which companies help engineering firms license simulation data for physics AI models?
Two groups work with simulation data in public, and they do different things. Simulation-AI companies help a firm use its own runs. DimensionLab says "Every engineering organisation that relies on simulation is sitting on an archive: years of CFD and CAE runs" and offers to turn them "into physics models you can train, evaluate and deploy on your own infrastructure" (DimensionLab). PhysicsX, which raised a $300 million Series C in June 2026 (M&G), says "Its AI models predict physical behavior in seconds rather than hours or days" (PhysicsX). These are reported company claims, and neither publishes a programme that pays firms for archives.
The second group is about licensing archives to AI developers: hardware data programmes take simulation inside a wider archive, and independent data partners help owners license their own archives. We help you license your data. Origin Data Partners is an independent data partner. Your company decides what is shared and signs any license itself, and a buyer may pay Origin when a deal closes. If your firm holds validated runs on designs it owns, the estimate is the first step. The hardware engineering data hub lists every buyer that names engineering records.
An AI data company offered to buy our old simulation data. Is that legit, and what should we check before we sign?
It can be legitimate, but check the offer harder than usual, because no buyer publishes a simulation archive price you can compare it with.
- Who is the end user, and will the files be resold or pooled with other firms' data?
- Which projects, years and clients are in scope, and has each client consented?
- Is it a license for training or an outright sale of the files? Keep ownership where you can.
- What happens to the files and to trained models when the license ends?
- When is payment made, and is any part held back until the buyer accepts the data?
The data offer checker and how to review a data licensing offer go through each term. This page is general information, not legal advice.
Which simulation archive is worth more? A worked example
This is an illustration written for this page, not real firms or offers. Two engineering firms each hold about 4,000 CFD and FEA runs from ten years of work.
| Archive | What it holds | How a buyer would read it |
|---|---|---|
| Firm A: in-house product team | Runs on its own pumps and valves, each with CAD revision, mesh, inputs and solver settings; 300 runs matched to lab tests | Owned, complete and partly validated. Strong as part of a hardware archive with the native CAD |
| Firm B: consultancy | Runs for 60 clients, results files kept, meshes deleted to save space, no test data | Mostly client-owned and not reproducible. Little can be licensed without consent, and what remains is hard to check |
The difference is not the run count. It is ownership, completeness and validation, the same three things that decide every simulation deal discussed on this page.
What is the first step for a firm that wants to license its simulation archive?
Count the runs you own outright, mark which were validated against physical tests, and check whether the meshes, inputs and solver settings were kept. Then decide whether the runs are worth more inside a wider hardware archive with the native CAD, which is how buyers name them today; see the hardware engineering data hub. Origin's estimate asks the same questions, and nothing leaves your systems until you decide.
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
Do AI companies pay for old Ansys or Abaqus result files on their own?
No buyer publishes a price for result files on their own. Results without the geometry, mesh, loads and solver settings cannot be rerun or checked, so they are worth little. Simulation is bought today as expert work, at a posted hourly rate of $70 to $90 on one project, or as part of a wider hardware archive with native CAD.
I am the lawyer for an engineering consultancy. Can it license its simulation archive for physics AI, and what should the contract say?
It can license runs it owns, such as internal studies and method development, and client work only with written consent. The contract should list projects and years, exclude client-confidential and export-controlled studies, limit use to training, bar resale and pooling, require deletion at the end, and say who answers if restricted data is found later.
Does a simulation-AI company buy our runs when we use its software?
Not in the cases we found. DimensionLab describes training physics models on a firm's own runs, deployed on the firm's own infrastructure. That turns your archive into a tool for your engineers rather than a sale. Read any software agreement for clauses that let the vendor use your data to improve its products.
Are public CFD datasets a threat to the value of our archive?
They set the floor. DrivAerNet++ alone offers 8,000 car designs with high-fidelity CFD and more than 39 TB of public engineering data. A private archive earns its price by offering what public sets lack: real products, real loads, validation against physical tests, and the design decisions each study supported.
Which simulation records should a firm keep if it may license them later?
Keep each run with its CAD revision, mesh, boundary conditions, material data, solver version and settings, and results, plus any physical test that validated it and the decision it supported. Store them by project and date. A complete, validated run can be reused; a results file on its own usually cannot.
A data company offered $50,000 for our CFD archive. Is the offer fair?
There is no public benchmark, because no buyer has published a price for a simulation archive. Judge the offer by what it covers: how many runs, whether they are validated and complete, whether any belong to clients, and whether the buyer may resell or pool them. Ask whether the runs would be worth more licensed later with your CAD archive.
Is FEA of a defense part export controlled when shared with an AI developer?
Yes. ITAR technical data covers information required for the design, testing and other work on defense articles, and a structural or thermal study of a defense part is that kind of information. Leave defense programs out. For commercial work, check whether the item has an ECCN before sharing.
Sources and further reading
- Mercor: Hardware Engineering Company listing (listing data, read November 19, 2026)
- Hilco Global: Project Beacon, wireless powering tech code and technical data assets
- DimensionLab: Introducing Frontier (November 3, 2026)
- PhysicsX: $300M Series C announcement (June 8, 2026)
- M&G plc: Catalyst strategy invests in PhysicsX (June 8, 2026)
- arXiv: DrivAerNet++, a large-scale multimodal car dataset with CFD simulations
- NASA Glenn Research Center: CFD validation assessment
- U.S. Copyright Office: Circular 30, Works Made for Hire
- eCFR: 22 CFR 120.33, technical data (ITAR)
