How much is CAD data worth to AI companies?

How much is CAD data worth? Until recently, almost nothing outside your own company. That has changed. AI companies are building text-to-CAD tools and engineering agents, and they need real design files to train them. This guide is the CAD-specific version of our general guide on how much company data is worth. It covers why buyers want CAD, which formats count most, what raises and lowers value, and the price ranges that have actually been published.

Why AI companies want CAD data for AI training

Language models learned to write from billions of public web pages. There is no equivalent pile of professional engineering design on the open web. Most real assemblies, drawings and design histories never leave the companies that made them.

Meanwhile, demand is growing on several fronts:

• Text-to-CAD tools. Zoo, the platform launched by the company then called KittyCAD, offers Text-to-CAD that generates B-rep surfaces you can import as a STEP file and edit in existing CAD software, rather than a mesh you can't edit. For enterprise customers, Zoo advertises the ability to "train on your proprietary data."
• Research models. Autodesk trained an experimental model on about ten million 3D shapes from publicly available data, and has said such models become much more useful when trained on larger, higher-quality professional datasets.
• Buyer listings. A published broker price list citing Mercor shows buyers asking hardware engineering firms for CAD assemblies, circuit designs and test results from real products, and asking construction firms for project records, drawings and site imagery from completed projects.
• Trade commentary. One CAD and AI newsletter called CAD data and engineering RL tasks "the gold rush of 2026." Treat that as commentary, not market data, but it captures the mood.

The upshot: a professional archive that has sat untouched on a file server may now be worth assessing. Through a license, you keep ownership and choose what is included.

What public CAD datasets lack

To see why your archive could matter, look at what researchers train on today. The main public datasets each have clear limits:

• ABC holds over a million B-rep models from the Onshape platform, but according to the SldprtNet authors it lacks modeling sequences and any text annotations.
• DeepCAD contains about 170,000 CAD sequences, but it is limited to two command types: 2D sketching and extrusion.
• Fusion 360 Gallery includes real modeling histories, but focuses mainly on basic extrusion-based models.
• SldprtNet, a newer dataset of over 242,000 parts with native SolidWorks files, was collected from public repositories such as GrabCAD, McMaster-Carr and FreeCAD.

All of these come from publicly shared models. None of them shows how a real product is engineered. There are few large, multi-level assemblies. No drawing sits next to the model it documents. There are no BOMs, no GD&T, no test results and no revision history showing how a design changed after a failed test.

That is the gap a professional archive fills. Your files show real constraints, real tolerances and real trade-offs, and that is what buyers can't get from public sources.

Native CAD vs STEP: value tiers by format

Format matters a great deal. The key question is how much of the design intent survives in the file. Text-to-CAD models have to learn the sequence of operations behind a part, not just its final shape. The SldprtNet authors note that the native .sldprt format records the feature operations and parameters used to build the model.

Two notes on BIM data for AI training and on electronics. An IFC export is useful for exchange, but it usually flattens Revit families and the parametric relations between elements, which are part of what makes a native model informative. In the same way, Gerber files describe the layers to manufacture, while the native EDA project holds the schematic and the design logic behind the layout.

The practical advice is simple: don't convert anything before an assessment. Keep native files native, and don't merge drawing revisions or print models to PDF to "tidy up."

What multiplies the value

Format sets the baseline. Context multiplies it. These are the factors that move an archive from interesting to in demand:

• Pairing. A model on its own is worth less than a model with its drawing, BOM, GD&T or PMI, specifications, calculations, engineering change orders and test or simulation results. Pairing lets a buyer connect a requirement to a design to a result. The buyer listing for hardware firms says price depends on "assembly count and completeness."
• Revision history. Complete projects with every plan revision are worth more than single drawings without context. A design that changed after a failed test shows reasoning that a final file never does.
• Full project records. Projects that can be followed across systems, from quote to acceptance, including emails and tickets, are particularly sought after.
• Years and variety. Several product generations, many part types and different customers or applications give a buyer more to learn from than thousands of near-identical variants.
• Clean rights. The buyer needs to know the files can be licensed. Work made for customers, under contracts with IP clauses or with partners varies from case to case – see who owns CAD files for how it generally works.

What lowers the value

Some things pull value down, or rule files out entirely:

• Third-party library parts. Fasteners, bearings and other purchased components downloaded from vendor catalogs aren't your IP. They add volume without adding value and usually have to be filtered out.
• Unclear ownership. If customers own the designs, or contracts are silent or restrictive, those files may have to be excluded.
• Export-controlled content. Designs covered by ITAR or the EAR, or defense-related work, need careful screening and are often excluded. Our guide on preparing CAD files for AI licensing covers how to sort them.
• Duplicates. Copies, near-copies and files saved under many names inflate the count without adding information.
• Only exports. An archive that exists only as PDFs, or where native files were thrown away after STEP export, sits lower in the format tiers.

The price ranges published so far

There is no public price list for CAD data, and no one has published a per-file or per-GB price. Anyone quoting you a figure per drawing or per gigabyte is guessing. What does exist is a handful of buyer ranges per company.

A published broker price list citing Mercor shows these buyer listings, all per company:

Read these figures carefully:

• They are ranges, not quotes. They are what buyers say they may pay, a best case rather than realized prices.
• They are per company or engagement. They are not per file, per drawing or per project.
• They are not guaranteed. Every archive and every deal is unique. Your archive can land below, above or outside these ranges, and some archives can't be licensed at all. The hardware listing in particular is based on few referrals so far.

A license also shapes the economics. By default it is non-exclusive and limited in scope, time and use, so the same archive can in principle be licensed to more than one buyer over time. An exclusive license may pay more per deal but closes the door to others.

To see where your archive might land, describe it in our form on what is your data worth. You get an indicative estimate per buyer, built on these published ranges. It is an estimate, not an offer.

Is my CAD data valuable? A quick self-check

You don't need to send any files to get an assessment. Go through the questions below. The more you can answer yes, the stronger your position:

1. Native files? Do you have SolidWorks, Inventor, Creo, NX, Fusion, Revit or EDA files with the feature history intact, not just exports?
2. Real assemblies? Are there assemblies from shipped products, not only single parts?
3. Volume? Roughly how many assemblies or completed projects do you have? The estimator counts native CAD assemblies for hardware firms and completed projects with full records for construction.
4. Paired records? Do drawings, BOMs, specs, calculations, ECOs and test or simulation results sit alongside the models?
5. Revisions? Is the revision history kept, or only the final version?
6. History and variety? How many years and product generations does the archive cover?
7. Ownership? Did you design the files yourselves, and do customer contracts allow you to license them?
8. Third-party content? How much is vendor library parts or customer-supplied geometry?
9. Export control? Is any of it covered by ITAR, the EAR or defense contracts?
10. Order? Can you find what belongs to which project, or are files scattered with duplicates?

If you are unsure about several of these, that is normal. Working them out is part of the assessment.

How a license works

You license the data. You don't sell it outright. You keep ownership of your files, the license is non-exclusive by default, and you approve every deal before anything happens. A non-disclosure agreement is in place before anything is shared, and nothing leaves your systems until you have approved a concrete offer. The assessment is free, and we are paid only when a deal goes through.

When you are ready, read how to sell CAD data to AI companies, or start by describing what you have.

Summary

CAD data is in demand because text-to-CAD tools and engineering agents need professional designs, and public datasets are mostly simple, publicly shared parts with no drawings, BOMs, test results or revision history. Native files with feature history sit at the top of the value tiers, and pairing, revisions and full project records multiply that value. Library parts, unclear ownership, export-controlled content and duplicates pull it down. Published buyer ranges run from tens of thousands to over a million dollars per company. They are ranges, not quotes, and no per-file price is public. If your archive passes the self-check, it may be time to license your CAD archive.