Selling CAD files and engineering drawings for AI training
25 min readAIData.Broker research29 Oct 202653 records, 52 sources
AIData.Broker is a sell-side broker: it values a company's data free and describes it to every buyer that fits at the same time, without the company's name, for a fee only out of a sale.
- No data upload
- No account
- Free
Short answers · figures from the market's records, last reviewed 29 Oct 2026
Can a company sell its CAD files and engineering drawings to AI companies, and what are they worth?
Yes, the designs the company owns, if their release is not a controlled export. Buyers take CAD as part of a company's records: Mercor's hardware listing asks for native assemblies with feature trees, not only exported meshes, with drawings and test results, and puts what such a company earns at $45K to $1.3M; Telegraph Lab reports paying an engineering firm $450K for records that included its CAD drawings. Customers' designs stay out without their consent, and CAD files are technology under US export rules, so releasing controlled designs to a buyer's foreign staff is a deemed export; defense drawings are ITAR technical data and stay out.
Yes, a company can license its CAD files and engineering drawings to an AI buyer, if it owns the designs, the export-control rules on what they describe allow the release, and the files leave as native assemblies with their history rather than as loose parts. Buyers now ask for CAD by name. Mercor's listing for hardware engineering companies says " License CAD assemblies, circuit designs and test results from real products" and puts what such a company earns at $45K to $1.3M; its listing for construction companies asks for drawings, specs, RFIs, submittals, change orders and daily logs from completed projects; Handshake AI lists AutoCAD among the sources it buys; and Telegraph Lab says it paid $450K to a 70-person engineering firm for records that included its CAD drawings, calculations and revisions. What they pay for is depth: full multi-part assemblies in native CAD with feature trees, not only exported meshes, tied to the drawings, simulations and test results of real products. Two things stop more CAD sales than price does: designs that belong to the company's customers, and designs whose release is an export.
This guide is for the owner, engineering lead, officer or trustee of a company that holds CAD: a hardware or product design company, a machine builder, an automotive, aerospace or robotics supplier, a contract manufacturer, an architecture or engineering firm, or a closed company whose design vault is still on a server or in the cloud. It covers what buyers value in a CAD archive, who buys it and what they pay, whose designs a company may license, the export-control rules that apply to drawings and CAD files, what the CAD vendors' own terms say, how to prepare a lot, the case of a closed company, and how a sale through AIData.Broker works. Every quote comes from the page, regulation or licence it names, read on the day shown in its record.
This guide is research from public sources, not legal advice.
Contents
- What buyers want in a CAD archive
- Who buys CAD files and engineering drawings
- What a CAD archive is worth
- Whose designs are they?
- Export controls: when licensing a drawing is an export
- What your CAD software's terms say
- Preparing a CAD lot
- A closed company's design vault
- How a sale through AIData.Broker works
- Questions people ask
- What this means for you
1. What buyers want in a CAD archive
A CAD archive is more than its part files. A native file, saved in the CAD system's own format, keeps the feature tree: the ordered list of sketches, extrusions, holes, fillets and constraints the engineer used to build the part, so the model can be edited and its intent read. An exported mesh (an STL or OBJ, the triangles a 3D printer or a viewer uses) keeps only the final shape. Around the files sit the assemblies that put parts together, the drawings with their dimensions and tolerances, the bills of materials, the circuit boards and their Gerber files, the simulations (FEA for stress, CFD for flow) and the test results, and in a product data management vault (PDM) the revision history of every one of them.
Buyers want the whole of that, joined up. Mercor's hardware listing asks for engineering teams with native CAD vaults and wants full multi-part assemblies in native CAD with feature trees, not only exported meshes, with drawings, circuit designs, simulation files and measured test results, across several products or product generations; it names automotive, aerospace, robotics and industrial equipment as its priority. It lists single parts or parts catalogs as not a fit. The reason is what a model learns from each: a finished shape teaches geometry, while a feature tree, a revision history and a test result teach how an engineer gets from a requirement to a part that works. Human-data platforms show the same appetite from the other side: Prolific recruits mechanical engineers to complete engineering design tasks while their tool usage, clicks and design decisions are captured to train AI on engineering problem-solving.
Why the free CAD collections do not meet it
There is a lot of public CAD, and almost none of it can be used to train a commercial model. Autodesk's Fusion 360 Gallery dataset, the best-known public set of designs with their construction steps, may be used only for non-commercial research purposes. The ABC dataset collects one million models from public Onshape documents, and its own records say the copyright of these CAD models is owned by their creators, not by the dataset. Sellers of compiled collections publish samples, not prices: InfoBay AI's CAD listing on AWS is a free sample for research, with the full collection licensed on request. A company's own vault of real products, with their history and their test results, is the thing that is scarce, which is why buyers name depth rather than file counts.
| Part of the archive | What it is | What a buyer values in it |
|---|---|---|
| Native part and assembly files | SolidWorks, Creo, NX, Inventor, Fusion or Onshape files with their feature trees | How the design was built, not only its shape |
| Revision history in the PDM vault | Every released version, with who changed what and why | How designs change in response to tests, failures and customers |
| Drawings | Dimensioned, toleranced 2D drawings, often as PDF beside the native file | How engineers communicate a part to manufacturing |
| Circuit designs | Schematics, board layouts and Gerber files from Altium, KiCad or similar | Electronics design, which is scarcer than mechanical |
| Simulation and test results | FEA and CFD setups and results, lab and field test data tied to a design | The outcome: whether the design worked |
| CAM programs and setup sheets | How a shop turned the model into toolpaths | Manufacturing knowledge, which Mercor asks for among records from the shop floor |
| Project records around the files | Specs, change orders, RFIs, tickets, email | Why each change was made |
2. Who buys CAD files and engineering drawings
CAD is bought in two ways. Some buyers take it as one part of a company's whole record, priced per company; others publish listings for the kinds of company that hold it. No buyer we found publishes a price per file, per part or per assembly. The table is every buyer whose own page names CAD or engineering drawings. For every buyer and its terms, see who buys AI training data.
| Buyer | What its page names | What it says it pays | What to know |
|---|---|---|---|
| Mercor, hardware | " CAD assemblies, circuit designs and test results from real products" | $45K to $1.3M a company | Native assemblies with feature trees; designs the company owns |
| Mercor, construction | Drawings and specs from completed projects, with BIM and Revit models | $115K to $1M a company | Tied to site photos and project records; asks about owner contracts |
| Mercor, manufacturing | " Production, maintenance and inspection records from the shop floor", CAM programs among them | $100K to $1.5M a company | Excludes data restricted by customer contracts |
| Handshake AI | AutoCAD among the sources it buys | $100K to $4M+ a partnership | Buys a company's records across its systems |
| Polyshares | Design files among a company's operational data | $100K to $2M and above a licence | Its licence is exclusive by default |
| Telegraph Lab | CAD among the formats in company drives | Paid an engineering firm $450K | Its own report of a past deal |
Two reported purchases show what else moves. A buyer-side referral network reported in October 2026 that a lab buyer took an industrial parts manufacturer's archive with 11 years of CRM, accounting, inventory and orders, and CAD drawings of client orders, and, from a builder of engineering training environments, CAD and PCB design tasks, all graded automatically. The second is a different product: tasks made from designs, each with a check of whether a model's answer is right, which is how training environments are sold. A company with a deep vault can be the source such tasks are built from.
Who trains on CAD but does not buy it
The CAD vendors themselves are building AI into their products, and their own terms say what they may do with customers' designs (section 6). None of them publishes a programme for buying a company's archive. A company should read its vendor's terms before offering its CAD elsewhere, because a buyer will ask whether the same designs have already gone into someone else's training.
3. What a CAD archive is worth
There is no public price per CAD file, so a holder prices its archive the way buyers do: per company, from what the archive shows about real products. Mercor's three listings that name CAD, drawings or CAM put what a company earns at $45K to $1.3M for hardware engineering, $100K to $1.5M for manufacturing and $115K to $1M for construction, each depending on how many assemblies, records or projects the company holds. Telegraph Lab's $450K engineering firm is the one disclosed payment for records that included CAD drawings. Across all buyers, published offers for one company's records have a median of $280K per company and the middle half of public prices runs $132K to $485K per company; completed sales are fewer and lower, with a median of $72.5K per company across the handful on record, so read the offer figures as what buyers say before they have seen the data.
What moves the price
- Native files with history. An archive of native assemblies with feature trees and PDM revisions is what the listings ask for; exported meshes and PDF drawings alone compete with free data.
- Outcomes joined to designs. Simulation results, test reports, field failures and change orders tied to each design are what turn geometry into engineering. Mercor's hardware listing asks for the drawings and the test results that go with them.
- Several product generations. The same product redesigned over the years shows why it changed. One generation of one product is a thin lot.
- Scarce domains. Electronics, aerospace, medical devices and robotics are rarer than brackets and enclosures, and Mercor names automotive, aerospace, robotics and industrial equipment as its priority. Scarcity raises price across the market ( scarcity beats volume).
- Clean ownership and classification. A lot where every design is the company's own and classified for export can go to every buyer at once. A lot that has to be sorted after a buyer asks loses time and leverage.
A worked example
An example, not a real holder. A robotics company with its own engineering team has designed three generations of a mobile robot over eight years. Its SolidWorks PDM vault holds the native assemblies with their revision history, its Altium projects hold the boards, and its test lab kept results for each prototype. It also built parts to order for two customers. Sold as a hardware engineering archive, the robot's own designs with their tests and revisions are what the hardware listing describes, and the customer parts come out before anything is offered. If any part of the robot was designed for a defense programme, those files are sorted out under section 5 first. The valuation gives a range for such a company from its own profile, with no upload.
4. Whose designs are they?
Ownership is the first question every buyer asks about CAD, and the one most often answered badly. A company holds three kinds of design files, and only the first is plainly its own to license.
Designs the company made for its own products. Drawings and models made by employees in the course of their work belong to the company: a work "prepared by an employee within the scope of his or her employment" is a work made for hire, and for such a work "the employer or other person for whom the work was prepared is considered the author". Work by contractors and design consultancies belongs to the company only if they signed an assignment; buyers' lawyers look for "Signed invention and IP assignment agreements for all contributors, including contractors and agencies; gaps listed and excluded." Our design files guide works through employees, contractors and agencies in detail.
Designs the company made or holds for its customers. A contract manufacturer, a machine shop, an engineering consultancy or an architecture firm holds a great many drawings that its customers supplied or paid for. Their ownership and use are set by the customer contracts and the confidentiality agreements signed with each customer, and many of them belong to the customer outright. Mercor's hardware listing puts it plainly: it wants designs the company owns, not customer IP, and its manufacturing listing excludes data restricted by customer contracts. A buyer's rights check asks whether the data contains other companies' confidential information, and the answer for a job shop's vault is usually yes. One reported sale included CAD drawings of client orders; whether those customers agreed is not public, and a holder should not assume a buyer will take the risk.
Content the company bought or downloaded. Vaults fill up with things nobody at the company made: manufacturers' models of standard parts, purchased component libraries, models from public sharing sites, templates and sample files that came with the software. Each comes with its own licence. Buyers expect that bought and licensed material is identified and removed before packaging, and the vendors' own terms on their libraries are in section 6.
Trade secrets travel with the files. Most of a design archive's value to its owner is that competitors cannot see it, and US law protects it as a trade secret only while the owner has taken reasonable measures to keep it secret. A licence to an AI buyer is a disclosure. It needs a confidentiality agreement before any file is shown, terms that limit the buyer to training and evaluating models, and an answer to whether the buyer's models could reproduce a design. Our guide on proving data quality without sharing it covers what a buyer can be shown before anything leaves.
5. Export controls: when licensing a drawing is an export
CAD files are what US export controls were written about. The Export Administration Regulations (EAR), which cover most commercial and dual-use items, define the technology they control to include "blueprints, drawings, photographs, plans, diagrams, models, formulae, tables, engineering designs and specifications, computer-aided design files". The International Traffic in Arms Regulations (ITAR), which cover defense articles, define technical data as information "required for the design, development, production, manufacture, assembly, operation, repair, testing, maintenance, or modification of defense articles", including blueprints, drawings, photographs, plans, instructions, or documentation.
An export does not need a border. Under both sets of rules, handing controlled files to a foreign person inside the United States is an export: the EAR treat releasing or otherwise transferring "technology" or source code (but not object code) to a foreign person in the United States as a "deemed export", and the ITAR do the same for releasing or otherwise transferring technical data to a foreign person in the United States. An AI buyer's engineers, annotators and contractors are often foreign nationals or abroad, and a training licence makes the files available to all of them. A licence of controlled CAD to such a buyer is an export, made by the company that licensed it.
Most commercial designs are lightly controlled, but that is a classification, not an assumption. Items subject to the EAR that are not listed on the Commerce Control List are designated as "EAR99", the lowest level of control, where much of the technology for ordinary consumer and industrial products sits. Technology for items that are listed can need a licence, depending on the item and on the foreign person's country. ITAR technical data needs an export authorization before a foreign person sees it, whatever the country. What decides is what each design is for, not the file format.
The practical rule for a holder:
- Classify before you count. Before an archive is described to any buyer, sort its designs by what they are for, and have the export classification of each product line confirmed by whoever does it for the company's products today.
- Keep ITAR files out. Defense articles' drawings and models do not go into an AI training lot without an export authorization; in practice they stay out.
- Ask the buyer where its people are. A licence for EAR-controlled technology should state who at the buyer may access the files, from which countries, and that the buyer will not re-export them; a buyer that cannot say is not a buyer for that part of the archive.
6. What your CAD software's terms say
The software a design was made in has terms of its own, and they decide three things a buyer will ask: whether the company owns what it made, whether the vendor has already trained on it, and whether anything in the licence limits what a buyer may build with it. We read the terms of the five largest CAD and circuit-design vendors on 29 October 2026.
Ownership is the company's in every case. Autodesk's terms say " You will retain Your ownership rights to files, designs, models, data sets, images, documents or similar material created by You"; Dassault Systèmes' licence says the customer remains the owner of its data in Dassault's online services; PTC's Onshape terms say the customer maintains ownership of its files, designs and models; Siemens will not acquire any title to or ownership of Customer Content, output included; and Altium's agreement leaves the customer the rights in files, designs, models, data sets, images, and documents created by You. Autodesk's own summary of its December 2025 update adds that " you may train on your models and data, provided you do not use Autodesk offerings to copy or reverse engineer Autodesk products".
Two vendors limit what the files may be used to build. Dassault's licence bars a customer from using, or letting a third party use, any output of its software, " including content, data or results generated by the use of any 3DS Offering", to develop software competing with Dassault's, " including through artificial intelligence model or system training". Altium bars using its products or their output to develop or enhance any product that competes with an Altium product or service. A buyer training a general model is not obviously building a CAD or circuit-design product; a buyer training a design assistant may be. A company licensing SOLIDWORKS, CATIA or Altium files should ask what the buyer is building, and put a term in the licence that keeps the files out of products competing with the vendor whose software made them. Autodesk's restriction is narrower: it bars using its software to replicate the functionality of any Offering, including through training of an AI model.
The vendor may already have trained on your designs, or may not. Autodesk lets an administrator decide, hub by hub, whether " your Customer Content in that scope will be used for future Advanced AI training"; PTC commits that it will not use Customer Data to train its AI models; Altium's agreement licenses it to use customers' data to develop, enhance, support and improve its products without naming AI training. A buyer paying for an exclusive licence will ask, so check the settings before you offer one.
The way out matters. PTC's cloud agreement (Creo+, Windchill+ and its other hosted services, not Onshape) forbids using its API to extract data from the Service for the purpose of training, fine-tuning, or creating an artificial intelligence (AI) model without PTC's consent. As with every system, the export should be the company's own, made through the vendor's export, never through a buyer's connector.
Files made on free, student and academic seats are not commercial files. Onshape makes every document created on its free plan public, with a licence to anyone who sees it to use, copy and " sell copies of the Document", and its free versions are for creating and editing intellectual property for non-commercial purposes. Autodesk's education users may not access or use the Education Offering for commercial or professional purposes, and its free personal Fusion licence is for personal, non-commercial projects only. Dassault's education licences bar commercial purposes of Customer or any third party and may watermark what they produce, and Siemens warns that academic work product carries limitations that make it unusable outside the educational use or public research area, which can spread to data it is combined with. Find these files (an intern's student seat, a startup's first months on a free plan) and take them out.
| Vendor | Who owns the files | Limit on what the files may build | The vendor's own AI training on them | Free, education and academic files |
|---|---|---|---|---|
| Autodesk (AutoCAD, Inventor, Fusion, Revit, Vault) | The customer | Not to replicate Autodesk's functionality | An administrator's setting per hub | No commercial use |
| Dassault Systèmes (SOLIDWORKS, CATIA, 3DEXPERIENCE) | The customer | No competing software, AI training named | Not stated | No commercial use; may be watermarked |
| PTC (Onshape, Creo, Windchill) | The customer | No API extraction for AI training from its hosted services | Not on customer data | Free-plan documents are public |
| Siemens (NX, Solid Edge) | The customer, output included | Not on the files | Not stated as AI training | Academic work product is restricted |
| Altium (Altium Designer and its cloud workspace) | The customer | No competing products from its output | A licence to improve its products | Not read |
7. Preparing a CAD lot
- Inventory by vault and product line. List each CAD and PDM system, the products and generations it holds, roughly how many assemblies, parts and drawings, and which simulations and test records go with them. Buyers price from a profile before any file moves: no buyer of company records states a sample size, and Mercor's first scoping questions are which CAD system holds roughly how many assemblies, and who owns the designs.
- Sort by owner. Separate the company's own designs from customer designs and from bought or downloaded content (section 4). Read the customer contracts and confidentiality agreements for anything you plan to keep in.
- Classify for export. Mark every product line EAR99, listed, or ITAR, and take out what cannot be released (section 5).
- Export natively, with history. Take the native files and the PDM revision history, not only STEP files or meshes, and the drawings beside them. Neutral formats can go too, as a convenience for the buyer, never instead of the native files.
- Join the outcomes. Link each design to its simulations, test reports, change orders and field failures, by part number and revision. This join is what the listings call test results for designs.
- Remove identities. Title blocks, file properties and PDM records carry the names of engineers, customers, suppliers and sites. Remove or replace customer and supplier names in everything that stays in, and decide how engineers' names are handled; our de-identification tool does the text work on your own computer.
- Describe it. One document a buyer reads first: systems, products and generations, counts, what is native, what history and test data come with it, how ownership and export classification were checked. Our datasheet builder writes it, and the rights check runs through the questions a buyer's lawyers will ask.
- Offer it with the company's records. CAD is worth most beside the tickets, specs, email and test records that explain it, and the buyers above take it that way. Our buyer match shows which buyers' published rules fit what you hold.
8. A closed company's design vault
A closed company's CAD can still be licensed, and closed companies are holders buyers ask for: micro1 runs a separate programme for companies that have shut down. The difficulty is practical. CAD and PDM subscriptions lapse when the company stops paying, cloud vaults are deleted on the vendor's schedule, and the engineers who knew which designs were whose have left. Export the vault while access lasts, keep a list of which customer each job was for, and settle who can sign for the company: an officer winding it up, or an assignee, liquidator or trustee once it is in an insolvency process. Our deletion countdown shows how long common systems keep data after a subscription ends, and our page on closed and insolvent companies covers who can sell.
9. How a sale through AIData.Broker works
AIData.Broker represents companies selling their data to AI buyers. We act only for the holder, as its agent in the sale, for holders based in the United States, the United Kingdom and Canada. For a holder of CAD and engineering records, a sale through us runs like this.
The valuation comes first, with no upload. The valuation asks about your systems, products, years and people and returns a range from the public records in section 3. If you add files, they are measured on your own computer. The valuation binds no one.
You keep the files and deliver them yourself. We do not hold your vault, show any part of it to a buyer, or license it in our name. You deliver directly to the buyer, after the ownership and export checks above, and the buyer pays you directly.
Every buyer that fits sees it at once. Every buyer whose published requirements fit your records receives a description that does not identify you; each that gives its confidentiality undertaking receives the same datasheet, with one deadline for offers. Offers come back on the same terms, so you can compare price, exclusivity, who may access the files and payment timing side by side. Our free offer check reads any offer you already have against the published terms of the buyers we track.
The fee is due only out of money you receive. It is a success fee on what you are actually paid, band by band: 12% of the first $250,000, 10% up to $1 million, 8% up to $3 million and 6% above $3 million. On a $300,000 sale the fee is $35,000. Mandates are not exclusive: our fee is due only on a sale to a buyer we introduced, during the mandate and for a tail of 12 months after it, and we do not take mandates where the expected sale is below $23,000. How it works sets out every step and term.
Questions people ask
Can I sell my company's CAD files to AI companies? Yes, the designs your company owns, if their release is not a controlled export. Buyers take CAD as part of a company's records: Mercor's hardware listing asks for native assemblies with feature trees with drawings and test results, and Handshake lists AutoCAD among the sources it buys.
How much are CAD files worth to AI companies? There is no price per file. Mercor's listing puts what a hardware engineering company earns at $45K to $1.3M, and Telegraph Lab reports paying an engineering firm $450K for records that included its CAD drawings. Native files with revision history and test results are worth the most.
Can a contract manufacturer sell its customers' drawings? Usually not without each customer's written consent. Their ownership and use are set by the customer contracts and confidentiality agreements, and buyers exclude them: Mercor's listing wants designs the company owns, not customer IP.
Does licensing CAD files to an AI company need an export licence? It can. CAD files are "technology" under the EAR, and releasing controlled technology to a foreign person, even in the United States, is a deemed export. Designs for ordinary commercial products are often EAR99, but defense drawings are ITAR technical data and stay out.
Do STEP files or meshes sell? Less well. Buyers ask for native CAD with feature trees, not only exported meshes, because a feature tree shows how a part was designed. Neutral files can go alongside the native ones as a convenience.
Is there free CAD data AI companies already use? Some, but little for commercial training: Autodesk's Fusion 360 Gallery dataset is for non-commercial research only, and the ABC dataset's models are public Onshape documents whose copyright stays with their creators. A company's own vault of real products is what is scarce.
What this means for you
- Sort your vault by owner before anything else. Your own designs can be licensed; your customers' cannot, without them.
- Classify for export. Know which product lines are EAR99, listed or ITAR, and keep what cannot be released out of any lot.
- Keep the native files and their history. Feature trees, revisions and test results are what buyers pay for; exported parts are not.
- Read your CAD vendor's terms on its own AI training and on libraries and educational licences (section 6).
- Get a range before any buyer sees anything. The valuation is free and needs no upload; our hub guide on selling company data to AI labs covers the whole sale, and what labs buy shows where engineering records sit among the kinds of data in demand.
AIData.Broker acts for holders in a sale, never for the buyer: if you want several buyers to compete for your engineering records, start with the valuation.