Skip to content
AI Data Payout

Guides

Why do AI companies want CAD files?

AI companies want CAD files because real design work barely exists in public training data, and working drawings show how things actually get built. Those files sit on firms' own servers, and Ventra says labs look for CAD libraries specifically, often one kind in volume, while Polyshares lists firms with heavy CAD work as in demand.1,2,3

By Lead analyst (to be named) · reviewed by Reviewing editor (to be named) · published 1 Nov 2026 · updated 1 Nov 2026

Real design files never made it onto the open web

Real design files sit on private servers, which is why design tools lag behind AI tools for code and writing. Fast Company reported in August 2026 that the drawings, 3D models and sketches that could train an architecture model are in firms' servers and filing cabinets, and that none of the big AI labs was taking on the problem itself when it wrote.1

Public research sets show how thin the supply is. The ABC dataset holds one million CAD models, each stored as curves and surfaces, which teaches a model about shape. The Fusion 360 Gallery authors said no dataset of realistic CAD models in a short programmatic form existed before theirs, and theirs has 8,625 design sequences limited to two operations, sketch and extrude.4,5

A working firm's archive holds what those sets leave out, in our view, because every file was drawn to meet a real budget, code or customer, and the revisions show what changed and why.6

Buyers ask for one kind of design file, in volume

Buyers ask for a lot of one kind of design. Ventra says labs often want one kind in volume, like a library of building designs, and told us one lab partner asked it for 10,000 of them, which is Ventra's own account. Its public catalog shows 20 of its 31 proprietary datasets as design, drawing or 3D scan data.2,7,8

Depth matters too. Polyshares says drawing sets record layers, blocks, dimensions and revisions tied to actual projects, and show how a design moved from early layouts to issued sheets. It also says the value is in reasoning captured while the work happened, not in volume, so a smaller archive with its history intact can still interest a buyer.6,9

What labs likely do with design files

Labs rarely say what they build from a given archive, so this part is our reading. The Fusion 360 Gallery authors tested a model on recovering the steps of a CAD program from a finished shape, which points to the goal of AI that can build a part step by step, not only picture one.5

The pull from users is plain in the AIA's research, where 84% of architects are optimistic that AI can automate manual tasks and only 6% use it regularly in their jobs. Some firms are already building their own tools from their portfolios, Fast Company reported, so demand is not limited to the big labs.10,1

Pros and cons

Pros

  • CAD is hard to find, and Ventra says labs look for it specifically, which puts a real archive in a better spot than records every company holds.2
  • Your firm's size matters less here. Ventra told us CAD datasets don't need its 30-person minimum and could even come from one person.11
  • A license is a copy, so your team keeps its files. Mode says the company keeps its originals, and Ventra says a licensing deal lets the company keep ownership.12,13

Cons

  • No buyer publishes a price for CAD files, and none of the 31 listings in Ventra's catalog shows one, so you can't value an archive from a website.14
  • Rights slow things down. In Ventra's catalog, 14 of the 20 design, drawing and scan listings still show rights as being confirmed with the owner.15
  • Some of an archive has to come out. Troveo's engineering guide lists designs clients own, licensed standards, export-controlled data and anything under a confidentiality agreement.16

What it means for a business like yours

If your business has years of drawings or models, count them before anyone calls: projects, files, GB, formats and the oldest year. Those are the rows on Ventra's datasheet, which it heads "What a lab prices on", and they let a buyer tell you quickly whether your archive fits.14

Should your business get in now?

Businesses are already licensing their records: micro1 has committed $1B over 12 months to buying company data, and at least 18 firms started advertising for it between September and mid-October 2026.17,18 Our view, not a promise: the money will not be around forever, because prices tend to come down as more businesses sell.17,19 Take the quiz to see which firms fit your business.

Plus a free Seller's Guide when you finish.

Take the quizWho the brokerages are

We feature companies that pay us a referral fee. Our scores and reviews don't change for it. How we earn.

FAQ

Questions, answered with sources

  • AI data firms are asking for CAD files by name. Ventra names CAD files in its ads, Polyshares lists firms with heavy CAD work as in demand, and Mode lists AutoCAD among its sources. These are the firms' own statements, and few closed CAD deals have been made public.20,3,21

  • Free CAD sets exist but they are limited. The ABC dataset has one million models, and the Fusion 360 Gallery has 8,625 design sequences built from two operations. In our view neither shows the budgets, reviews and revisions that sit behind a real job.4,5

  • Buyers most often describe one kind of file in volume. Ventra lists architectural plans, building designs, mechanical parts and assemblies, and its catalog groups them as building and architecture with 11 datasets, mechanical with 7, and 3D scans with 2.22,8

  • No buyer publishes a per-file or per-drawing price. Pewter's licensing partner reports a 70-person engineering firm at $450K for CAD drawings, calculations, reports and specifications, which is one firm's whole set of records, and Pewter says it did not advise on the deal.23,14

  • Buyers don't publish an answer to this, so ask before you sign. Ventra says confidential details are masked or removed before the agreed use, and Mode says its agreement sets the scope, but you want to know what a trained model may produce and who else gets your files.13,12

  • A small firm can have what buyers want, since 6 of the 31 listings in Ventra's catalog show a core team of 1 to 10 people. Those listings show what Ventra is gathering, not closed deals.24

  • Every business is different, so this one is best talked through with the buyer you're working with. Their team goes over it with you before anything is signed.

  • Every business is different, so this one is best talked through with the buyer you're working with. Their team goes over it with you before anything is signed.

Before you sign

Questions to talk through with your buyer

Every business is different, so these are best talked through with the buyer you're working with. Their team goes over them with you before anything is signed.

Questions for this kind of business

  • Which of our files belong to clients, and which can be licensed?
  • Do our Autodesk or Dassault terms say anything about licensing files for AI training?
  • Who owns this data: the company, an estate, or someone else?
  • Do customer, supplier or platform contracts, NDAs or API terms restrict this?

Questions for any buyer

  • Is the license exclusive, and for how long?
  • Can you sublicense the data, and to whom?
  • Which AI labs will receive it, and will you tell us?
  • How long do you keep it, and how is it deleted?
  • What is your liability cap, and who indemnifies whom?
  • Will the contract ban re-identification of people in the data?
  • Who removes personal data, by what method, and do we approve it first?

What changed since we first published

  • 1 Nov 2026 First published.

How this page was made: drafted with AI help from the sources below, each checked; a buyer's claims about itself are its own. Updated November 2026. How we rank.

Sources (24)

  1. 1.Fast Company (Nate Berg), published 24 Aug 2026. Accessed 19 Oct 2026.
  2. 2.Ventra company data page. Accessed 15 Oct 2026. Company's own claim.
  3. 3.Polyshares broker portal, market criteria update (2026-10-22), published 22 Oct 2026. Accessed 22 Oct 2026. Company's own claim.
  4. 4.Koch et al., ABC: A Big CAD Model Dataset For Geometric Deep Learning, published 30 Apr 2019. Accessed 1 Nov 2026.
  5. 5.Willis et al., Fusion 360 Gallery, published 17 May 2021. Accessed 1 Nov 2026.
  6. 6.Polyshares AutoCAD data guide. Accessed 1 Nov 2026. Company's own claim.
  7. 7.Ventra founder, call with AI Data Payout (2026-10-19), published 19 Oct 2026. Accessed 19 Oct 2026. Company's own claim.
  8. 8.Ventra dataset catalog. Accessed 1 Nov 2026. Company's own claim.
  9. 9.Polyshares data licensing page. Accessed 1 Nov 2026. Company's own claim.
  10. 10.The American Institute of Architects, published 11 Mar 2025. Accessed 19 Oct 2026.
  11. 11.Kyle at Ventra, email to AI Data Payout (2026-11-01), published 1 Nov 2026. Accessed 1 Nov 2026. Company's own claim.
  12. 12.Mode company data site. Accessed 14 Oct 2026. Company's own claim.
  13. 13.Ventra company data page. Accessed 15 Oct 2026. Company's own claim.
  14. 14.Ventra dataset catalog. Accessed 1 Nov 2026. Company's own claim.
  15. 15.Ventra dataset catalog. Accessed 1 Nov 2026. Company's own claim.
  16. 16.Troveo, engineering firm data licensing, published 12 Oct 2026. Accessed 14 Oct 2026. Company's own claim.
  17. 17.The Next Web, published 9 Oct 2026. Accessed 14 Oct 2026.
  18. 18.Meta Ad Library (our review). Accessed 14 Oct 2026.
  19. 19.TechCrunch, published 16 Sep 2025. Accessed 14 Oct 2026.
  20. 20.Meta Ad Library, Ventra page. Accessed 14 Oct 2026. Company's own claim.
  21. 21.Mode company data page. Accessed 15 Oct 2026. Company's own claim.
  22. 22.Ventra company data page. Accessed 1 Nov 2026. Company's own claim.
  23. 23.Pewter website. Accessed 1 Nov 2026. Company's own claim.
  24. 24.Ventra dataset catalog. Accessed 1 Nov 2026. Company's own claim.