A data broker has $20M for factory CAD. The valuable part isn’t the CAD.

Short answer: On Oct 28, 2026, Michael Davidson, VP of Partnerships at Polyshares, a company that licenses data to AI developers, posted, "We have $20M to spend next month on CAD" (post). Two factory founders pushed back hard, warning about lawsuits and export controls. Our view: CAD is the theory; the factory is reality. The scarce asset is the record of how a design changed on its way to being made, and most of that record belongs to the people who make things, not to brokers. For job shops, selling customer CAD looks like a liability, not a revenue line.

What happened

  • Polyshares licenses operational data to AI developers. It describes deals as typically $100K to $2M under bounded licenses, and it doesn't name the labs (Polyshares). The $20M budget comes from one social post. No frontier lab has confirmed buying shop CAD, so the end buyers are unknown.
  • Aaron Slodov of Atomic Industries: "insane lawsuits waiting to happen. don't fork over customer CAD" (post).
  • Hadrian CEO Chris Power: "Everyone selling cad to the labs is going to jail" (post), and "Most of it is ITAR as well" (post). "Going to jail" is a prediction. As of Oct 31, 2026, we found no lawsuit, enforcement action or regulator statement about CAD sales. That's a negative search result, not proof that none exists.

What manufacturing data do AI labs actually need?

One buyer says it itself. Fallow Labs, another data buyer, prizes "rejected designs, revisions and the reasons behind them," plus design-for-manufacturing (DFM) notes (Fallow Labs). Slodov put it simply: "you can't skip the actual making part" (post).

Think of the CAD file as the answer key. The making record (what changed, what was rejected, why a part couldn't be made as drawn) is the worked solution. A model can learn the shape of a part from CAD. It can only learn how to make it from the factory.

There's a fair counterpoint. Synthetic CAD may close much of the geometry gap without buying anyone's files. CAD-Recode trains on a million procedurally generated sequences (arXiv), and Text-to-CadQuery fine-tunes general LLMs on CAD code (arXiv). That weakens the case for buying geometry. It doesn't touch the making record.

Is it legal to sell CAD files to AI labs?

This is a reading, not legal advice. Under the International Traffic in Arms Regulations (ITAR), CAD of parts on the US Munitions List (USML) is "technical data" (22 CFR 120.33). Releasing it to a foreign person, even inside the US, is a deemed export (22 CFR 120.50). Boeing's $51M settlement in 2024 shows that unauthorized access to drawings can itself be the violation (Reuters), though that was a different fact pattern. Our reading is that anonymizing a file doesn't fix ITAR, because the geometry itself is the controlled data.

What it means for you

  • Investors: treat any "we bought CAD" training-data story as a diligence item: whose files, what rights, what export screening. The flip side is a possible re-rating of factories that make their own parts and own their making data, such as Hadrian and Atomic (our view).
  • Founders in AI for manufacturing: "clean data" can be a position. Partnerships with factories, under written data-use terms, look more durable than data purchases.
  • Job shops, including LA's large ITAR supplier base: your customers' CAD usually isn't yours to sell. Your own quotes and DFM notes might be licensable, if your contracts allow it. Check with an export-control attorney first.
  • Engineers in ML: the open problem is learning manufacturing intent, not geometry.

What we're watching

  • Whether Polyshares and Fallow publish their export-control screening.
  • The first enforcement case or civil suit, which would likely chill brokers.
  • Whether primes start adding "no AI-training use" clauses to purchase orders (our speculation).

The question we're asking: how does a US factory turn its making data into AI training data without breaking ITAR or customer contracts?

Your turn: if you run a shop or license manufacturing data, reply and tell us how you're handling it. Follow LA Robotics Network for the first enforcement case, if it comes.

Previous
Previous

San Bernardino County Fire didn’t wait for the next crisis

Next
Next

LA starts the factories. Where does line two go?