A semiconductor or materials company sells its data to AI by licensing run records that tie each recipe to a measured result, to a buyer like Mercor.
Mercor pays semiconductor and materials companies $100,000 to $750,000. It wants 1,000 or more runs from a company with 10 to 2,000 employees, anywhere.
The catch is ownership. Foundry design kits, customer designs and licensed IP often sit next to your own run data, and none of them are yours to license.
For the full process across every industry, start with our guide to how to sell your company’s data to AI companies.
Key takeaways
- Mercor’s semiconductor listing pays $100,000 to $750,000 and wants recipes linked to measured results.
- Failed builds count. Mercor wants every build kept, with lab notes on why it failed.
- It rules out data owned by a customer or foundry partner, licensed third-party IP, and anything already published.
- Releasing controlled technology to a foreign person in the US is a deemed export under the EAR.
- Companies with CHIPS Act awards have an extra rule on licensing technology to foreign entities of concern.
What semiconductor and materials data do AI companies buy?
Mercor’s semiconductor and materials listing, read November 13, 2026, splits it up like this:
| Data | Mercor’s view | Main risk |
|---|---|---|
| Run and build records with measurements | Core of the license | Customer or foundry ownership |
| Lab notebooks | Core of the license | Names of staff and customer programs |
| Process, yield and equipment data | Core of the license | Export-controlled process technology |
| Design and test files | Core, where available | Licensed IP and foundry design kits |
| Notes on why builds failed | Adds to what you earn | Little, if they stay internal |
| The same recipe run again on different days | Adds to what you earn | Little |
| Published papers and datasheets only | Not a fit | Rejected outright |
The link between recipe and result is the point. A model learns how process changes move yield or performance, and a result with no recipe teaches it nothing.
Mercor names the companies it wants:
- Fabs and foundries
- Device makers in solar, photonics, quantum and power electronics
- Fabless chip designers
- Process and test equipment makers
- Materials and battery R&D labs
It works with MES, SPC and yield systems, LIMS, electronic lab notebooks, Cadence, Synopsys and OSIsoft PI. Labs and startups that have wound down also qualify.
How much is semiconductor data worth to AI companies?
| Mercor listing | Company earns | Size bar |
|---|---|---|
| Manufacturing and industrial | $100,000 to $1.5 million | 20 to 2,000 employees |
| Hardware engineering | $45,000 to $1.3 million | 20 to 2,000 employees |
| Biotech and life sciences lab | $50,000 to $875,000 | 10 to 1,000 employees |
| Semiconductor and materials | $100,000 to $750,000 | 10 to 2,000 employees |
| Chemistry and materials science R&D records | $10,000 to $300,000 | Not stated |
Run count and linkage drive the number. Mercor says the same recipe built again on different days, and measurements under more than one condition, add to what you earn.
Mercor also shows a referral fee of up to $110,000 for semiconductor companies. That’s paid to someone who refers a company, not to the company.
Our guide to how much company data is worth to AI compares published payouts across industries.
Can you license data that involves a foundry’s design kit?
Fabless designers feel this most. Their layouts sit on top of a foundry’s design rules, device models and standard cells, which the foundry licensed to them under its own terms.
Sort the archive into three piles before you talk to a buyer:
- Your own records: run logs, measurements, lab notebooks and test results you generated on your own programs.
- Foundry material: design kits, design rules, device models and anything the foundry supplied.
- Licensed IP: third-party cores, libraries and EDA content under someone else’s license.
Only the first pile is a clear fit. Read your foundry agreement before including anything that came from a shared process run.
Do export controls apply to semiconductor data?
15 CFR 734.13 treats any release in the US of technology or source code to a foreign person as an export. It counts as an export to their most recent country of citizenship or residence.
The October 13, 2022 BIS rule added license requirements for technology tied to certain semiconductor manufacturing equipment destined to China.
It also told US persons that some support for producing advanced chips in China needs a license.
An AI buyer’s annotation team is usually global, so the people who read your files are the export question.
Have your export lead classify process and equipment data before it goes in scope.
Our guide to selling hardware engineering data to AI covers deemed exports and ITAR in more detail.
What if your company received CHIPS Act funding?
The rule comes from Commerce’s September 25, 2023 final rule, effective November 24, 2023. It covers joint research and technology licensing with a foreign entity of concern.
A US data buyer isn’t a foreign entity of concern. The risk is downstream: who the buyer can relicense your data to.
If you hold a CHIPS award, ask the buyer to name the labs it sells to, and put a ban on relicensing to foreign entities of concern in the contract.
Can you license data from customer programs?
Read the IP and confidentiality clauses in each customer, development and joint-research agreement. Government-funded research can carry its own data rights terms too.
Published work is out for a different reason. Mercor wants data that hasn’t been published, so papers and datasheets earn nothing.
What should you leave out of a semiconductor data deal?
- Foundry material: design kits, device models and design rules
- Customer-owned results without written consent
- Licensed IP: third-party cores, libraries and EDA content
- Controlled technology until your export lead clears it
- Current recipes you protect as trade secrets
- Inventions you plan to patent but haven’t filed
The trade secret point is commercial, not only legal. A licensed process history teaches a model how you tune yield, and that model may be sold to your competitors.
How do you prepare run records and lab notebooks?
| Item | What to do |
|---|---|
| Lab notebook entries | Remove staff names and customer programs, keep the reasoning |
| Lot, wafer and run IDs | Replace with consistent codes across every system |
| Equipment and tool IDs | Code them if a serial number points to a site or customer |
| Recipes and process settings | Keep, unless they’re trade secrets or export-controlled |
| Yield and metrology data | Keep the measurements, remove customer part numbers |
| File metadata | Clear authors, edit history and comments |
Consistent codes matter more here than in most industries. A model learns from the same recipe run again on different days, which only works if the runs still match up.
Mercor says it signs an NDA, runs the export and removes or protects personal information. An engineer on your side should still check a sample before delivery.
Our guide to anonymizing data for AI training covers metadata tools in more detail.
Who buys semiconductor and materials data?
- Mercor: run records, lab notebooks, process, yield and equipment data, for companies with 10 to 2,000 employees. Pays $100,000 to $750,000.
- micro1: engineering workflows, technical documentation and operational data, listed on its data partnerships page. It wants 30 or more full-time staff and publishes tiers from $100,000 to over $3 million. You can apply to micro1 here.
Our list of companies that buy business data for AI has each buyer’s terms.
What are the steps for a semiconductor company, start to finish?
- Count linked runs. Runs with recipe, settings and measured result together, failures included, over several years.
- Split yours from foundry and customer material. Read the foundry agreement and each customer and development contract.
- Screen for export controls. Have your export lead classify process and equipment technology first.
- Check CHIPS award terms. If you hold an award, limit who the buyer can relicense to.
- Apply to a buyer. Mercor asks which systems hold your records, how many runs you have and whether any were published.
- Code names and IDs. Staff, customers, lots and tools, consistently across every system.
- Have counsel review the license. Our AI data licensing agreement checklist lists the clauses to check, including relicensing and deletion.
- Check a sample. Have an engineer try to name the customer or program from cleaned records.
Our guide to selling manufacturing data to AI covers shop-floor records if you also run production.
Common mistakes when selling semiconductor data to AI
The most common mistake is including foundry or customer material because it sits in the same system as your own runs. These are the others:
- Deleting failed builds, which Mercor values
- Sending results without the recipes that produced them
- Breaking the links between runs when you code lot and wafer IDs
- Ignoring deemed exports because nothing leaves the country
- Forgetting CHIPS award terms when the buyer can relicense
- Licensing current recipes you’d rather keep as trade secrets
- Reading Mercor’s $110,000 referral fee as what the company gets paid
Frequently asked questions
Can a semiconductor company sell its data to AI?
Yes, for run and build records it owns. Mercor’s listing rules out data owned by a customer or foundry partner, and licensed third-party IP, so foundry PDKs and customer designs usually stay out.
How much do AI companies pay for semiconductor data?
Mercor’s semiconductor and materials listing pays $100,000 to $750,000, depending on run count and how well each run links to its measured result. The referral fee of up to $110,000 goes to whoever refers the company.
What semiconductor data do AI companies want?
Run and build records that tie each recipe or process setting to a measured result, with failed builds and repeat runs kept. Mercor wants 1,000 or more runs, builds, wafers or experiments.
Do failed experiments have value to AI buyers?
Yes. Mercor asks for every build, failed ones included, and says lab notebooks explaining why a build failed are especially useful.
Can a wound-down lab or startup sell its data?
Yes. Mercor’s listing says labs and startups that have wound down qualify, as long as they own the records and haven’t published them.
Do export controls apply to semiconductor data?
They can. Releasing controlled technology to a foreign person in the US is a deemed export under 15 CFR 734.13, and BIS added controls on semiconductor manufacturing technology in October 2022.
Last updated: November 13, 2026
Disclosure: some links on this page are affiliate links. If you sign up through one, we earn a commission at no extra cost to you.
This isn’t legal advice. Have counsel review any data deal, and your export control lead review any process or design data, before you sign.






