Don't Do It Yourself

How to Sell Semiconductor and Materials Data to AI

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?

AI companies buy run records: each build’s recipe or process settings linked to its measured performance. Mercor wants 1,000 or more runs, builds, wafers or experiments, with dates, and it wants the failed builds and repeat runs kept, not cleaned out.

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’s semiconductor and materials listing pays $100,000 to $750,000, depending on run count and how well runs link to results. The $100,000 floor is one of the higher floors on Mercor’s 26 listings, so a qualifying archive starts at a solid number.
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?

Not the design kit itself. Process design kits come from the foundry under license and confidentiality terms, and Mercor lists data owned by a foundry partner and licensed third-party IP as not a fit. Your own measurements and notes may still be licensable once the foundry’s material is removed.

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?

They can. Releasing controlled technology to a foreign person in the US is a deemed export under 15 CFR 734.13. In October 2022, BIS added controls on technology for certain semiconductor manufacturing equipment and on US persons supporting advanced chip production in China.

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?

CHIPS Act award recipients face a technology clawback. Under 15 CFR 231.203, knowingly licensing technology that raises national security concerns to a foreign entity of concern can make Commerce recover the full award, once Commerce has flagged that technology to you.

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?

Usually not without the customer’s written agreement. Development contracts and foundry service agreements often give the customer the results or bind you to confidentiality, and Mercor lists customer-owned data as not a fit. Your internal process development is what you can license.

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?

Leave out foundry design kits, customer-owned results, licensed third-party IP, export-controlled technology you haven’t cleared, and process recipes you rely on as trade secrets. Older process generations give a buyer the recipe-to-result link while giving away less.
  • 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?

Keep every run linked to its recipe, settings and measured result, and keep the failures. Strip staff names, customer and program names, lot numbers that point to a customer, and file metadata. Replace them with consistent codes so the links between runs survive.
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 is the only buyer publishing a payout range for semiconductor run records. micro1 buys engineering process data and documentation from companies with 30 or more staff, which can fit a fab’s or lab’s procedures and reporting workflows.
  • 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?

A semiconductor or materials company licenses its data in eight steps: count linked runs, split your data from foundry and customer material, screen for export controls, check any CHIPS award terms, apply to a buyer, code the IDs, have counsel review the license, then check a sample.
  1. Count linked runs. Runs with recipe, settings and measured result together, failures included, over several years.
  2. Split yours from foundry and customer material. Read the foundry agreement and each customer and development contract.
  3. Screen for export controls. Have your export lead classify process and equipment technology first.
  4. Check CHIPS award terms. If you hold an award, limit who the buyer can relicense to.
  5. Apply to a buyer. Mercor asks which systems hold your records, how many runs you have and whether any were published.
  6. Code names and IDs. Staff, customers, lots and tools, consistently across every system.
  7. Have counsel review the license. Our AI data licensing agreement checklist lists the clauses to check, including relicensing and deletion.
  8. 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.

Don't Do It Yourself
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.