Own the data AI still needs.

A starting plan and two paths · November 2026

  1. StartRefer, then brokerReferral fees first. Then Traine sells data to the buyers directly.
  2. Path AIndustrial Simulation StrategyCreate new data on purpose, and own every run.
  3. Path BSector Aggregation StrategyBecome one industry's shared data source.

Start with revenue. Then choose the path.

The SaaS era built convenience and connection for people. This era builds them for machines. We aren't scaling to reach an audience. We're scaling to augment intelligence.

The market right now

AI labs have moved from scraping the web to paying for records of real work.

Money is flowing to data that shows how work gets done and checked. Buyers use it to build realistic practice environments for AI agents.1

$1Bcommitted by one buyer, micro1, to buy and license company operating data over 12 months1
18×growth in one data vendor's run rate in 12 months, to $375M, on demand for high-end training data2
$10MGoogle's winning bid for one bankrupt airline's internal messages and records, pending court approval3
$1.5Mtop of Mercor's listed range for a single manufacturer's records, which starts at $100K3

The same trackers show which data commands the highest listed ranges: linked operating records from physical and engineering work, where each step is tied to a measured result.3

Chasing the moment loses

The mainstream data rush belongs to the largest players. The opening is in small pockets they can't efficiently reach.

What everyone is buying today is already being priced like a commodity.

  • Company data is priced by formula. One buyer's public estimator runs on $140,000 base plus $2,333 per employee, and a tracker reports a second major buyer uses the same math.4
  • Big buyers need big pools. Their programs ask for 20 to 30 or more employees per seller,5 and they are deploying budgets in the billions.1 Small, specialized holders fall below their line of sight.
  • The crowded categories are saturated. A lab-side buyer calls coding and standard computer-use data saturated, and warns that suppliers without a specialty “are going to be commoditized.”6
  • What labs still lack is hard to fake. The same source points to long tasks that are hard to verify as the unmet need.6
Our screen: three traits of data worth owning

1. A verified outcome is attached. A scan with what was found. A run with how it ended.
2. It is costly or impossible to recreate.
3. It is about things, not people. Cleaner rights, fewer privacy problems.

Data that passes all three is scattered across small firms. Each archive alone is a scrap. Pooled into one deep catalogue of a single type, it becomes something a buyer cannot assemble on its own.

Start here

Refer, then broker

Two steps that bring in revenue and buyer relationships before any large commitment.

The first step costs nothing to begin and can start this week. The second turns the relationships it opens into a brokerage.

Phase 0

PMG refers

PMG is an established company. It registers as a referral partner and introduces companies to the data buyers.

$50Kmicro1's referral fee on a company deal up to $700K, with 7.5% of deal value above that7
$100Ktop of Mercor's referral range per company, paid in full for companies with 30 or more employees8
$65Ktop referral fee on Mercor's open listing for manufacturers, including a $10K bonus9
$100K+what micro1 lists for the company itself, rising past $3M for approved data packages7
  1. PMG registers with micro1 and Mercor as a referral partner.
  2. A company owner applies through PMG's referral link.
  3. The buyer qualifies the company and agrees a deal with it directly.
  4. The company delivers its data.
  5. The buyer pays the company, and pays PMG a separate referral fee.
  • The fee comes from the buyer. Nothing is taken out of what the company earns.
  • There is no broker in the middle. Going direct keeps a broker's commission out of every deal.
  • Payment follows delivery. At micro1 the fee enters payout approval only after the company's data is transferred and approved,7 so the first payment is likely months out.
  • The buyers set the terms. micro1's fee was up to $25K when its program launched and is $50K now.8 It can change again.
The first sale is shared

The first company referred is the partner shop itself. The shop is paid for the records it already holds, and PMG is paid the referral fee on the same deal. Both partners earn from the first data sale, before Traine has spent anything.

PMG holds the referral relationship. Traine is the partnership for everything that follows. Every company PMG introduces is told that PMG earns a fee.

Phase 1

Traine brokers

Traine is the partnership. It begins now, selling data to the same buyers directly.

Traine's first inventory is data that already exists. It leases overlooked archives from small holders, pools them by type, and licenses them. The referrals open the buyer relationships that Traine then sells into.

The proof case is aerial data. An Australian mapping firm licensed survey data it had already captured to two AI companies for a combined $1.1M, as off-the-shelf licences.10 One tracker lists it as the only disclosed seller in the category.3

  1. Ask the buyers what they cannot find enough of. There is “no price list, so every price is negotiated.”4
  2. Collect leases in one data type. Depth in one type beats scattered archives.
  3. Build the catalogue. One described entry per dataset. Files stay with the holder.
  4. License directly to buyers, renewing only what sells.

Selling as a supplier is negotiated. The buyers publish referral programs, not a program for buying pooled archives. The first referral deals give Traine a named contact and a record of delivered value to open that conversation.

Where the gaps are

Six pockets that pass the screen. In each, researchers or the market have documented that the data is scarce.

Data pocketWhat the research showsWho holds it
Aerial, LiDAR and 3D survey archives$1.1M in licences for existing data, no new flying.10 Costly to re-fly.3Regional surveyors, aerial mapping and drone firms
Ultrasonic inspection scans“Training data for ultrasonic inspection is scarce.”11 Defects are “rare by design.”12Weld and materials testing firms
Machine failure recordsThe shortage of real run-to-failure data is called “the single greatest threat” to predictive-maintenance AI.13 Vibration analysis has no large labeled dataset.14Reliability contractors, motor shops, oil-analysis labs
Ground-penetrating radarResearchers cite a “lack of available datasets” for training detection models.15Utility-locating and concrete-scanning firms
Weld X-raysAvailability is “limited by non-disclosure contractual terms.”16Industrial testing labs
Thermal inspection imageryAlready sold commercially, collected by certified thermographers.17Thermography contractors

What this evidence does and doesn't show. Outside aerial, the proof is documented scarcity, not disclosed sales. That is why Traine talks to buyers before it signs any lease.

How we capture it

A short-term lease. The owner is paid regardless. We carry the risk and keep the upside.

  • AI use is named up front. The lease states the data may be licensed for AI development.
  • The selling window is limited. Our right to license runs monthly or for 60 to 90 days.
  • Buyer use is permanent. Once data is used in training, it “cannot practically be pulled back out,”18 and the lease says so plainly.
  • The owner warrants ownership. No client-owned, personal, or export-controlled material.
  • Licences are non-exclusive by default. The same catalogue can be licensed more than once. Founders report exclusive deals cost buyers roughly four to five times more,4 so exclusivity is priced when asked for.

What the start pays for

Three things the next step needs.

Revenue

Referral fees to PMG, and the first licence income to Traine.

Relationships

Named contacts on the buyers' partnership teams.

Knowledge

What the buyers cannot find, in what format, and what they pay.

Two paths from here

Two separate business models. The starting step funds and informs whichever is chosen first.

Both paths create data that does not exist yet. They differ in how it is made, who takes part, and how long it takes to pay.

These are separate businesses. Each has its own page and can be judged without the other. Choosing one first does not rule out the other later.

Path A

Industrial Simulation Strategy

Produce new training data on purpose. Real production work is run under full recording, every run is owned outright, and the same data can be licensed more than once.

Open Path A
Path B

Sector Aggregation Strategy

Become the shared data source for one underserved industry. Operating data is collected across many participants, pooled, and licensed as access for recurring income.

Open Path B

Risks and next steps

For the starting step only. Each path lists its own.

What could go wrong

  • Referral payouts are approved deal by deal. A company can sign and still stall before it delivers.
  • Open programs are crowded. Anyone can register, so the edge is knowing owners personally.
  • Demand is unproven for the industrial data types. Scarcity is documented. Purchases mostly are not.
  • Client ownership. Much of what holders have was captured for their customers and cannot be leased.
  • Traine becomes the licensor. If rights turn out to be bad, that is Traine's problem.

Next steps

  1. Register PMG with micro1 and Mercor, and read both sets of referral terms.
  2. Refer the first company to whichever buyer fits its size and industry.
  3. Form the Traine partnership and agree each partner's role.
  4. Have counsel draft the standard lease.
  5. Ask the buyers' partnership contacts what they cannot find enough of.
  6. Choose the first path.

Sources

Public sources retrieved November 16 to 17, 2026.

  1. Let's Data Science: micro1 commits $1B to buy and license enterprise data
  2. AI Training Jobs: What AI labs are buying (citing TechCrunch on Snorkel)
  3. Origin Data Partners: AI training data buyers by data type
  4. AIData.Broker: How AI training data prices are negotiated
  5. DDIY: Companies that buy business data for AI, compared
  6. Interview with Sean Cai on selling data and environments to AI labs
  7. micro1: Refer a company (referral program terms)
  8. Regulr: micro1 vs Mercor vs Handshake, data program pay and terms
  9. Mercor Data: open company listings
  10. Grafa: Aerometrex secures $1.1M in AI data contracts
  11. Journal of Nondestructive Evaluation: Deep learning flaw detection in ultrasonic data
  12. University of Florida: Why AI struggles with nondestructive testing
  13. arXiv: Neuro-symbolic AI for predictive maintenance, review and recommendations
  14. PHM Society: Towards a universal vibration analysis dataset
  15. Data article: A ground-penetrating radar dataset for subsurface utilities and voids
  16. Preprints.org: Synthesizing training data for digital X-ray radiography
  17. Claru: Thermal industrial dataset
  18. Frankfurt Kurnit: Selling company data for AI training

For discussion only. This presentation describes concepts under evaluation. It is not an offer to sell or a solicitation to buy any security or investment, and nothing in it promises a return. Market figures come from the public sources listed, several of which are published by companies active in this market. Referral fees and terms are set by the buyers and can change. Legal structures require review by counsel.