Skip to content

AI data market

An AI data investment committee market map that runs on two numbers

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

An AI data investment committee market map should place each revenue line, not each company, using two numbers taken from documents: the share of delivery cost that is new labor for every sale, and the share of revenue earned on rights the company controls. The four zones that result, licensing assets, prepared licensing, services, and tools or venues, each get their own comparables and metric. A line no document can place earns no premium.

Value drivers

AI data investment committee

This page covers an AI data investment committee market map that runs on two numbers.Illustrative

Key takeaways

  • Place revenue lines, not companies: one AI data firm can mix licensing, services and software under a single label.
  • Two documented numbers do the placing: labor per sale and the share of revenue earned on rights the company controls.
  • Four zones follow, licensing assets, prepared licensing, services, and tools or venues, each with its own metric and comparables.
  • Filings show zone signatures: ratable content licensing at Reddit, net marketplace revenue at LiveRamp, one customer near 58 percent at Innodata in 2025.
  • Price the zone the documents prove; a line no document can place earns no premium.

What decision does an AI data market map serve at investment committee?

It serves one choice: which comparables and which metric apply to each part of the company in front of the committee. An error there is hard to see later, because a services line priced like a licensing asset, or a brokered venue priced on gross bookings, carries the mistake into every number that follows.

AI data companies can run more than one kind of line at once. A firm can license an archive it owns, take on custom collection for one model developer and host evaluation software, then present all three under a single platform label. The map in this article is built for that situation. It ignores the label, measures each line with two numbers the data room can supply, and sends each line to the comparables that fit it.

This is a synthesis of a series on AI data business models. It borrows the series' tests as worked examples, adds its own map and decision rule, and cites company figures only from filings checked on 27 October 2026.

Two numbers that place any revenue line

Labor per sale and rights controlled. Together they describe what a valuation multiple silently assumes: how a line behaves when it grows, when it loses a buyer and when it is sold a second time.

Labor per sale is the share of a line's delivery cost spent on people or contractors whose work starts again for each new sale: collection, labeling, preparation to one buyer's specification, review. It comes from an hours ledger coded by sale, or from the cost of revenue note split by line. Labor spent once per asset version does not count, because the next sale reuses it.

Rights controlled is the share of a line's revenue earned by granting rights the company owns or is permitted to sublicense, rather than by delivering work the buyer will own or by passing through someone else's listing. It comes from a contract sample and the revenue recognition note. ASU 2016-10 helps read that note: a license to functional intellectual property is usually a right to use it as it exists when granted, recognized at a point in time, while a license to symbolic intellectual property is a right to access it, recognized over time. The standard also says a license's restrictions on time, geography or use define its scope rather than its timing, which is why metrics should be compared per unit of licensed scope.

Where to draw each line between low and high is an editable assumption; this article uses half as a default and asks the committee to record the value it chooses.

The investment committee map: four zones and the disguises that cross them

The two numbers produce four zones, and each zone has its own engine, its own repeat purchase and its own metric. The arrows in the figure are four disguises this series examines: a claimed position that the documents move elsewhere.

Licensing assets earn on rights with little new labor. Prepared licensing earns on rights but needs real work for every delivery. Services earn on labor the buyer will own. Tools and venues earn on software use or on transactions between other parties, with neither rights nor much labor of their own. A fifth status, unplaced, is for any line whose two numbers no document can support.

Each zone, the repeat purchase that drives it and the metric to compare it on:

The investment committee map: four zones and the disguises that cross them
ZoneWhat repeats when revenue repeatsScope-normalized metricWhere the number livesEvidence gap to close
Licensing asset: rights high, labor lowPermission to use an existing asset: renewal, a second licensee, a refreshLicenses and renewals per dataset version; contract value per unit of licensed scopeLicense schedule, revenue recognition note, remaining performance obligationsRenewal terms and how many licensees one version can legally carry
Prepared licensing: rights high, labor highRights repeat, but each delivery needs selection, preparation and reviewPreparation hours per accepted dataset; licenses per prepared versionHours ledger by dataset version, acceptance recordsWhether preparation is reused or rebuilt for each buyer
Services: rights low, labor highThe buyer orders more workGross margin per delivery hour; top-customer share of revenueCost of revenue note, customer concentration note, statements of workHours on the next sale versus the first
Tools and venues: rights low, labor lowCustomers use the software or transact with each other againNet revenue retention for tools; net take per matched transaction for venuesGross versus net presentation, usage and transaction disclosuresMatching between distinct buyers and sellers, not supplier counts
Unplaced: neither number documentedUnknownNone until the hours ledger and contract sample existManagement presentation onlyEverything above; no premium meanwhile
A two-by-two map with labor per sale across and rights controlled upward, zones for tools and venues, services, licensing assets and prepared licensing, and four arrows from a claimed position to where documents often place a line: platform to services, marketplace to services, data asset to prepared licensing and proprietary data to tools and venues.
The investment committee map. Illustrative: arrow positions show common disguises, not measured companies.

What public filings show in three of the zones

Filings show the zones' signatures in licensing assets, tools and venues, and services; no public filer we checked reports a prepared-licensing line on its own. Each example below shows one disclosure, not a ranking, and each leaves a gap the data room has to fill.

Licensing asset. Reddit's 10-K for 2025 says it recognizes content licensing revenue as partners consume and benefit from the licensed content, generally ratably over the license period, for a fixed or usage-based fee. Its quarterly report for June 2025 put remaining performance obligations on contracts longer than one year at $194.9 million as of 30 June 2025, consisting primarily of long-term content licensing contracts. That shows contracted, ratable rights revenue; it does not show renewal terms or what a second licensee of the same content would pay.

Tools and venues. LiveRamp's 10-K for fiscal 2026 says Data Marketplace revenue is recognized net of the share of revenue earned by the data seller, and that it comes mainly from revenue-sharing arrangements with sellers monetizing their data through its marketplace platform. A venue that reports net looks far smaller than one that books the same trades gross, so headline revenue cannot be compared across the two.

Services. Innodata's 10-K for 2025 says one customer in its digital data solutions segment generated approximately 58 percent of total revenues in 2025, against 48 percent in 2024, and ties the segment's growth to higher volume of its data engineering and AI systems services. Revenue that rises with one buyer's programs carries that buyer's program risk, whatever the line is called. The filing does not give hours per program or how much work carries from one project to the next.

The decision rule: price the zone the documents prove

Price each revenue line at the comparables of the zone its documents prove, and give the company's label no weight of its own. The rule takes five steps, each of which a deal team can finish before the committee meets.

Illustrative: a company presents itself as an AI data platform with three lines. Line one, 40 percent of revenue, licenses an archive it owns, with labor near a tenth of delivery cost. Line two, 45 percent, collects custom data for one model developer, who owns the output, with labor near 70 percent of delivery cost. Line three, 15 percent, hosts evaluation software on customers' own data. Under the rule, line one sits with licensing assets, line two with services and line three with tools, so nearly half the revenue is valued on gross margin per hour rather than on licensing multiples, and the platform label adds nothing. The percentages are invented to show the mechanics.

The five steps, in order:

  • Split revenue by the document that creates it: license, statement of work, subscription or usage order, marketplace terms.
  • Measure the two numbers for each line from the hours ledger, cost of revenue detail, a contract sample and the recognition note, and record the cut-off value used.
  • Place each line in a zone, or mark it unplaced when either number rests only on a management estimate.
  • Compare each line only with that zone's comparables and metric, normalized to licensed scope where rights are sold.
  • Revisit placement at every new filing or contract sample, because labels and contract mixes change.

Is the map just the label game with new names?

Only if the two numbers are taken on trust. The strongest objection to any market map is that words such as platform or marketplace hide services-heavy economics, and a new set of zone names could do the same. The defense is that placement depends on measured shares, not on the company's description, and that a line without documents cannot be placed at all.

The numbers still have soft spots. Cost allocation choices can move labor between delivery and overhead, recognition policy can make a rights line look steadier than its renewals, and a single large contract can flip a line across a cut-off. Ask for the ledger behind each share, test it on a sample of sales, and report the zone with the share beside it rather than the zone alone.

The map also says nothing about quality, rights validity or demand. A line can sit firmly in the licensing-asset zone and still hold an asset nobody needs. Those questions belong to the tests in the next section.

Reading this series as worked examples

Each disguise on the map has a dedicated test in this series, and the committee can run them as a sequence once a line is placed. The titles below are the series' working titles:

  • Platform that turns out to be services: Data or labor? How to spot a data startup services business model, and Data platform versus data product investment: who answers for the rows?
  • Marketplace that turns out to be aggregation: 448 suppliers, no market: data aggregation marketplace economics, and Data marketplace standardization investment: 6 standards, 0 require quality.
  • Licensing asset that turns out to need per-buyer work: Custom data or repeatable product strategy? Start with same-company margins, and Data inventory or on-demand sourcing? The fit-rate hurdle that decides.
  • Proprietary data that open substitutes erode: Open dataset commercial competition: where paid data keeps its edge, and Data infrastructure and data ownership: two investment bets, not one.
  • Rights claims that need proof: The rights-cleared enterprise data model sells what public licenses disclaim, and Direct data licensor business model vs data broker: a four-question test.
  • Quality, evaluation and documentation sold as products: Liability decides if an AI dataset quality assurance business stands alone, A dataset acceptance infrastructure business sells pass marks, not checks, and AI evaluation infrastructure economics, tested on 129 open evals.

Risks that could make the map mislead

Five developments could make a correct placement produce a wrong valuation. Each has a public example an investor can watch.

If the committee can name which of these would move a company's largest line across a zone boundary, it has stress-tested the map:

  • Labels drift between filings. Reddit's quarterly report for March 2024 called the line data licensing and said the opportunity had only recently generated revenue; its 10-K for 2025 calls it content licensing. Placement has to be redone from documents each time.
  • Recognition can flatter recurrence. Ratable recognition and large remaining performance obligations, as in Reddit's filings, show contracted revenue, not renewal; a licensing asset whose licenses end without renewal behaves like a sequence of projects.
  • Concentration overrides the zone. A services line with one customer near 58 percent of company revenue, as Innodata reported for 2025, carries that customer's program risk whatever its margin per hour.
  • Net and gross presentation change apparent scale. A venue reporting net of seller payouts, as LiveRamp does for Data Marketplace, cannot be compared on revenue with one reporting gross.
  • Exclusivity caps repeat licensing. Where a supplier grants exclusivity for a field of use, licenses per version depend on what the licensor may still sublicense, which only the contract shows.

Placing SourceX, and the questions for any company on the map

On its published terms, SourceX sits in the prepared-licensing zone. It runs the chain from sourcing and qualification through rights review, contracting, preparation, delivery and payment for businesses licensing selected operating records to AI developers, and it sources each dataset to a buyer's request instead of holding inventory. No records reach it until the supplying company has agreed the scope and terms in a signed agreement.

Its how-it-works page adds two terms that matter for the map. Licenses are exclusive for AI training in the sense that the supplier agrees not to license the selected data to anyone else for that purpose, while SourceX's own right to distribute or sublicense the de-identified dataset is unaffected; the term and any alternatives are set in each deal's Dataset Deal Confirmation. Suppliers are usually paid once per dataset, with some agreements setting recurring deliveries, and no royalties or usage fees. Licenses per prepared dataset and preparation hours per accepted dataset are therefore the numbers that decide its economics.

None of those numbers is public. SourceX shares company and commercial metrics with investors directly rather than publishing them, and an investor can ask for its materials on the investors page. Separately, the company's current raise is open only to verified accredited investors under Rule 506(c). The same questions apply to SourceX and to any company on the map:

  • Which documents produced each line's labor share and rights share, and what cut-off did you use?
  • How many hours did the second sale of a dataset version take compared with the first?
  • How many licensees does a typical version carry, and what do exclusivity terms leave you free to sublicense?
  • What share of licensing revenue is recognized at a point in time versus over time, and how much renewed at the last expiry?
  • Which line would cross a zone boundary if your largest customer left?
  • Which open dataset or synthetic source would most reduce what a buyer pays for your largest rights line?

Methods note: how the map was built and what it leaves out

The map, the two numbers, the default cut-off of one half and the five-step rule were drawn up for this article on 27 October 2026 as editorial tools; the worked example's percentages are illustrative. Filing statements come from Reddit's 10-K for 2025 and its quarterly reports for March 2024 and June 2025, LiveRamp's 10-K for fiscal 2026 and Innodata's 10-K for 2025, verified through search excerpts on sec.gov; the accounting reading comes from FASB's ASU 2016-10. SourceX's terms come from its how-it-works and investors pages, read the same day.

Not measured: the labor share or rights share of any named company's lines, the distribution of companies across zones, valuation multiples by zone, and any SourceX metric. No company is placed by measurement here; the filing examples illustrate one zone signature each, and the series titles cited may change before publication.

Frequently asked questions

What is an AI data investment committee market map?

It is a one-page way to decide which comparables and metrics apply to each revenue line of an AI data company. This version places lines by two documented numbers, labor per sale and rights controlled, instead of by the company's own category label.

Why place revenue lines rather than whole companies?

Because one AI data company can run several kinds of business at once, such as licensing an archive, doing custom collection and hosting software. Averaging them into one company-level label can price every line at the wrong comparables.

How does revenue recognition help place a licensing line?

The recognition note shows whether license revenue is earned at a point in time or over the license period, and whether a line is reported gross or net. ASU 2016-10 ties point-in-time recognition to a right to use intellectual property as it exists and over-time recognition to a right to access it.

What should a committee do with a line it cannot place?

Treat it as unplaced and give it no premium until the company produces an hours ledger and a contract sample. A management estimate of either share is a hypothesis, not evidence.

Where does a company like SourceX sit on the map?

On its published terms, in prepared licensing: it licenses rights to selected business records and performs preparation for each dataset. Its licenses per dataset and preparation hours per dataset are not public; the company says it shares company and commercial metrics with investors directly.

Sources

  • Reddit recognizes content licensing revenue as partners consume and benefit from licensed content, generally ratably over the license period, for fixed or usage-based fees (10-K for 2025). Source
  • Reddit's remaining performance obligations on contracts longer than one year were $194.9 million at 30 June 2025, primarily long-term content licensing contracts. Source
  • Reddit's quarterly report for March 2024 described data licensing as an opportunity that had only recently generated revenue. Source
  • LiveRamp recognizes Data Marketplace revenue net of the share earned by data sellers, mainly through revenue-sharing arrangements (10-K for fiscal 2026). Source
  • One customer in Innodata's digital data solutions segment generated approximately 58 percent of total revenues in 2025 and 48 percent in 2024; segment growth came from higher volume of data engineering and AI systems services. Source
  • ASU 2016-10 treats a license to functional intellectual property as a right to use, recognized at a point in time, and a license to symbolic intellectual property as a right to access, recognized over time; license attributes define scope, not timing. Source
  • SourceX licenses are exclusive for AI training: the supplier agrees not to license the selected data to anyone else for that purpose, SourceX's right to distribute or sublicense is unaffected, and suppliers are usually paid once per dataset with no royalties or usage fees. Source
  • SourceX shares company and commercial metrics with investors directly and takes investor materials requests on its investors page. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies