AI data market
AI data company business models: eight models, three questions
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
AI data company business models differ in who owns the source rights, who performs the recurring labor and who carries the risk that a buyer rejects the data. Archive licensing, direct business-records licensing, expert data, annotation services, data software, marketplaces, evaluation and synthetic generation each earn money differently, so judge each on its own unit: contract value and renewal, margin after labor, retention or the share of transactions kept.
Value drivers
AI data company business
Key takeaways
- Sort AI data companies by who owns source rights, who does recurring labor and who carries acceptance risk, not by the label they use.
- Labor-heavy rows grow with contributor pay and headcount; rows built on owned rights or tooling can resell an asset, so different metrics apply.
- Contract value is not annual revenue: Reddit's $203.0 million in January 2024 licensing agreements ran over terms of two to three years.
- Buyer concentration appears across rows, from Reddit's two licensing partners to Innodata's single customer at about 58 percent of 2025 revenues.
- SourceX sits in the direct business-records licensing row and has not published operating metrics; it shares them with investors directly.
Why one label hides several AI data business models
Because the label describes what a company sells, not how it makes money. Two companies that both call themselves AI data platforms can sit at opposite ends of the economics: one licenses rights it already owns with little recurring work, the other sells the hours of contributors it recruits, manages and pays. Comparing them on one revenue multiple or one growth rate mixes businesses that scale, fail and get repriced for different reasons.
The drift is visible in companies' own descriptions. Snorkel AI built a data development platform, then launched Snorkel Evaluate and an Expert Data-as-a-Service offering alongside its Series D, and its FAQ now describes the company as a frontier AI data lab rather than a self-serve labeling tool. Nothing about that shift is wrong; it means the software label no longer tells an investor where the labor sits.
The testable thesis: an AI data company's economics are set more reliably by three facts, who owns the source rights, who performs the recurring labor and who carries acceptance risk, than by its product category. If that is right, companies sorted by those facts will show more similar margins, concentration patterns and failure modes than companies sorted by label. The atlas below makes the sorting reproducible; the evidence to test the thesis fully is not yet public.
Three questions that place any AI data company on the map
Ask who owns the source rights, who does the recurring labor and who absorbs the cost when a buyer rejects the data. The answers place a company in the atlas faster than any pitch-deck category, and each one changes which numbers matter.
At a glance
Three questions that place any AI data company on the map
A fourth question follows from the first three: what unit revenue should be measured in. Contract value suits licensing, contributor hours suit services, usage and retention suit software, and the share of transaction value kept suits marketplaces. Forcing one unit on all of them is how services-heavy economics end up valued as software.
- Source rights: who owns the material, or holds a license to sublicense it, before any buyer touches it. A company that owns its archive can license it again; a company working on a customer's data owns nothing it can resell.
- Recurring labor: who performs the work repeated for each new dataset or contract, such as collection, annotation, expert writing, cleaning or grading. Recurring labor makes revenue track headcount and contributor pay.
- Acceptance risk: who absorbs the cost when delivered data fails the buyer's test. A company paid only on acceptance carries rework and write-off risk that a company paid for access does not.
The AI Data Business Model Atlas
The atlas maps eight business models against the four roles that drive their economics. Cells describe the usual arrangement; any single contract can move a role, which is why diligence reads the contracts rather than the category.
Read it column by column. Where the company carries recurring labor, as in expert data, annotation, evaluation and per-dataset preparation in direct licensing, revenue grows with people and contributor pay, so margin after labor and utilization matter more than headline growth. Where the company owns rights or tooling, as in archives, software, marketplaces and synthetic generation, the same asset can be sold again, so renewal, retention and concentration matter more.
The classification is editorial, built for this article on 2026-10-26 from the company filings, announcements and documentation cited below. The methods note at the end of the page says what it does not measure.
| Model | Source rights | Recurring labor | Tooling | Acceptance risk usually sits with |
|---|---|---|---|---|
| Archive licensing | The company, which licenses an archive it already holds | Light: packaging, delivery and renewals | The buyer's own | The buyer, which licenses a known corpus |
| Direct business-records licensing | The supplying business owns them; the intermediary holds a license to sublicense | The intermediary, directly or through a processing partner | Shared between the intermediary and its partner | Shared, as set in each deal's agreement |
| Commissioned collection and expert data | Usually assigned to the buyer by contract | The company's recruited contributors and experts | Shared: company workflow, buyer specification | The company, against the buyer's specification |
| Annotation and data engineering services | The buyer, which supplies the raw data | The company's workforce | Shared: company platform, buyer guidelines | The company, against agreed quality targets |
| Data development software | The customer | The customer's own team, using the software | The company | The customer |
| Marketplaces and exchanges | Each listing provider | Providers, for their own listings | The company runs listings and entitlements | The subscriber, under each provider's terms |
| Evaluation services | Shared: the company owns its test sets, clients own their data | The company's experts, who refresh and grade | The company | The company, for the validity of its tests |
| Synthetic data generation | The company, for its outputs, subject to seed data and model terms | Light: mostly compute, plus verification | The company | The buyer, after its own testing |
Representative disclosed models, with dates and limits
Public disclosures illustrate the rows; they do not rank anyone. Each entry states what the company itself disclosed, where and when, and what the disclosure cannot show.
Three patterns stand out without any ranking. Contract value and recognized revenue are different numbers: Reddit's $203.0 million was contract value spread over terms of two to three years, not one year of revenue. Concentration appears in more than one row, from Reddit's two licensing partners to Innodata's single customer above half of revenues. And category labels move, as Snorkel's own description shows. Shutterstock's data, distribution and services line was about 19 percent of its $935.3 million total revenue in 2024, a reminder that archive licensing usually sits inside a larger business and has to be read by segment.
| Company | Row in the atlas | What it disclosed | What it does not show |
|---|---|---|---|
| Archive licensing | January 2024 data licensing agreements: $203.0 million of aggregate contract value over two-to-three-year terms (S-1, 2024); two partners behind substantially all licensing contract value (10-K for 2024) | Renewal terms, or what licensees pay per unit of data | |
| Shutterstock | Archive licensing | Data, distribution and services revenue of $175.3 million in 2024, up 28 percent, including access to metadata for machine learning and generative AI model training (10-K for 2024) | How much of that revenue came from AI training licenses |
| Innodata | Annotation and data engineering services | One customer generated about 58 percent of total revenues in 2025 and about 48 percent in 2024, with no other customer at 10 percent or more (10-K for 2025) | Margin by program, or that customer's renewal terms |
| Scale AI | Annotation and expert data services | A Meta investment announced in June 2025 valued Scale at over $29 billion, with Meta holding a minority stake (company announcement) | Revenue, margin or concentration; Scale is private |
| Snorkel AI | Software moving into expert data and evaluation | Launched Snorkel Evaluate and Expert Data-as-a-Service with its Series D; describes itself as a frontier AI data lab rather than a self-serve labeling tool (company pages) | The revenue split between software and services |
| AWS Data Exchange | Marketplaces and exchanges | Providers publish offers with prices, durations, a data subscription agreement and a refund policy, and AWS reviews products before listing (AWS documentation) | Volumes, the platform's share or the AI training share |
Which revenue, margin and retention measures fit which model
Use the unit each model's economics actually run on, and ask for the denominator with it. The table names the revenue measure, the margin question and the retention signal for each family, plus the figure that most often misleads.
These measures are not interchangeable, and a company that spans several rows should report them by segment. A blended margin across licensing and services hides which half carries the business, and a blended growth rate hides whether growth came from reusable assets or from added contributor hours.
| Model family | Revenue measure to ask for | Margin question | Retention or repeat signal | Figure that misleads |
|---|---|---|---|---|
| Archive and direct licensing | Contract value by term, and recognized revenue by period | Cost to prepare and deliver each dataset | Renewals, and repeat buyers per dataset | Total contract value presented as annual revenue |
| Expert data and annotation services | Revenue per contributor hour and per program | Margin after contributor pay, recruiting and rework | Program renewals and top-customer share | Gross bookings with contributor pay left in |
| Data development software | Recurring software revenue and usage | Gross margin excluding bundled services | Net revenue retention by customer cohort | Services revenue counted as software |
| Marketplaces and exchanges | Transaction value and the share the platform keeps | Curation and review cost per listing | Repeat subscribers, and listings with paying subscribers | Listings counted as supply |
| Evaluation services | Revenue per evaluation cycle and refresh | Expert grading cost per cycle | Clients renewing after a model release | One-off launch evaluations treated as recurring |
| Synthetic generation | Usage or compute-based revenue | Compute plus verification cost per unit | Share of the buyer's data mix over time | Volume generated rather than volume accepted |
Where buyer demand points, and what it does not prove
Model developers say they want material they cannot reach on the open web, which supports the licensing rows of the atlas. OpenAI's data partnerships page describes two tracks, an open-source archive and private datasets for training proprietary models, and says it is not looking for datasets that contain sensitive or personal information, or information owned by a third party. OpenAI's o1 system card says its partnerships give access to non-public data such as paywalled content, specialized archives and other domain-specific datasets.
The same developers also generate their own data, which caps what an outside supplier can charge for some uses. NVIDIA's Nemotron-4 340B technical report (2024) says more than 98 percent of the data used in its model alignment process was synthetically generated. Demand statements show interest in non-public material; they do not show volume, price or which business model captures the value.
Where SourceX sits on the map, and the execution questions it faces
SourceX sits in the direct business-records licensing row. Across that row it handles the full chain for businesses licensing operational records to AI developers, from sourcing and qualification through rights review, contracting, data preparation and delivery to payment, so that selected records become licensed datasets. Rights pass in a fixed order: the supplying company first approves the scope and terms and signs, and only then is data delivered; from that point SourceX's use, sublicensing and distribution of the de-identified dataset are bounded by what the agreement allows. Buyer datasets are not held in stock but sourced from U.S. businesses once a request arrives, and SourceX does not host multi-terabyte sets: big deliveries either stay put in the seller's storage or go out on encrypted drives.
SourceX has not published metrics on the execution questions its row raises; it gives company and commercial figures to investors directly. Its capital raise runs under Rule 506(c) and is limited to accredited investors, who can request materials on the investors page.
On the atlas, that places source rights with the supplying business under a license, recurring preparation work with SourceX or its processing partner, and acceptance terms in each deal's documents. The execution questions follow directly from the row, and they apply to any company in it:
- How many supplier qualifications does it take to produce one accepted dataset, and where do candidates drop out?
- Which preparation steps recur for every dataset, and which can be reused across datasets from the same kind of system?
- How do the deal documents handle acceptance and rework, and who absorbs the cost of a rejected delivery?
- How concentrated are buyers, and does the same dataset sell to more than one?
- How long does a dataset take from request to accepted delivery, with open cases reported rather than dropped?
What could blur the eight business models or move their value
Four developments could blur the categories or change which rows hold value. Each has a dated public marker an investor can track, and a company that cannot say which one would hurt it most has not tested its own model.
- Concentration could matter more than model choice. Reddit attributes substantially all licensing contract value to two partners in its 10-K for 2024, and Innodata reports one customer at about 58 percent of 2025 revenues. When a few buyers set prices for every row, the atlas explains less than buyer behavior does.
- Synthetic data could absorb part of the demand. NVIDIA's 2024 report of more than 98 percent synthetic alignment data shows developers can replace purchased data for some training stages, which hits labor-heavy rows first.
- Categories could converge. Snorkel's move from software toward expert data and evaluation shows companies migrating across rows, so a classification made today can be wrong within a year; reclassify on each new disclosure.
- Regulation could raise costs in every row while favoring documented supply. Providers of general-purpose AI models placed on the EU market from 2 August 2025 must publish a training-content summary using the European Commission's template, published on 24 July 2025, with models already on the market given until 2 August 2027. That rewards sources with clear provenance and adds documentation work to each license.
Questions to ask any AI data company, row by row
Ask for definitions with every number: the denominator, the period and whether open cases are counted. A company that can answer these per row, rather than in a blended total, understands its own model, and the same questions apply to SourceX.
Checklist · 6 items
- Which row of the atlas produces most of your revenue today, and how has that share moved over the last four quarters?
- For each dataset or program, who owns the source rights, and can the same asset be licensed again?
- What share of revenue requires recurring labor, and what is margin after contributor pay and rework?
- How is acceptance defined in your contracts, and what happened to the last delivery that failed it?
- What share of revenue or contract value comes from your largest buyer, and when do those contracts renew?
- Which synthetic or in-house alternative would a buyer use if your price rose, and how do you know?
0 of 6 ticked. Tick each item as you go.
Methods note: how the atlas was built and what it does not measure
The atlas, the three questions and the metric table are editorial frameworks built for this article on 2026-10-26. Company examples were chosen to illustrate each row from primary disclosures available on that date: SEC filings by Reddit, Shutterstock and Innodata; announcements and pages from Scale AI and Snorkel AI; AWS documentation; OpenAI's data partnerships page and o1 system card; NVIDIA's Nemotron-4 340B report; and the European Commission's template notice. Inclusion is not a ranking or an endorsement, and each company is described in its own words.
Not measured: margins, growth rates or retention for any company, the share of each company's revenue in each row, private-company financials, and SourceX's own operating metrics, none of which are public in comparable form. The acceptance-risk column describes usual arrangements, not the terms of any specific contract. The 19 percent Shutterstock share is this article's arithmetic from the two reported figures.
Frequently asked questions
What are the main AI data company business models?
This atlas uses eight: archive licensing, direct business-records licensing, commissioned collection and expert data, annotation and data engineering services, data development software, marketplaces and exchanges, evaluation services, and synthetic data generation. They differ in who owns the source rights, who does recurring labor and who carries acceptance risk.
How is a data licensing business different from a data labeling business?
A licensing business sells rights to material it owns or may sublicense, so the same asset can earn revenue more than once. A labeling business sells work on data the customer usually owns, so revenue tracks contributor hours and margin depends on labor cost and rework. The two should be judged on different metrics.
Why do some AI data platforms have services-heavy economics?
Because the label describes the interface, not the work behind it. If a company recruits contributors, prepares each dataset or carries the cost of rejected deliveries, its revenue follows labor, whatever its product is called. The atlas's recurring-labor and acceptance-risk columns make that visible.
Which metrics should investors ask an AI data company for?
Ask for the unit that fits each row: contract value by term and renewals for licensing, margin after contributor pay for services, net revenue retention for software, repeat subscribers and the platform's share of transactions for marketplaces. Ask for each by segment, with denominators and periods.
Does SourceX publish its operating metrics?
No. SourceX has not published operating or commercial metrics and shares them with investors directly. Investor materials can be requested through sourcex.si/investors, and the questions in the business-model test are a fair place to start; any offering is open only to verified accredited investors.
Sources
- Snorkel AI announced Snorkel Evaluate and Expert Data-as-a-Service, a white-glove service for AI evaluation and post-training datasets, alongside its Series D. Source
- Snorkel AI's FAQ describes the company as a frontier AI data lab rather than a self-serve labeling tool, delivering datasets, benchmarks and evaluations built with its experts. Source
- Reddit's amended S-1 disclosed data licensing agreements entered in January 2024 with an aggregate contract value of $203.0 million and terms ranging from two to three years. Source
- Reddit's 10-K for 2024 says that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
- Shutterstock's 10-K for 2024 reports Data, Distribution, and Services revenue of $175.3 million, up 28 percent, total revenue of $935.3 million, and data offerings that include access to metadata for machine learning and generative AI model training. Source
- Innodata's 10-K for 2025 says one customer generated approximately 58 percent and 48 percent of total revenues in 2025 and 2024, with no other customer at 10 percent or more. Source
- Scale AI announced in June 2025 a significant investment from Meta that values Scale at over $29 billion, with Meta holding a minority stake. Source
- AWS documentation says AWS Data Exchange product offers define prices and durations, a data subscription agreement and a refund policy, and that AWS reviews data products against its guidelines before listing. Source
- OpenAI's data partnerships page describes an open-source archive and private datasets for training proprietary models, and says OpenAI is not seeking datasets with sensitive or personal information or information that belongs to a third party. Source
- OpenAI's o1 system card says its data partnerships provide access to non-publicly available data such as paywalled content, specialized archives and other domain-specific datasets. Source
- NVIDIA's Nemotron-4 340B technical report says over 98 percent of the data used in its model alignment process was synthetically generated. Source
- The European Commission published the template for the public summary of training content for general-purpose AI models on 24 July 2025; the obligation applies from 2 August 2025, and models already on the market have until 2 August 2027. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.