Deal economics
A curated data subscription business model renews only what goes stale
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A curated data subscription business model recurs only when stopping payment costs the buyer something it still needs: updates whose absence makes its copy wrong, or the right to keep using what it holds. AI training data is the hard case, because a trained model keeps what it learned. Check the refresh obligation, the decay rate, what cancelling removes and how retention is measured before calling the revenue recurring.
Value drivers
Curated data subscription
Key takeaways
- A subscription is recurring only if cancelling takes something away; otherwise it is a one-time sale billed in installments.
- In OpenSanctions' repository at 27 October 2026, 230 of 454 active crawlers declare daily or weekly updates, and the cadence follows the cost of staleness: 63 of 78 sanctions sources refresh daily.
- OpenAlex releases its data under CC0 and charges for monthly snapshots, daily change files and API capacity, which isolates freshness as the product.
- Training buyers keep what models learned, and IFRS 15 treats a license without ongoing activities as a right to use, so AI data subscriptions must earn each renewal with new records.
- Ask for retention by value and by client count with services removed, plus refresh hours per source and per buyer.
When does a curated data subscription business model actually recur?
It recurs when stopping payment costs the buyer something it still needs: updates whose absence makes its copy wrong, or the right to keep using what it already holds. When neither applies, the subscription is a one-time sale paid in installments, and each renewal depends on the seller finding something new to sell.
Curated here means the seller selects, cleans and maintains the records instead of passing through a raw feed. The investor question is which part of that work repeats. A list of sanctioned parties can go out of date within days if nobody maintains it, so the buyer's need renews by itself. A training corpus does not go wrong the same way, because a model trained on it keeps what it learned after the license ends.
The test below checks four things in order, then applies them to two curated datasets whose refresh promises and license terms are published in full, OpenSanctions and OpenAlex, before turning to AI training data.
The cancellation test: four questions that separate renewals from installments
Ask what the buyer would lose on the day it stopped paying, and how soon it would notice. Each question needs a document behind the answer, not a sales description:
- Refresh obligation: what the seller must deliver each period, stated per source or per table, with the cadence written down. A promise to update regularly is not an obligation.
- Decay rate: how fast an unrefreshed copy becomes wrong for the buyer's use, measured as the share of delivered records that change, appear or disappear per period.
- Cancellation scope: whether cancelling ends only future updates, also ends the right to use delivered copies, or reaches models and outputs built from them.
- Retention basis: whether renewal is reported by contract value or by number of clients, with services removed, because the two can tell different stories.
What 454 OpenSanctions sources declare about refresh
OpenSanctions, a curated database of sanctions lists, politically exposed persons and watchlists, publishes a metadata file for every source it crawls. At commit 1a0089c (27 October 2026) its repository holds 547 of them; after setting aside collection and internal folders (88 files) and disabled sources (5), 454 active crawlers remain. Of those, 145 declare a daily update, 85 weekly and 154 monthly; 30 are marked never, which a comment in the repository says the website shows as not automated, and 40 declare no cadence.
The cadence tracks the cost of being stale. Of the 78 crawlers whose first list tag is list.sanction, 63 refresh daily, 80.8 percent; of the 212 whose first list tag is list.pep, for politically exposed persons, 140 refresh monthly, 66.0 percent. A screening customer that misses yesterday's designation risks a compliance failure, while one holding a month-old roster of officeholders has a slightly dated list. One subscription therefore carries several refresh obligations at once, and its renewal case is strongest in the segment that decays fastest.
The license sits on top of the refresh. The repository licenses its data under Creative Commons Attribution-NonCommercial 4.0 and states that commercial use requires a separate license, and the project's FAQ tells businesses that use the data to acquire one. Anyone can read the rows; what renews is the right to use them commercially plus the crawling that keeps them current.
OpenAlex gives the rows away and charges for the sync
OpenAlex, an index of scholarly works, shows the refresh premium with the rights question taken away. Its documentation says the data is released under CC0, in the public domain and free to use in any way. The same documentation says the free snapshot is updated quarterly and the API is capped at 100,000 calls a day, and that monthly snapshots, daily change files that keep a local copy in sync, and a higher API limit require OpenAlex Premium.
Everything paid for there is time and capacity. A buyer who stops paying keeps every record it downloaded, under a public-domain dedication, and loses only the faster path to new and corrected records. That is the purest form of a curated subscription: each renewal is a bet that the gap between quarterly and daily matters to the buyer's work.
Why AI training data strains the subscription logic
Training data is consumed once. A model trained on a delivery keeps what it learned whether or not the license renews, so the cost of cancelling is the value of the next delivery, not the loss of the last one. SourceX's own supply terms show the norm from the licensee's side: once it has paid for a dataset, its license to the de-identified data is permanent, survives termination, and models already trained on it are not retrained or deleted.
Accounting points the same way. Under IFRS 15, a license is a right to access, recognized over time, only when, among other conditions, the licensor is expected to carry out activities that significantly affect the licensed property and the customer is exposed to their effects; otherwise it is a right to use the property as it exists when granted, recognized at a point in time. A training snapshot with no duty to change it fits the second description even when the invoice arrives monthly.
An AI data subscription therefore has to earn each renewal the way OpenAlex does, by delivering records the buyer's next training or evaluation run needs: events after the last cutoff, new edge cases, replacement test items. The diligence question is whether the buyer's retraining schedule creates that need every period or only every few quarters.
Five curated data models compared on scope-normalized metrics
Normalize each structure to what one buyer receives per period, so a monthly invoice cannot pass for a refresh. The rows describe structures rather than rank companies, and the last column names the document that shows each feature:
| Model | What recurs each period | What cancelling removes | Metric to ask for | Where it is documented |
|---|---|---|---|---|
| One-time snapshot license | Nothing, unless a further delivery is agreed | Nothing already delivered | Share of relationships that sign a second delivery | SourceX supply terms: usually one payment, some agreements set recurring deliveries |
| Open data with paid freshness | Faster snapshots, change files and API capacity | Only the faster path to updates | Paying users against users of the free snapshot | OpenAlex documentation |
| Curated list with a commercial license | Crawled updates plus the right to commercial use | The commercial right and future updates | Renewal by value, split by each segment's refresh cadence | OpenSanctions repository and FAQ |
| Live-access data subscription | Access to a maintained database used in daily work | Access itself | Retention by contract value and by client count, services excluded | FactSet annual reports on Form 10-K |
| Training-data subscription in batches | New batches, if the contract obliges them | Future batches only; trained models stay | Records added or changed per batch, and whether the buyer retrains on them | Contract schedules; no public example reviewed here |
Is the subscription label covering a services business?
Sometimes. A refresh is product or labor depending on whom it serves. OpenSanctions maintains one crawler per source, and each run serves every licensee at once; a team that re-cuts, re-labels or re-checks a dataset separately for each buyer every quarter is selling services on a subscription invoice. The cost line looks identical in both cases until it is divided by customers.
Mature data subscription companies draw the line inside their own metrics. FactSet defines annual subscription value as forward-looking revenue from subscription services and leaves out professional service fees, which it says are not subscription-based, and its annual reports state client retention both as a share of that value and as a share of clients. Ask an AI data company for the same split: refresh hours per source and per buyer, and retention measured both ways with services removed.
Where SourceX sits, read from its published terms
SourceX is a labeled application of the test, described only from its published pages. It acquires licenses to business records directly from the companies that hold them and supplies licensed datasets to AI developers, handling the work from rights review and contracting through preparation, delivery and payment. Buyer datasets are sourced on request from U.S. businesses rather than kept in inventory, so it does not sell a catalog subscription.
On the supply side its terms sit in the first row of the table. SourceX usually pays a company once for a dataset, and some agreements set recurring deliveries, each paid as the agreement states. Exclusivity for AI training covers only the selected data, the term and any alternatives are fixed in each deal's Dataset Deal Confirmation, and future records are outside a deal unless it adds them. Recurring revenue, where it arises, would come from further deliveries and onward licenses rather than from a subscription to a fixed asset.
SourceX has not published how many suppliers deliver more than once, how often one dataset is licensed onward more than once, or any retention figure. It shares company and commercial metrics with investors directly, and investor materials are available on request through its investors page; those are the numbers the cancellation test needs.
What could break the curated subscription thesis
Five developments could weaken the case for curated data subscriptions, and each has a visible signal:
At a glance
What could break the curated subscription thesis
- Free rows: OpenAlex releases a curated index under CC0, so where a public alternative is close enough, a seller can charge only for speed and capacity.
- One-time consumption: training buyers keep what their models learned, so renewal rests on each period delivering records the next run needs, not on continued access.
- Slow decay: in the OpenSanctions count, 154 of 454 active sources refresh monthly and 70 are not automated or declare no cadence, so much of a curated product can change slowly and give buyers of those segments less reason to renew on schedule.
- Reclassification: under IFRS 15, a contract billed monthly with no duty to change the licensed data can still be a right to use, recognized when granted, which removes the recurring label from reported revenue.
- Labor creep: refreshes customized for each buyer convert margin into hours; FactSet's practice of excluding professional services from its subscription measure is the model for separating them.
Methods behind the count, and the questions to ask next
The OpenSanctions count reads metadata files only, at commit 1a0089c6 of github.com/opensanctions/opensanctions (27 October 2026); no crawler code was run. Population: every .yml file under datasets/. Dropped, in order: files in top-level folders whose names start with an underscore (88 files in _analysis, _collections, _externals, _funes, _global, _testing and _wikidata), files marked disabled (5) and files without an entry point (0). For the 454 that remain, coverage.frequency is tabulated as written, with a missing value counted as not declared; list type is the first tag beginning with list. Percentages are rounded once, and the counting script stays in the batch's scratch folder so a reviewer can rerun it.
The OpenAlex terms come from its documentation repository at commit 8cec9db (May 2026); FactSet's definitions come from its fiscal 2017 and 2019 annual reports, read through search excerpts of the filings on sec.gov; the IFRS 15 test comes from the standard's published text. Not measured: how often any source actually changed, renewal or churn at any company, prices, or the refresh labor behind a crawler. The comparison table describes structures; it does not rate companies.
Questions to put to any curated data seller, SourceX included:
Checklist · 5 items
- Which deliverables are contractually owed each period, per source or table, and at what cadence?
- What share of delivered records changed, appeared or disappeared in each of the last four periods?
- What does a buyer lose on cancellation: updates, the right to use delivered copies, or rights in trained models?
- Is retention reported by contract value and by client count, with services removed from both?
- How many refresh hours are spent per source, and how many per buyer?
0 of 5 ticked. Tick each item as you go.
Frequently asked questions
What is a curated dataset subscription?
It is a contract under which a seller maintains a selected, cleaned dataset and keeps delivering updates for as long as the buyer pays. It differs from a raw feed, where nobody curates the records, and from a one-time license, where nothing more is owed after delivery. Its economics depend on how quickly the buyer's copy goes out of date.
How is a data subscription different from paying for a dataset in installments?
A real subscription leaves the buyer worse off when it stops paying, because updates stop or the right to use the data ends. Installment billing of a snapshot leaves the buyer with everything it needs after the last payment. The cancellation test in this article checks which one a contract is.
Do AI training data subscriptions renew like other data subscriptions?
No public source reviewed for this article reports renewal rates for AI training data subscriptions. The structural difference is that a trained model keeps what it learned after cancellation, so renewal depends on each period delivering new records the next training or evaluation run needs.
Which retention metric should an investor ask a data subscription company for?
Ask for retention by contract value and by number of clients, with professional services removed from both, and split by how often each segment refreshes. FactSet's annual reports have stated retention both ways, which shows the split is practical for a data business.
Why do sanctions data refresh daily while politically exposed person data refresh monthly?
Because a missed sanctions designation can mean a compliance failure the next day, while a month-old roster of officeholders is only slightly dated. In the OpenSanctions repository at 27 October 2026, 63 of 78 sanctions crawlers declare daily updates and 140 of 212 politically exposed person crawlers declare monthly ones.
Sources
- The OpenSanctions source metadata files at commit 1a0089c: 454 active crawlers and the update frequency each declares, counted by the article's script. Source
- A repository comment says a frequency of never shows as not automated on the OpenSanctions website. Source
- OpenSanctions licenses its data under CC BY-NC 4.0, separately from its MIT-licensed code, and requires a separate license for commercial use. Source
- The OpenSanctions FAQ says businesses that use the data must acquire a data license. Source
- OpenAlex data is made available under CC0, in the public domain and free to use in any way. Source
- OpenAlex's free snapshot is updated quarterly and its API allows 100,000 calls a day; monthly snapshots, daily change files and higher limits require OpenAlex Premium. Source
- IFRS 15 treats a license as a right to access, recognized over time, when the licensor's activities significantly affect the intellectual property; otherwise as a right to use, recognized at a point in time. Source
- FactSet's fiscal 2019 annual report defines annual subscription value as forward-looking revenue from subscription services, excluding professional service fees. Source
- FactSet's fiscal 2017 annual report states client retention both as a percentage of annual subscription value and as a percentage of clients. Source
- SourceX's published deal terms: usually a one-time payment, some agreements with recurring deliveries, exclusivity limited to the selected data with the term set in each Dataset Deal Confirmation, and a permanent license once paid. Source
- SourceX's investors page says it shares company and commercial metrics with investors directly and offers a materials request form. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.