Deal economics
The rights-cleared enterprise data model sells what public licenses disclaim
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
The rights-cleared enterprise data model earns money on what public data licenses refuse to give: a named party standing behind authority, personal-data handling and acceptance. In a SourceX count of 31 data licenses on the SPDX License List, 29 disclaim warranties and none makes the licensor indemnify users. Map qualification, licensing, preparation and acceptance to who pays, who carries the risk and which asset survives the sale.
License terms
Rights cleared enterprise
Key takeaways
- Public data licenses grant use and little else: 29 of 31 on the SPDX list disclaim warranties, two are silent, and none has the licensor indemnify users.
- CDLA-Permissive 1.0 promised reasonable care over where data came from; version 2.0 dropped that promise and disclaims title and non-infringement.
- Fourteen of the 31 carve privacy, publicity or personal data out of the grant, the layer enterprise records most need cleared.
- Qualification and preparation behave like labor; licensing and acceptance decide who holds a reusable, sellable rights position.
- Compare companies per accepted dataset: who stands behind the rights, what survives the sale and how much labor the next dataset needs.
What does a rights-cleared enterprise data business actually sell?
It sells assurance about the rows more than the rows themselves. A buyer of enterprise records for AI training wants to know that someone with authority granted the use, that personal data was dealt with before delivery, that the same records will not be licensed to a rival for the same purpose, and that the data passed an agreed test before anyone was paid. The rights-cleared enterprise data model is the business of producing those four things and charging for them.
That framing gives an investor a sharper question than whether a company calls itself a platform or a marketplace. At each stage of a deal, qualification, licensing, preparation and acceptance, ask who pays for the work, who carries the loss if it fails and whether anything durable remains once the dataset ships. Value accrues where a durable, sellable rights position is created; cost accrues wherever the work must be repeated for the next dataset.
Rights for machine learning are divisible, which is why one cleared position can be worth more than one sale. The Montreal Data License proposal of 2019 set out a taxonomy for licensing data in AI that separates, among other rights, using a dataset to evaluate a model from other uses of it, so a single body of records can carry several distinct grants with different prices.
What 31 public data licenses promise about rights
Almost nothing beyond permission, and that gap is what a paid model fills. SourceX counted every non-deprecated license on the SPDX License List, version 3.29.0 (repository spdx/license-list-data, commit 31ba1a5 of 16 September 2026), whose name contains Data, Database or Open Government, or whose text mentions database rights or sui generis rights. Thirty-one licenses qualified, from the Creative Commons 4.0 family and the Open Data Commons licenses to the Community Data License Agreements and government licences from the UK, Canada, Norway, Germany, France and Taiwan.
Twenty-nine of the 31 disclaim warranties or liability, and the two Data licence Germany texts are silent on the point. CDLA-Permissive 2.0 is typical: its section 4.1 supplies the data "WITHOUT REPRESENTATIONS, WARRANTIES OR CONDITIONS OF ANY KIND", naming title and non-infringement among the warranties excluded. None of the 31 has the licensor indemnify the user. The only text that uses the word, the Solderpad Hardware License 0.51, requires a redistributor who sells a warranty to indemnify the original contributors.
Three texts promise something about rights. In CDLA-Permissive 1.0 and CDLA-Sharing 1.0, section 4.1 says each data provider "has exercised reasonable care" that the data was created by it or obtained with the right to publish it, and that publishing it breaks no privacy or confidentiality obligation the provider took on. Version 2.0 of the permissive license dropped that section, so one of the few licenses that names machine learning models moved away from assurance, not toward it. The third, France's Etalab Open License 2.0, promises the reuser that intellectual property rights held by third parties or by the licensor do not stand in the way of the rights it grants, a promise limited to intellectual property.
Personal data is the second gap. Fourteen of the 31 carve privacy, publicity or personal data out of the grant: the six Creative Commons 4.0 licenses say publicity, privacy and similar personality rights are not licensed, and the UK, Canadian and Norwegian government licences exclude personal data or personal information from what they cover. ODbL 1.0 and ODC-By 1.0 also leave privacy rights in contents unlicensed: their section 2.4 says the license does not cover rights in a database's individual contents other than database rights, and names data protection, privacy and personality rights among those the contents may carry. Only five of the 31 mention computational use, machine learning or artificial intelligence at all.
The count, test by test:
| Clause test on the license text | Licenses | Examples | What a buyer of enterprise records still needs |
|---|---|---|---|
| Disclaims warranties or liability | 29 of 31 | CC BY 4.0, ODbL 1.0, CDLA-Permissive 2.0, OGL-UK 3.0 | A party that stands behind authority and title |
| Excludes privacy, publicity or personal data from the grant | 14 of 31 | Creative Commons 4.0 family, OGL-UK 3.0, OGL-Canada 2.0 | Personal data handled before delivery, on a stated basis |
| Mentions computational use, machine learning or AI | 5 of 31 | CDLA family, C-UDA 1.0, O-UDA 1.0 | A grant that names training and evaluation |
| Provider represents something about rights | 3 of 31 | CDLA-Permissive 1.0 and CDLA-Sharing 1.0 (reasonable care only), Etalab 2.0 | A representation backed by a contract and a remedy |
| Licensor indemnifies the user | 0 of 31 | None; the one indemnity clause runs from a redistributor | Someone who pays if the rights fail |
Four deal stages, and the one that leaves an asset behind
Qualification and preparation are labor repeated for every new dataset; licensing and acceptance decide whether a sellable asset exists afterward. That split, more than any label, sets the economics of the rights-cleared enterprise data model.
Qualification finds a company whose records fit a buyer's need and whose contracts allow licensing; whoever runs the search pays for it, and most of the effort is lost when a candidate fails. Licensing produces the grants: if the intermediary takes a license it may pass on, plus a promise that the supplier will not license the same records to others for the same purpose, the rights position can outlive the first sale. Preparation, meaning de-identification, documentation and formatting, is skilled work whose cost scales with records and systems. Acceptance turns delivered records into earned revenue; until it happens nothing is owed, and whoever bears rejection risk carries the gap.
For each stage, what it produces, who usually carries it and what survives:
| Stage | What it produces | Who usually carries cost or risk | What survives the sale | Metric per accepted dataset |
|---|---|---|---|---|
| Qualification | A supplier whose records fit and may be licensed | The intermediary; failed candidates are sunk cost | A supplier relationship, if the contract can be reused | Candidates screened |
| Licensing | The grant from the supplier and the onward grant to the buyer | Supplier warrants authority; intermediary carries onward terms | A sublicensable, exclusive position, where the contract allows | Licenses sold per acquired scope |
| Preparation | De-identified, documented records in an agreed format | The intermediary or its processing partner | Little: tooling improves, the labor repeats | Labor hours, first dataset versus repeat |
| Acceptance | A written decision that the records meet the test | Whoever is paid only after acceptance | Quality evidence that supports the next sale | Share accepted on first delivery |
| Onward licensing | Revenue from buyers | The intermediary, as principal or as agent | Repeat revenue only if the license allows reuse | Revenue retained after supplier payment |
Which models carry the assurance, and which pass it along
Comparing rights-cleared data companies on revenue alone mixes businesses that carry the assurance with businesses that hand it to someone else. Normalize instead to one accepted dataset and ask who stands behind the rights, whether personal data is dealt with before delivery and what the next dataset costs to produce.
Marketplaces show how assurance can be templated without being carried. AWS Data Exchange supplies a default data subscription agreement, designed to set baseline clauses on use, warranty and indemnification, which providers may edit or replace with their own before it attaches to an offer. What a subscriber is promised therefore depends on each provider's terms.
A comparison on those terms, an editorial classification from the cited documents and common contract structures rather than a rating of any named company:
| Model | Who stands behind the rights | Personal data before delivery | Normalize revenue by | What the label can hide |
|---|---|---|---|---|
| Publicly licensed dataset | No one: warranties disclaimed in 29 of 31 licenses | Usually left out of the grant and to the user | Not a revenue line; the user's own review is the cost | Free use read as cleared use |
| Marketplace listing | Each provider, under the agreement on its offer | As each provider's terms say | Transaction value kept per listing that sells | Platform revenue read as data revenue |
| Commissioned collection | The collector, which usually assigns output to the buyer | Designed into the collection | Revenue per labor hour | Services sold as datasets |
| Direct rights-cleared licensing | The supplier warrants; the intermediary licenses onward | Before delivery, by the intermediary or a partner | Revenue retained per accepted dataset | One-off preparation read as recurring licensing |
Licensing asset or repeat labor? Three tests
The strongest objection to the model is that it is labor with a licensing label: if every dataset needs fresh qualification and fresh preparation, margins look like consulting and the rights position adds little. The objection holds whenever the license cannot be reused and the buyer owns the output.
Three tests separate the cases. Does the next dataset from a returning supplier take less work than the first, because the master agreement and qualification carry over? Can the cleared dataset be licensed again, to another buyer or for another permitted use, without new permission? Is pricing set per dataset and rights scope, or per hour of work? A company that fails all three is selling labor, whatever its pitch says.
Exclusivity cuts both ways. A supplier's promise not to license the same records to others for AI training makes the cleared dataset scarce, which supports price; the asset is still only as reusable as the intermediary's own onward rights permit. Ask how often a cleared scope has produced more than one sale, and on what terms.
Applied to SourceX: where its published terms put each stage
SourceX is a labeled application of this question, described only from its published pages. It licenses business data directly from companies and licenses it onward to AI developers in its own name, and it runs each deal end to end, from sourcing and rights review to contracting, preparation, delivery and payment. Nothing reaches it until the company has approved the scope and terms and signed; from then on its use, sublicensing and distribution of the de-identified dataset stay within what that agreement allows.
Its standard deal terms place each stage on the ledger. A supplier signs one master agreement after qualification and a short Dataset Deal Confirmation for each deal, and a returning supplier reuses the master. The supplier keeps ownership of its original data, while SourceX owns the de-identified dataset, subject to the supplier's underlying intellectual property. Licenses are exclusive for AI training, a restriction that binds the supplier and leaves SourceX's right to distribute or sublicense the de-identified dataset untouched; the term and any alternatives are fixed in each Confirmation. De-identification is done by SourceX or its processing partner.
Acceptance risk and the rights warranty sit with the supplier until Closing. SourceX reviews delivered data and closes only once it has accepted the data in writing; data that never reaches Closing earns nothing. The supplier confirms that it may license the data, that personal data was lawfully collected, that the data has not been licensed for AI training before and that it was not created or altered with AI for the sale. Payment is usually a one-time fee without royalties, and once it is paid SourceX's license to the de-identified dataset is permanent.
On the ledger, that puts the authority warranty and rejection risk with the supplier, qualification and preparation labor with SourceX and its partner, and a sublicensable rights position with SourceX. SourceX has not published how often a cleared dataset is licensed more than once, its labor per dataset or its acceptance rates; it gives company and commercial metrics to investors directly. Its raise is open to accredited investors under Rule 506(c), and materials can be requested on its investors page.
What could shrink the clearance premium
Each of these would narrow what buyers pay for clearance, and each can be checked against the figures in the diligence list that follows:
- Public licenses could start carrying assurance. The CDLA history points the other way, but a widely adopted license with a real rights representation would compress what a paid seller adds.
- The law could lower the cost of skipping clearance. If courts or legislators in a buyer's main markets treat training on lawfully accessed data as permitted without a license, demand would narrow to records whose value lies in privacy handling and exclusivity rather than permission.
- Supplier warranties may not be collectible. A warranty from a mid-sized company is worth what that company can pay on a claim, and buyers may demand intermediary indemnities that move the risk onto the thinnest margin in the chain.
- Labor may not fall. If preparation hours stay flat as suppliers return, the business scales like a services firm and the licensing label flatters it.
- Exclusivity may cost more than it earns. A sublicensable position pays only if second buyers exist; in a market with few buyers, exclusivity is a restriction the supplier is paid once to accept.
- Templated marketplace terms could become good enough. Default agreements such as AWS Data Exchange's already template warranty and indemnification clauses; if buyers accept templated provider promises, intermediated clearance becomes harder to price.
Diligence questions that price the clearance
Put each question per accepted dataset, name the denominator, and ask every company in the comparison for the same figures:
- Who signed the authority warranty on each dataset sold last year, and what remedy does the buyer have if it fails?
- How many candidates were screened for each accepted dataset, and at which stage did the failures stop?
- How many labor hours did a first dataset take, and a repeat dataset from the same supplier?
- What share of delivered datasets passed acceptance on first delivery?
- How many licenses has each cleared scope produced, and what limits further licensing?
- Which indemnities does the company give buyers, how are they capped and what backs them?
How the license count was made, and what it cannot show
The count used a sparse clone of github.com/spdx/license-list-data at commit 31ba1a50e5397e00a304dbadc76531740e89ee48 (16 September 2026; License List 3.29.0), read on 27 October 2026 by a script that opens only json/licenses.json and the license text files. A license was in scope if it was not deprecated and either its name contained the whole word Data or Database or the phrase Open Government (case-sensitive), or its text contained database right or sui generis (any case). Each clause test is a case-insensitive pattern run on the text with whitespace collapsed, and a license counts once per test. Disclaimer: as is or as-is, without representations, without warranty, no warranty, any word starting disclaim, or the French sans autre garantie. Representation: represents that or warrants that. Indemnity: any word starting indemnif. Personal-data carve-out: not licensed, not cover, not apply or not grant followed within 120 characters, in the same sentence, by privacy, publicity, personal data or personal information, or exemption followed within 80 characters by personal data or personal information. Computational: machine learning, artificial intelligence, computational use or computational analysis. The patterns are English-only, and the carve-out test needs the negation to come before the privacy word, so two results were coded by hand. Etalab Open License 2.0, whose French text promises the reuser that intellectual property rights do not stand in the way of the grant, is counted as a representation, giving 3 rather than the pattern's 2. ODbL 1.0 and ODC-By 1.0, whose section 2.4 names data protection, privacy and personality rights before saying the license does not cover rights in the contents, fall outside the carve-out rule as written, so they are reported beside the 14, not in it. Every hit on the representation, indemnity, carve-out and computational tests was read by hand, and the script and a per-license results file are kept for reviewers.
The method has limits. A pattern can miss a clause phrased another way, and the Solderpad Hardware License entered scope only because it mentions database rights. SPDX texts are canonical versions, not the copies attached to particular datasets, and a license says nothing about whether a given dataset was lawfully assembled. The stage ledger and model comparison are editorial frameworks built from the cited documents and SourceX's published terms; no SourceX metrics were used, and no row rates a named company.
Frequently asked questions
What does rights-cleared data mean in a business model?
It means the seller has confirmed authority to license the records for a stated use, dealt with third-party and personal data before delivery, and stands behind that confirmation in a contract. The business model is the work and risk of producing that confirmation, priced into each dataset.
Are public data licenses a substitute for rights clearance?
Rarely for enterprise records. In a SourceX count of 31 data licenses on the SPDX list, 29 disclaim warranties, none has the licensor indemnify the user and 14 exclude privacy, publicity or personal data from the grant. They grant permission, not assurance.
Which stage of a rights-cleared data deal creates the most value?
Licensing, when it produces a rights position the intermediary may license again within its agreement. Qualification and preparation are necessary but repeat for every new dataset, so they behave like labor costs.
How should investors compare rights-cleared data companies?
Per accepted dataset: who stands behind the rights, how many labor hours the dataset took, whether it passed acceptance on first delivery and how many licenses the cleared scope has produced. Gross revenue mixes companies that carry assurance with companies that pass it along.
Does exclusivity make a rights-cleared dataset more valuable?
It makes the dataset scarcer, which can support price, but its value depends on whether the holder may license onward and whether second buyers exist. Exclusivity without a reusable onward license earns once.
Sources
- The SPDX License List data repository at commit 31ba1a5 (16 September 2026, License List 3.29.0) holds the license metadata and texts used for the 31-license count. Source
- CDLA-Permissive 2.0 section 4.1 provides data without representations, warranties or conditions of any kind, including title and non-infringement. Source
- CDLA-Permissive 1.0 section 4.1: each data provider represents it has exercised reasonable care that the data was created by it or obtained with the right to publish, and that publication violates no privacy or confidentiality obligation it undertook. Source
- CC BY 4.0 states that moral rights are not licensed, nor are publicity, privacy and similar personality rights. Source
- The UK Open Government Licence 3.0 does not cover personal data in the information. Source
- The Solderpad Hardware License 0.51 lets a redistributor charge for warranty or indemnity only if it indemnifies each contributor. Source
- The Etalab Open License 2.0 promises the reuser that intellectual property rights held by third parties or by the licensor do not stand in the way of the rights it grants. Source
- ODbL 1.0 section 2.4 says the license does not cover rights in a database's individual contents other than database rights, and names data protection, privacy and personality rights among those the contents may carry. Source
- ODC-By 1.0 section 2.4 carries the same contents clause as ODbL 1.0. Source
- SourceX's How it works page sets out its standard deal terms: master agreement and Dataset Deal Confirmation, supplier ownership, de-identified dataset ownership, exclusivity for AI training, acceptance and Closing, one-time fee, permanent license and supplier confirmations. Source
- SourceX's investors page says it licenses business data directly from companies and licenses it onward to AI developers in its own name, and shares company and commercial metrics with investors directly. Source
- AWS Data Exchange provides a default data subscription agreement designed to set baseline clauses on use, warranty and indemnification, which providers can edit or replace. Source
- The Montreal Data License paper (2019) proposes a taxonomy for licensing data in AI and machine learning, including a right to use data to evaluate models separate from other rights. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.