AI data market
Data marketplace standardization investment: 6 standards, 0 require quality
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Data marketplace standardization investment rests on a thin layer. In six public listing and data-product standards checked on 2026-10-27, identifiers and descriptive fields are well covered, but only two require a licence or usage policy, one requires it in machine-readable form, and none requires a quality measure. Credit a marketplace's standardization only where rights scope and acceptance terms are required and checkable; price the rest as services.
Checklist
Data marketplace
Key takeaways
- Public data standards cover identity, description and access; of six checked, two require a licence or usage policy and none requires a quality measure or a field-level schema.
- Rights travel in incompatible shapes: a licence URL, optional free text, or an ODRL offer whose action words are left open.
- Buyers judge AI data by task-specific acceptance tests that generic quality dimensions cannot encode, so quality is settled per deal.
- Measure any marketplace per accepted dataset: the share closed on unmodified standard terms and the hours each one takes, by cohort.
- A services-heavy model behind a platform label is a valid business, but it deserves a services cost curve, not a software one.
What should an investor test in a data marketplace standardization pitch?
Test whether the company standardizes the terms that set price and acceptance, meaning the scope of rights a buyer receives and the test a dataset must pass, or only the description of the data. Descriptions are cheap to standardize and already are. A marketplace earns software-like economics only if the expensive part of each deal, settling rights and quality, stops being negotiated one transaction at a time.
The decision here is narrow: whether to back a data marketplace whose investment case depends on standardization lowering the cost of each transaction as volume grows. Three kinds of evidence would support it. The listing format would require machine-checkable rights terms. Buyers would reuse acceptance tests across sellers. Seller-side work per accepted dataset would fall from one cohort to the next. Evidence against is the mirror image: optional rights fields, bespoke acceptance and flat labor per deal.
Public evidence can only speak to the first kind, so this page starts there. It counts what six open data-listing and data-product standards actually require, explains why rights and quality resist a common format, compares five marketplace routes by where the leftover work lands, and ends with the questions that test the other two kinds of evidence inside a company.
What do six public data standards require of a listing?
Very little beyond identity. Across six public standards checked on 2026-10-27 at pinned commits, a valid listing must identify the dataset or product, and only some formats also demand a title, a description or an access point. Two of the six require any licence or usage policy, one requires that policy in a machine-readable form, and none requires a quality measure or a field-level schema.
The six were chosen to cover the formats a data marketplace or data exchange would plausibly adopt: two data-product descriptors (Bitol's Open Data Product Standard and Open Data Mesh's Data Product Descriptor Specification), one data contract (Bitol's Open Data Contract Standard), one public-sector catalog profile (DCAT-AP 3.0.1, the European application profile of the W3C catalog vocabulary), one machine-learning dataset format (MLCommons Croissant 1.1) and one data-space catalog (the Eclipse Dataspace Protocol's 2025-1 schemas). Required means a document fails validation without the property; the full counting rules are in the methods note.
These are formats, not marketplaces. A marketplace can add required fields of its own, and some surely do. The count shows what the shared, public layer settles and therefore what any marketplace must settle privately, deal by deal or with its own proprietary fields. The table summarizes the minimum each standard demands:
| Standard (version, steward) | A valid listing must contain | Licence or usage policy | Quality measure | Field-level schema |
|---|---|---|---|---|
| ODCS v3.2.0 data contract (Bitol) | 4 of 23 top-level properties: version, apiVersion, kind, id | No licence property; usage, purpose and limitations are optional text | Optional checks with 7 named dimensions | Optional |
| ODPS v1.1.0 data product (Bitol) | 3 of 22: apiVersion, kind, id | No licence property; usage text optional | None; output ports cite a contract ID | None at product level |
| DPDS v1.0.0 descriptor (Open Data Mesh) | 3 of 7, plus 5 info fields: name, qualified name, version, domain, owner | Optional terms object that points to any specification | Optional SLO object, same pattern | Optional API definition per port |
| DCAT-AP 3.0.1 catalog profile (SEMIC) | Dataset: title and description (2 of 36); distribution: access URL (1 of 24) | Licence recommended; access rights, rights and ODRL policy optional | None in the shapes | None |
| Croissant 1.1 ML dataset (MLCommons) | 10 properties, including license, creator and distribution | Licence required as a URL or CreativeWork; use conditions optional | None in the core spec | Optional record sets |
| Dataspace Protocol 2025-1 (Eclipse) | @context, @id, a distribution and at least one policy offer | Required ODRL offer with a permission or prohibition | None | None |
Rights arrive as a pointer, a paragraph or a policy
The formats disagree on what rights even are, so two listings can both be valid and still be impossible to compare. Croissant 1.1 requires a license, but its expected type is a URL or a CreativeWork, which is a pointer to a document. The two Bitol formats have no licence property at all; intended usage, purpose and limitations are optional strings. DCAT-AP recommends a licence on each distribution and leaves an ODRL policy optional.
The Dataspace Protocol goes furthest. Its dataset schema requires hasPolicy with at least one offer, and the offer must carry a permission or a prohibition, so every catalog entry states its usage terms as data. Yet the schema types the action itself as a plain string. Two compliant catalogs can describe training a model with different words, and a buyer's software cannot tell whether they mean the same thing.
This matters more for AI training data than for open data. A licence link works when the terms are a public licence everyone recognizes. Licences for business records turn on terms that no URL captures: training versus evaluation use, exclusivity, duration, what happens to models already trained, and which records were excluded. Those terms live in negotiated contracts, and a format that cannot hold them pushes the negotiation back to people.
Regulation points the other way, at least inside European data spaces. Article 33(1)(a) of the EU Data Act (Regulation 2023/2854) requires participants in data spaces that offer data to describe dataset content, use restrictions, licences, collection methodology, data quality and uncertainty, where applicable in a machine-readable format. The phrase where applicable, and the article's data-space scope, leave room; the six public formats make most of those items optional, and none requires quality or uncertainty.
Why quality has no common unit across the formats
Because only two of the six formats have a quality slot, neither requires it, and only one fixes a vocabulary. ODCS v3.2.0 lets a contract list quality checks against seven dimensions: accuracy, completeness, conformity, consistency, coverage, timeliness and uniqueness. DPDS offers a service-level objective object that, like its terms object, simply points to whatever specification the producer chooses. The other four have no quality property in the core schema.
A buyer of AI training or evaluation data rarely defines quality in those generic terms. Its test is whether records support a named task: whether the outcome field is filled for the cases that matter, whether decision timestamps survive de-identification, whether rare exceptions appear often enough to grade. Seven dimensions can describe a table; they cannot say whether that table will pass one buyer's acceptance test and fail another's.
Croissant's Responsible AI extension adds documentation properties such as rai:dataLimitations and rai:dataBiases, and the specification types dataBiases as text. Narrative fields help a reviewer, but they are not pass-or-fail criteria. Quality, in other words, is still settled per transaction, which is exactly the work a standardization thesis claims to remove.
Where the standardizing work lands: five marketplace routes compared
In most routes the format standardizes discovery and access, and in every route the leftover work on rights and acceptance lands on someone's payroll; the investment question is whose. The table compares five routes a company can call a marketplace, by what the format settles, what stays negotiated, who absorbs the remainder and which metric shows whether that remainder is shrinking.
Normalize every metric to the accepted dataset, not to the listing. Listings that never transact flatter any per-listing figure, and a marketplace with ten thousand listings and twelve accepted datasets has a catalog, not a market. Ask for hours as well as counts, since labor that a company books as onboarding or customer success is still per-deal labor. The comparison:
| Route | What the format standardizes | What stays negotiated | Who absorbs the leftover work | Scope-normalized metric to request |
|---|---|---|---|---|
| Catalog listing marketplace | Titles, descriptions, formats and access links | Rights scope for AI use; acceptance tests | Seller and buyer legal and data teams, or marketplace staff | Share of accepted datasets closed on unmodified standard terms |
| In-place sharing exchange | Access path and table formats within one platform | Licence terms and quality acceptance | The provider's sales and legal staff | Contract negotiation hours per accepted dataset |
| Data-space connector network | Catalog entries and machine-readable usage offers | What each action term means; quality | Participants who write policies; connector operators | Share of offers using a shared action vocabulary; negotiation rounds per agreement |
| Data-contract software | Schemas and quality checks between teams | Licensing, which sits outside the format | Producer and consumer teams inside one company | Contracts in production per customer; not a licensing metric |
| Direct licensing with per-deal preparation | Nothing by format; each dataset is scoped to a request | Rights scope, preparation and acceptance | The licensor's own staff | Preparation hours per accepted dataset and the share reused for the next buyer |
Is the platform label covering services-heavy economics?
Often, and the six-standard count explains why. If the format cannot hold rights scope or acceptance criteria, someone has to write them for each deal, and that someone is usually on the company's payroll. A business that does this work is a services business with a catalog in front of it. That is not a flaw, but it deserves a services multiple and a services cost curve, not a platform one.
The strongest counter-case comes from the data-space route. The Dataspace Protocol shows that a format can make usage terms mandatory and machine-readable. In a vertical where participants agree on the meaning of a short list of action terms, the leftover work could shrink deal by deal, and a marketplace built on that agreement would start to look like software. The evidence to watch is not the number of listings but the share of agreements closed without a human editing the policy.
The decision rule follows. Credit standardization only for the part of the transaction the company has made both required and checkable for its target buyers. Treat everything else as labor, measure it per accepted dataset, and expect it to fall across cohorts before paying for it as leverage.
Developments that would weaken or strengthen the thesis
Four developments could move the thesis, and an investor should know which one a company is betting on; a fifth item marks the limit of this page's own count. Each development is tied to a dated public source where one exists.
- Mandates could standardize descriptions faster than markets do. The Data Act's Article 33 essential requirements already ask data-space participants to describe use restrictions, licences and quality, machine-readably where applicable; harmonized standards under it would raise the floor for everyone, which helps marketplaces and commoditizes the description layer at once.
- Intermediary rules could cap how value is captured. Article 12(a) of the Data Governance Act (Regulation 2022/868) bars data intermediation service providers from using the data for purposes other than making it available to data users, and requires the service to run through a separate legal person. A 2025 Commission staff working document describes a proposal to make that regime voluntary and remove the separate-entity requirement; this page did not confirm whether it has been adopted.
- Formats keep moving. ODCS's own schema accepts eight apiVersion values, from v2.2.0 to v3.2.0, and the DPDS schema stored in its v1.0.0 folder still declares a 1.0.0-DRAFT identifier. A marketplace that bets on one version carries migration cost when the format changes under it.
- Buyers may never accept a shared quality definition. If acceptance stays tied to each buyer's own model and task, no format will absorb it, and the per-deal labor becomes the product rather than a temporary cost.
- The count is not adoption. This page says what six formats require, not which marketplaces use them or how complete real listings are; a marketplace with strict proprietary fields could do better than any public standard.
Running the same test on SourceX
SourceX belongs in the last row of the routes table, and it should be judged as per-deal standardization rather than format-level standardization. It turns business records that a company selects into licensed AI datasets and handles the whole chain itself: sourcing and qualification, rights review, contracting, data preparation, delivery and payment. Datasets are sourced from U.S. businesses when a buyer asks, not held in stock, and the supplying company approves the scope and terms and signs its agreement before any data reaches SourceX.
That design accepts the labor the six-standard count predicts and puts it inside one company, so the questions that matter are about reuse and cost per accepted dataset, not about listing volume. SourceX has not published those figures or any other company metrics. Because its raise runs under Rule 506(c), any purchase is limited to verified accredited investors; the investor page takes requests for its materials, and the questions below can be put to the company there.
Diligence questions for any data marketplace, SourceX included
Ask each question against the accepted dataset as the unit, and ask for the trend across cohorts rather than a single figure:
- Which fields must a seller complete before a listing goes live, and does any of them state AI-use scope, exclusivity or term in a form your software can check?
- What share of last year's accepted datasets closed on your standard terms with no edits, and what share needed a negotiated amendment?
- How many buyer acceptance tests do you hold, and how many have been reused for a second buyer without changes?
- How many seller-side and staff hours does an accepted dataset take today, and how has that changed from your first cohort of sellers to your latest?
- Which public format, if any, do your listings follow, at which version, and what did the last format change cost you?
- If the Data Act's data-space requirements or a change to the Data Governance Act applied to you, which of your required fields would change?
Methods note: how the six standards were counted
The count was run on 2026-10-27 by a script that reads only data files from six repositories cloned at pinned commits: Bitol ODCS (04d1bfd, schema v3.2.0), Bitol ODPS (5f54a17, schema v1.1.0), Open Data Mesh DPDS (c220f69, schemas/v1.0.0), SEMIC DCAT-AP (4470b8e, release 3.0.1 SHACL shapes), MLCommons Croissant (ec24358, specification 1.1) and the Eclipse Dataspace Protocol (8c20a5b, 2025-1 catalog and negotiation schemas). For JSON Schemas, a property is required when it appears in a required list of its object or of an allOf branch, and it is required in every valid listing only if every property on its path from the root is required; alternatives inside anyOf or oneOf never count. For DCAT-AP, mandatory means a minimum count of at least one in shapes.ttl, and recommended means the same in shapes_recommended.ttl. For Croissant, required means listed in the specification's Required table or made required in its Modified and Added Properties section.
Slots were matched on the property's own name. A licence or usage policy slot is license, licence, rights, accessRights, hasPolicy, termsAndConditions, usageInfo or contractOffer (free-text usage notes not counted); a quality slot is quality, slo or hasQualityMeasurement; a field-level schema slot is schema, recordSet or api. A policy-language slot is one whose documented value is an ODRL policy or offer. Not measured: which marketplaces adopt these formats, how complete real listings are, proprietary fields a marketplace may add, prices, transaction volumes or any company's labor per dataset.
Frequently asked questions
What does standardization mean for a data marketplace?
It means the listing format itself carries the terms a buyer needs to decide, so deals close without bespoke negotiation. Public formats already standardize names, descriptions and access links. The open question for investors is whether rights scope and acceptance criteria are standardized too, because those are what drive cost per transaction.
Do any public data standards require a licence?
Two of the six checked on 2026-10-27 do. Croissant 1.1 requires a license, given as a URL or CreativeWork, and the Eclipse Dataspace Protocol requires at least one ODRL policy offer per dataset. DCAT-AP 3.0.1 recommends a licence, and the two Bitol formats have no licence property.
Why can't a marketplace simply adopt ODRL for rights?
It can, and the Dataspace Protocol does. ODRL gives terms a structure, but the action vocabulary is open, so two participants can describe the same use with different words. A shared, agreed list of actions for AI training and evaluation is what would make the terms comparable across sellers.
Is a services-heavy data marketplace a bad investment?
Not necessarily. It is a different business, with labor that scales with deals and margins that look like services. The mistake is paying a platform price for it. Measure staff and seller hours per accepted dataset and look for a decline across cohorts before treating the model as software.
Which single metric best shows standardization is working?
The share of accepted datasets that closed on unmodified standard terms, tracked by cohort. If that share rises while hours per accepted dataset fall, the format is absorbing work that people used to do. If it stays flat, the marketplace is selling negotiation capacity.
Sources
- Bitol's Open Data Contract Standard v3.2.0 JSON Schema requires four top-level properties, has no licence property, keeps usage, purpose and limitations as optional text and enumerates seven quality dimensions. Source
- Bitol's Open Data Product Standard v1.1.0 JSON Schema requires apiVersion, kind and id, and its output ports reference a contract ID rather than carrying licence or quality fields. Source
- The Open Data Mesh Data Product Descriptor schema (v1.0.0 folder) requires three top-level objects and five info fields; terms, billing, SLO and API entries are optional Standard Definition Objects. Source
- DCAT-AP 3.0.1 SHACL shapes make only title and description mandatory for a dataset and only the access URL mandatory for a distribution; licence, rights, access rights and ODRL policy are optional. Source
- DCAT-AP 3.0.1 recommended shapes list a licence as recommended, not mandatory, on each distribution. Source
- MLCommons Croissant 1.1 requires ten dataset-level properties including license, typed as CreativeWork or URL, and treats usage conditions (DUO terms or an ODRL offer in usageInfo) as optional. Source
- Croissant's Responsible AI extension adds documentation properties such as rai:dataLimitations and rai:dataBiases, with dataBiases typed as text. Source
- The Eclipse Dataspace Protocol 2025-1 dataset schema requires at least one policy offer (hasPolicy) and a distribution for every dataset. Source
- In the Dataspace Protocol contract schema, an offer must include a permission or a prohibition, and an action is typed as a plain string. Source
- Article 33(1)(a) of the EU Data Act requires data-space participants that offer data to describe dataset content, use restrictions, licences, collection methodology, data quality and uncertainty, where applicable in machine-readable format. Source
- Article 12(a) of the Data Governance Act bars data intermediation service providers from using the data for other purposes and requires the service to be provided through a separate legal person. Source
- A 2025 Commission staff working document describes a proposal to make the data intermediation regime voluntary and to remove the separate legal entity requirement. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.