AI data market
Data infrastructure and data ownership: two investment bets, not one
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Data infrastructure and data ownership are different investments. Infrastructure sells work on data customers already may use, and free tools now cover most of it: each core stage, from extraction to PII and quality scoring, has matching code in four of five Apache-2.0 curation toolkits, and none records who granted rights. Ownership sells the permission itself, so its risks are concentration, renewals and courts. Value each on its own unit and metrics.
License terms
Data infrastructure data
Key takeaways
- Infrastructure is paid for work done to data the customer already has rights to; ownership is paid for the rights.
- Each core processing stage has matching code in four of the five open-source curation toolkits counted here, all under Apache 2.0.
- No toolkit ships a stage that records who granted permission; the rights-adjacent steps only read labels already attached to the data.
- Cloud terms keep hosting and owning apart, and the EU Data Act ends switching charges for data processing services on 12 January 2027.
- A February 2025 U.S. trial-court opinion enforced an owner's rights against a competing non-generative AI tool, the mechanism that gives licenses their price.
Are data infrastructure and data ownership the same investment bet?
No. A data infrastructure company sells work done to data its customer already has the right to use; a data ownership company sells the right itself. Infrastructure revenue grows with the volume customers push through it and is exposed to free substitutes and to rules that lower switching costs. Ownership revenue grows with the number of rights it can license and licensees it can sign, and is exposed to buyer concentration, renewals and courts.
The two bets get blended because the same pitch deck can claim both. A pipeline vendor that touches large volumes of customer records starts to sound like a data owner, and a rights holder that runs its own cleaning stack starts to sound like a software company. An investor who values the first on the second's multiple, or the reverse, misprices the risk that actually decides the outcome.
This article separates them with evidence a reader can check: a count of what five open-source curation toolkits ship, the line the AWS Customer Agreement draws around customer data, an EU rule that removes one infrastructure moat, a court opinion on enforcing ownership against an AI use, and a comparison table normalized to the unit each bet actually sells.
Five free toolkits already cover most processing stages
Most of what an AI data pipeline does to records is now available free under a permissive license. Five open-source curation toolkits, read at commits dated between September 2025 and October 2026, all carry the Apache License 2.0: NVIDIA's NeMo Curator, Hugging Face's DataTrove, Data Prep Kit, the Allen Institute for AI's Dolma toolkit and Unstructured. For each of five core stages, extraction, deduplication, language identification, PII detection and quality scoring, four of the five toolkits ship matching code; two ship decontamination against benchmarks.
The projects describe themselves as general machinery. NeMo Curator's README says it helps teams build pipelines that "load, filter, deduplicate, and transform large text, image, video, and audio datasets for AI training", and that it powers the data pipelines behind NVIDIA's Nemotron models. DataTrove calls itself "a library to process, filter and deduplicate text data at a very large scale".
For an investor, that makes processing function a commodity input rather than a moat. A company selling processing has to win on what the free code does not provide: managed scale, accuracy on hard documents, integrations, support or compliance. The count below uses a file-path rule, so it is a floor; capabilities living in differently named files are missed. Toolkits with matching code, by stage:
| Stage | Toolkits with matching code | Which toolkits | Example file |
|---|---|---|---|
| Extract text from raw formats | 4 of 5 | NeMo Curator, DataTrove, Data Prep Kit, Unstructured | src/datatrove/pipeline/extractors/base.py |
| Deduplicate | 4 of 5 | NeMo Curator, DataTrove, Data Prep Kit, Dolma | nemo_curator/stages/deduplication/exact/identification.py |
| Identify language | 4 of 5 | NeMo Curator, DataTrove, Data Prep Kit, Dolma | python/dolma/taggers/language.py |
| Detect or redact PII | 4 of 5 | NeMo Curator (tutorial), DataTrove, Data Prep Kit, Dolma | src/datatrove/pipeline/formatters/pii.py |
| Score or filter quality | 4 of 5 | NeMo Curator, DataTrove, Data Prep Kit, Dolma | src/datatrove/pipeline/filters/gopher_quality_filter.py |
| Decontaminate against benchmarks | 2 of 5 | DataTrove, Dolma | src/datatrove/pipeline/stats/contamination_stats.py |
| Read a license label on the data | 2 of 5 | Data Prep Kit, Dolma | python/dolma/taggers/licenses.py |
| Record a grant of rights | 0 of 5 | None | No matching file |
The stage no toolkit ships: a record of who granted what
None of the five toolkits ships anything that records who granted permission to use the data, under which contract, for what purpose. No code file in any of them has a path naming consent, contract, permission, provenance, rights or grant.
The two rights-adjacent stages that do exist read a label that someone else attached. Data Prep Kit's License Select transform "checks if the license of input data is in approved/denied list", working from a license column supplied with the input. Dolma's cc_re tagger scans a page's HTML for links to Creative Commons licenses and tags the license type it finds. Both are useful filters for public data that arrives with a declared license; neither can create a right for records that arrive without one.
That gap is the boundary between the two bets. Infrastructure can deduplicate, redact and score a company's support tickets perfectly and still leave the question of whether anyone may train on them exactly where it started. The value that depends on that answer accrues to whoever holds, or can lawfully acquire, the right.
Hosting is not owning: what a cloud contract says about customer data
Holding a customer's data does not give the host the right to license it. The AWS Customer Agreement, in its section on customer content, says that apart from what that section provides "we obtain no rights under this Agreement from you" to the customer's content, and that the customer consents to its use to provide the services.
Clauses like this cap how far a hosting or processing business can move toward ownership. A provider that stores petabytes of business records holds them as a service supplier. Turning that position into a licensing business would require new rights from each customer, granted in a new contract, which is the ownership bet's core work rather than an extension of the infrastructure bet.
The same distinction applies to software vendors inside operating systems of record. A CRM, ERP or ticketing vendor sees its customers' records every day; whether it can license them for AI training depends on what its customer agreements grant, and under terms like the AWS clause above it is the customer, not the vendor, who holds something to license.
Two bets on one page: a scope-normalized comparison
The cleanest way to compare the two bets is to normalize each to the unit it sells: processing or hosting volume for infrastructure, a licensed scope over time for ownership. Two swap tests then sort any company. The bring-your-own-data test asks what the company would still sell if every customer arrived with fully licensed records; whatever remains is infrastructure. The free-tools test asks what it would still own if a rival offered identical processing at no charge; whatever remains is ownership.
The strongest alternative explanation for a company that passes neither test cleanly is services. A data platform that runs bespoke pipelines for each client, or a rights business that re-collects data to every buyer's specification, has labor economics whatever its label. Ask for revenue split into software, licensed rights and services, each with gross margin, before applying any multiple. The comparison below gives the metric to request for each bet:
| Bet | What the customer pays for | Revenue grows with | Main erosion risk | Scope-normalized metric to request |
|---|---|---|---|---|
| Processing tools | Transforming data the customer already holds | Customer data volume and compute | Free toolkits: each core stage in four of five | Paid usage per unit processed beyond the free baseline; margin after compute |
| Hosted processing and storage | Running and storing customer data | Stored volume and compute hours | EU Data Act: switching charges end on 12 January 2027 | Net revenue retention by cohort; revenue tied to egress or switching |
| Source ownership | Permission to use records others cannot reach | Licensable scopes and number of licensees | Open or synthetic substitutes, buyer concentration, court rulings | Contract value per licensed scope-year; renewal rate; licensee count |
| Licensed-rights intermediary | Rights acquired from owners, plus preparation | Owners signed and licenses per dataset | Owners licensing directly; exclusivity limits | Rights cost per accepted dataset; datasets with a second license |
| Blended data platform | A mix of tools, rights and labor | Whichever line the services team feeds | Services drift hidden by a platform label | Revenue and margin split into software, rights and services |
Where infrastructure moats come from once processing is free
With processing functions free, infrastructure moats come from quality on hard inputs, bundling and lock-in, and regulation is now removing part of the last one in Europe. The open-core pattern is explicit in Unstructured's README: the open-source library "is, and will stay, completely free", while higher-quality output is offered through the company's paid Transform API. That is a quality moat, measurable by how much paid volume the free version fails to keep.
Bundling is the second source. NVIDIA, which sells the GPUs that NeMo Curator is built to accelerate, gives the toolkit away under Apache 2.0; a supplier that earns on the hardware underneath can price the software layer at zero, which squeezes independent tool companies selling the same stage.
Lock-in is the third, and the EU Data Act targets it. Article 29 allows providers of data processing services only reduced switching charges until 12 January 2027 and none for the switching process after that date, and the regulation's definition of switching charges excludes standard service fees and early termination penalties. An infrastructure thesis that relies on the cost of leaving, especially data egress, should be stress-tested against that date for EU customers. These rules are written for providers of data processing services; a license to source material is a different contract.
Can ownership be enforced against AI use? One court's answer
Yes, on at least one set of facts, and a U.S. court has said so in detail. In Thomson Reuters v. Ross Intelligence, Judge Stephanos Bibas, sitting by designation in the District of Delaware, granted partial summary judgment to Thomson Reuters in February 2025 and rejected Ross's fair use defense. The opinion found that Ross used Thomson Reuters's Westlaw headnotes as AI data to build a legal research tool competing with Westlaw, and held that the use was not transformative.
The limits matter as much as the result. The opinion says only non-generative AI was before the court, it concerns curated editorial material rather than raw business records, and it is one trial-court ruling that appellate courts could see differently. It still shows the mechanism the ownership bet depends on: when material is protected and the AI use competes with the owner's own product, the right can be enforced, which is what gives a license its price.
For infrastructure, the same ruling is neutral to slightly negative. A processing vendor gains nothing from stronger ownership rights and may face more contract terms requiring it to prove what it processed and under whose permission.
What would make the two-bets distinction wrong
The distinction holds only while rights and processing stay separable. These are the developments that would weaken it, each something an investor can watch:
- Hosts acquire rights. If cloud or software vendors rewrite customer terms to obtain training or licensing rights, part of the ownership bet moves to infrastructure; check each vendor's customer agreement rather than assuming the AWS position is universal.
- Ownership loses its price. Open corpora, synthetic data or appellate rulings broader than Ross on fair use could shrink what permission is worth for the material in question.
- Free tools stall. The toolkit count measures availability, not quality; if hard inputs such as scanned forms keep needing paid models, processing quality remains a real moat, as Unstructured's own positioning suggests.
- Regulation reaches ownership. Rules that compel data sharing on regulated terms could cap the price of certain owned records, just as switching rules now cap one infrastructure moat.
- Concentration cuts both ways. Owners selling to a handful of model developers face renewal risk; infrastructure sold to a handful of hyperscale customers faces the same. Ask for the top-buyer share in either case.
SourceX on the map, and questions for any company claiming both
SourceX sits on the ownership side of the line by its own description, though it is not the original owner. Its how-it-works page says the supplying business keeps ownership of its original data, while under the signed agreement SourceX owns the de-identified dataset, subject to the business's underlying intellectual property rights, and may distribute or sublicense it. Its investor page describes licensing data directly from companies and onward to AI developers in its own name, with SourceX handling sourcing, rights review, contracting, preparation, delivery and payment.
On the comparison table that is the licensed-rights intermediary row: the bet is execution in acquiring rights from many owners, not a processing stack that free toolkits could replace. SourceX has not published rights costs, license counts or second-license rates, and this article does not estimate them; its investor page says company and commercial metrics are shared with investors directly. Questions that separate the two bets in any company, SourceX among them:
- If every customer arrived with fully licensed records tomorrow, which revenue lines would remain, and at what margin?
- If a rival offered your processing free, which assets would you still hold, and which contract grants you each one?
- For each licensed dataset, who owns the original records, what right do you hold in the derived dataset, and does it allow a second license?
- What share of revenue is software, licensed rights and services, and how has the services share moved by cohort?
- Which revenue depends on switching costs or data egress for EU customers, and what changes after 12 January 2027?
- Which court rulings or contract changes would most reduce what your rights are worth?
How the toolkit count was made, and what it misses
Toolkit count: clones of NVIDIA-NeMo/Curator at fdee8f8, huggingface/datatrove at 4c488a1, data-prep-kit/data-prep-kit at bbe6608, allenai/dolma at 669f534 and Unstructured-IO/unstructured at 53a56a0, with last commit dates from 18 September 2025 (Dolma) to 27 October 2026. The script toolkit_matrix.py reads git ls-files output and LICENSE files only. Universe: tracked paths ending in .py, .rs or .ipynb, lowercased, excluding any path containing test or benchmark. A stage counts for a toolkit when re.search finds its regular expression in at least one path. The expressions, verbatim: extraction extract|html|partition|pdf2; deduplication dedup|minhash|duplicate; language lang_id|langid|(^|/|_)lid(/|\.|_)|language_id|language_identification|fasttext_lang|(^|/)language\.py$|lang_detect; PII (^|/|_|-)pii(/|\.|_|-)|presidio|redact; quality quality|classif; decontamination decontam|contamin; license label licen[cs]e; rights grant consent|contract|permission|provenance|(^|/|_)rights(/|\.|_)|(^|/|_)grant. Matches are printed for review, and a stage is reported for a toolkit only by this rule.
The rule finds files by name, so it misses capabilities in differently named files; Unstructured's language detection, for example, sits in a file named lang.py and is not counted, and all eight NeMo Curator language matches sit in its audio modules, while its text language identifier is a class named FastTextLangId in a fastText filter file. Counts are therefore floors, and a match shows code exists, not that it is production quality. Public sources were checked on 27 October 2026: the AWS Customer Agreement, the Data Act on EUR-Lex and the Ross opinion through searches limited to each publisher's domain, and SourceX's pages by direct fetch.
Not measured: revenue, margins, retention or usage of any company named here, open-source adoption, or the quality of any toolkit. The comparison table states which metric to request, not values. The next evidence needed is company-level revenue split by software, rights and services, which only companies can provide.
Frequently asked questions
What is the difference between data infrastructure and data ownership?
Data infrastructure is the tooling and hosting that processes, stores or moves data a customer already has the right to use. Data ownership is holding, or lawfully acquiring, the right to use and license the source material itself. Infrastructure is paid per unit of work; ownership is paid per right granted.
Are open-source data curation tools replacing paid data processing?
For common stages they already compete on function. Extraction, deduplication, language identification, PII detection and quality scoring each have matching code in four of the five Apache-2.0 toolkits counted here. Paid vendors now compete on scale, accuracy on hard inputs, integration and support.
Does a cloud provider own the data its customers store?
Not under AWS's standard terms. The AWS Customer Agreement says AWS obtains no rights to customer content beyond what its content section provides, with the customer consenting to use needed to deliver the services. A host would need new rights from customers to license their data.
How does the EU Data Act affect data infrastructure investments?
Its switching rules target lock-in at providers of data processing services. Article 29 permits only reduced switching charges until 12 January 2027 and none for the switching process after that, so revenue that depends on the cost of leaving should be stress-tested for EU customers.
Which metrics show whether an AI data company is an infrastructure or an ownership business?
Ask for revenue split into software, licensed rights and services with margins; for infrastructure, paid usage beyond the free baseline and net revenue retention; for ownership, contract value per licensed scope-year, renewal rates, licensee counts and top-buyer share.
Sources
- NeMo Curator, Apache-2.0, builds pipelines that load, filter, deduplicate and transform datasets for AI training and powers NVIDIA's Nemotron data pipelines. Source
- DataTrove, Apache-2.0, is a library to process, filter and deduplicate text data at very large scale. Source
- Data Prep Kit's License Select transform checks a supplied license column against an approved or denied list. Source
- The Dolma toolkit's cc_re tagger detects Creative Commons license links in page HTML. Source
- Unstructured keeps its open-source library free and offers higher-quality output through a paid API. Source
- Under section 6.1 of the AWS Customer Agreement, AWS obtains no rights to customer content beyond what that section provides. Source
- Data Act Article 29 phases out switching charges for data processing services by 12 January 2027; Article 2(36) defines switching charges. Source
- The February 2025 Thomson Reuters v. Ross opinion rejected fair use for headnotes used as AI data in a competing non-generative tool. Source
- SourceX's published terms on ownership of original and de-identified data and its right to sublicense. Source
- SourceX licenses data from companies onward in its own name and shares company and commercial metrics with investors directly. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.