Building the licensing path for business data in AI.
SourceX licenses business data directly from companies and licenses it onward to AI developers in its own name, managing sourcing, rights review, contracting, preparation, delivery and payment.
Sourcing
Rights review
Contracting
Preparation
Delivery
Payment
X
Thesis
The next gains in AI depend on real business work, and businesses own it.
Why now
Public text is running out
Epoch AI estimates that, if current trends continue, language models will fully use the stock of human-generated public text between 2026 and 2032.
OpenAI's GDPval grades models on tasks built from real work products by experienced professionals in 44 occupations. That kind of work lives in business records, which only the businesses hold.
Software, BPO and contact centers, professional services, finance, healthcare administration and logistics, matched against documented demand from AI labs and model developers.
Direct licensing
SourceX licenses data directly from businesses and supplies licensed datasets to AI developers under documented rights and permitted uses.
Standardized licensing contract
One repeatable agreement shortens legal review for both sides on every deal.
Rights and privacy tooling
Ownership checks and removal of personal details are built into the workflow, not handled ad hoc.
Acquisition Partner anonymity
Labs can source data without revealing their roadmap or identity to the market.
What would have to be true for the SourceX investment thesis to hold?
Three conditions, and only the first has strong public evidence: AI developers keep paying for data they cannot collect from the open web, businesses that hold such data can license it with rights they can document, and an intermediary can do the rights, preparation and contracting work at a cost both sides accept. The second and third are where diligence belongs.
What would have to be tr…
tap a node
Condition
Developers pay for data they cannot scrape
Public evidence todayLlama 4 card names licensed data (April 2025); Reddit licensing arrangements of $203.0 million (Form S-1, 2024)
What would weaken itSynthetic or distilled data passes buyer acceptance tests at lower cost
Proof point to requestDistinct buyers with a written dataset request over the trailing four quarters
Each node is a row of the table below.From this section
Two dated signals support the first condition. Meta's model card for Llama 3.1, released July 23, 2024, describes about 15 trillion pretraining tokens from publicly available sources; its card for Llama 4, released April 5, 2025, describes a mix of publicly available data, licensed data and information from Meta's products and services. Reddit's Form S-1, filed in 2024, disclosed data licensing arrangements entered in January 2024 with an aggregate transaction price of $203.0 million.
Neither signal concerns business operating records or SourceX's buyers. Together they show that licensed data has become a named training input at one major developer and a disclosed revenue line at one public company. The Thesis Map below sets each condition beside its evidence, what would weaken it and the proof point to request.
What would have to be true for the SourceX investment thesis to hold?
Condition
Public evidence today
What would weaken it
Proof point to request
Developers pay for data they cannot scrape
Llama 4 card names licensed data (April 2025); Reddit licensing arrangements of $203.0 million (Form S-1, 2024)
Synthetic or distilled data passes buyer acceptance tests at lower cost
Distinct buyers with a written dataset request over the trailing four quarters
Records of real work are the input that matters
The GDPval card above: models graded on real professional work products
Evaluation moves to simulated tasks that need no company records
Share of buyer requests that name a task and an acceptance test
Companies will license, with rights they can document
None public at company level; this is the open question
Contracts, notices or vendor terms block most candidate datasets
Signed agreements per qualified company, by cohort, with rejection reasons
An intermediary can do the work at a cost buyers accept
None public; preparation cost is specific to each company
Rights review and de-identification stay manual and costly per deal
Preparation hours and cost per accepted dataset, failures included
How is a SourceX data transaction structured?
SourceX is licensee first and licensor second: it licenses selected records from a company under a signed agreement, prepares a de-identified dataset and licenses it onward to AI developers. Each step below comes from facts SourceX has published.
The steps below have three consequences for diligence. Rights risk sits in the signed agreements, so those documents, not summaries of them, are what to review. Revenue depends on onward licenses, so buyer breadth matters as much as supply. And because SourceX contracts with both sides, the gap between paying suppliers and collecting from buyers sets the working capital each transaction needs.
1Demand comes first. Datasets for AI buyers are sourced on request, not held in stock; a buyer states scope, volume, language, format, timing and licensing needs.
2Supply is qualified before data moves. A company starts with a short assessment that needs no data upload; as a general guide, the history buyers look for runs consistently for two years or more across tens of thousands of records or more in one system, assessed deal by deal.
3The contract comes before delivery. A company's approval of scope and terms, and its signature on the agreement, precede the first record SourceX receives; SourceX's right to use, sublicense and distribute the de-identified dataset flows from that agreement and stops where it stops.
4Preparation sits with SourceX, which handles de-identification in house or through its processing partner; large deliveries stay in the company's own storage or ship on encrypted drives.
5Terms are set deal by deal. Exclusivity and term are written into each deal's signed agreement, and SourceX publishes no prices, payouts or deal sizes.
Which SourceX investor metrics matter, and how is each defined?
Eight metrics keep demand, supply, acceptance, revenue and cash apart, each with a written definition, a denominator and a dated cohort. SourceX shares company and commercial metrics with investors directly; this page publishes the definitions, not the results.
Which SourceX investor m…
tap a node
Metric
Buyer breadth
Definition to insist onDistinct buyer organizations with a written dataset request naming scope and use
Denominator and windowAll buyer organizations in contact, trailing four quarters
What it exposesWhether demand rests on a few developers
Each node is a row of the table below.From this section
First reading rule: contract value is not revenue. Reddit's 2024 Form S-1 expected at least $66.4 million of its $203.0 million aggregate transaction price to be recognized as revenue in 2024, with the rest later. Ask whether revenue is reported gross or net of amounts owed to suppliers, because the answer changes every ratio built on it.
Second reading rule: treat any growth or ranking claim in this market as unsupported unless it names the metric, the baseline period, the comparison set, the observation window and an independent source. This page makes no growth or ranking claim about SourceX.
Which SourceX investor metrics matter, and how is each defined?
Metric
Definition to insist on
Denominator and window
What it exposes
Buyer breadth
Distinct buyer organizations with a written dataset request naming scope and use
All buyer organizations in contact, trailing four quarters
Whether demand rests on a few developers
Buyer concentration
Share of recognized licensing revenue from the largest buyer and from the largest two
Total recognized licensing revenue in the same period
Renewal and pricing exposure
Supply conversion
Companies with a signed licensing agreement
Companies qualified in the same cohort month, open cases shown as open
Whether rights or willingness limits supply
Acceptance rate
Datasets accepted against the agreed specification, with rejection reason codes
Datasets delivered in the same window
Whether preparation meets buyer standards
Contract value versus revenue
Signed contract value, revenue recognized and amounts owed to suppliers, side by side
The same contracts and period for all three
Whether contract value is presented as revenue
Cash cycle
Days between paying a supplier and collecting from the buyer for the same dataset
Every dataset with both payments made, plus those still open
Working capital each transaction ties up
Realized supplier earnings
Amounts actually paid to suppliers after acceptance, net of reversals
All suppliers with a signed agreement in the cohort, not only those paid
Supplier economics without estimates
Preparation cost per accepted dataset
Staff hours and outside processing cost, failed and abandoned datasets included
All datasets started in the cohort
Whether operating leverage is measured or assumed
What should SourceX investor due diligence ask for?
Documents and cohort tables, read against the definitions above, rather than summaries. SourceX is raising from accredited investors under Rule 506(c); it shares investor materials on request through this page, and any offering is made only to verified accredited investors through its offering documents.
SourceX has not published a completed-transaction case study. For any transaction described in investor materials, ask for the full trail: the supplier agreement, the buyer license, the delivery and acceptance records, the invoices and the cash receipts.
1Which Thesis Map condition does the company consider least proven, and what evidence would change its view?
2What do the supply conversion and acceptance cohorts show for the last four quarters, open cases and rejection reason codes included?
3How many distinct buyers sent written requests, and how much of recognized revenue came from the largest one?
4How does signed contract value reconcile to recognized revenue, amounts owed to suppliers and cash, and is revenue reported gross or net?
5Has any data moved before a signed agreement, and where are exceptions recorded?
6Which risk below would hurt most, and what early signal does the company track for it?
What are the SourceX investment risks, and what signals each one?
Five developments could make the thesis wrong. Each comes with a signal to monitor and, where one exists, a public source.
Synthetic data substitutes for real records. Meta's Llama 3.1 card (July 2024) says its fine-tuning data included over 25 million synthetically generated examples, and DeepSeek's R1 repository (2025) reports fine-tuning smaller open models on reasoning data generated by DeepSeek-R1, with strong benchmark results. Signal: buyer acceptance tests that synthetic sets pass at lower cost.
Demand stays concentrated. Reddit's annual report for 2024 ties substantially all of the contract value associated with its licensing revenue to two partners and warns they may not be renewed, or may be renewed on worse terms. Signal: the buyer concentration metric above.
Rights fail at the source. Customer contracts, privacy notices and software vendor terms can stop a company from licensing records it holds, whatever it owns. Signal: rising rejections at rights review in the supply conversion cohorts.
Law moves either way. Rulings or statutes that make training on unlicensed content easier would cut the premium on licensed public content, while stricter treatment of de-identified records would raise preparation cost and shrink usable supply. Unpublished operating records are less exposed to the first change and more exposed to the second. Signal: court rulings and regulator guidance on training data.
Execution and cash. Rights review and de-identification may not get cheaper per deal, and paying suppliers before buyers pay would turn each new transaction into a funding need. Signal: preparation cost per accepted dataset and the cash cycle, by cohort.
Methods note. The Thesis Map, the metric definitions and the questions above are editorial frameworks written for this page on 2026-10-26. Market evidence comes from Meta's Llama 3.1 and Llama 4 model cards, Reddit's Form S-1 and its annual report for 2024, and DeepSeek's R1 repository, each dated in the text; SourceX facts come only from its approved public facts. Not measured: any SourceX result for these metrics, demand for SourceX datasets, and the share of candidate datasets that clear rights review at any company. SourceX shares company and commercial metrics with investors directly; none is published here.
This article is for information only and is not an offer to sell or a solicitation of an offer to buy securities. Any offering by SourceX is made only to verified accredited investors through its offering documents.
Sources for this section
Meta's Llama 3.1 model card (released July 23, 2024) says the model was pretrained on about 15 trillion tokens from publicly available sources and that its fine-tuning data included over 25 million synthetically generated examples. Source
Meta's Llama 4 model card (released April 5, 2025) describes its pretraining data as a mix of publicly available data, licensed data and information from Meta's products and services. Source
Reddit's Form S-1 (2024) disclosed data licensing arrangements entered in January 2024 with an aggregate transaction price of $203.0 million, of which at least $66.4 million was expected to be recognized as revenue in 2024. Source
Reddit's annual report for 2024 says that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners and that the arrangements may not be renewed or may be renewed on worse terms. Source
DeepSeek reports fine-tuning several smaller open models on reasoning data generated by DeepSeek-R1 and says the distilled models perform well on benchmarks. Source