AI data market
AI data investment demand thesis: who pays, for which task, judged how
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
An AI data investment demand thesis holds only when each market it claims names who owns the budget, which task the data must improve and how the buyer accepts a delivery. Public evidence names all three for no segment yet. Reddit's January 2024 licensing deals prove paid demand exists, but its FY2024 10-K ties substantially all of that contract value to two partners. A large data stock proves none of the three.
Value drivers
AI data investment demand
Key takeaways
- A demand claim becomes a market claim only when it names a budget owner, a task and an acceptance process, each backed by a dated buyer-side document.
- Paid demand for non-public content is on the public record but concentrated: Reddit's FY2024 10-K attributes substantially all of its licensing contract value to two partners.
- Public evidence splits: filings show budgets without tasks or tests, while evaluation work shows tasks and tests without budgets.
- Rising enterprise AI use and a shrinking stock of public text are reasons a market might form; neither shows anyone paying for training data.
- In diligence, ask for a request ledger with one row per written request, so each of the three names can be counted against a denominator.
What must an AI data investment demand thesis explain?
It must explain who pays, for what, and how they decide the purchase worked. For every market it claims, a credible AI data investment demand thesis names a budget owner with authority to pay an outside supplier, a task the data is meant to improve, and an acceptance process that turns a delivery into a payment. Remove any one of the three and what remains is a list of organizations that might buy someday.
The three names are deliberately narrow. A budget owner is a role, not a company: the person at a model developer or an enterprise who holds a line item that can pay outsiders, such as a data partnerships lead, a post-training lead, an evaluation lead or the owner of an internal AI program. A task is a specific job with a failure someone can observe, such as routing insurance claims or matching invoices to purchase orders. An acceptance process says who judges a delivery, against which test, before money moves.
This page turns that standard into a working tool for an AI infrastructure investment thesis. It explains why the most quoted evidence for demand does not meet the standard, audits seven claims that recur in AI data pitches, gives a measurement design an investor can request in diligence, and lists the developments that would prove the demand case wrong.
Why a large data stock, or rising AI use, is not purchasing demand
Because both describe conditions that could lead to a purchase, not a purchase. The strongest alternative to any demand thesis is that abundant data and widespread interest in AI can coexist with very few organizations willing to pay for training or evaluation records.
Scarcity is the first stand-in. Epoch AI's 2024 study put the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens and projected that models would use it fully between 2026 and 2032 if trends held. Its later outlook, What will AI look like in 2030?, calls a data bottleneck hard to rule out but surmountable, in part because synthetic data can be generated in large quantities. Neither finding names a buyer, a task or a test. Scarcity can raise the value of the right records without saying whose budget pays for them.
Adoption is the second stand-in. A Census Bureau working paper released in April 2026, drawing on the AI supplement to the Business Trends and Outlook Survey, found that 18 percent of U.S. firms used AI in a business function between November 2025 and January 2026, or 32 percent weighted by employment. Sales and marketing was the most common function, used by 52 percent of AI-using firms. Those firms are buying AI tools and services; the survey does not ask whether any of them bought training data, and few have a reason to.
A quick screen follows. If a fact in the deck would be just as true in a world where nobody ever paid for a dataset, it is context for the thesis, not evidence of demand.
The three-name test, and what counts as evidence for each name
Grade each name as evidenced, asserted or missing, using documents rather than descriptions. A name is evidenced when a dated document from the buyer side states it, asserted when the seller describes it without such a document, and missing when nobody can say.
Budget owners are the hardest name to evidence from outside, because buyers rarely publish who signs. Filings sometimes help: a licensor that discloses its contracts and its counterparties' share shows that some budget paid, even when the buying role stays unnamed. Tasks are easier, since model developers now publish the work they want models to do. OpenAI's GDPval, released in 2025, defines 1,320 tasks across 44 occupations, drawn from the work of professionals with an average of 14 years of experience. Acceptance processes sit in between: the methods are public, while the thresholds a buyer applies to a purchase are not.
Anthropic's guidance on agent evaluations shows what a written acceptance process contains. In its terms a task is one test with stated inputs and a stated standard for success, a well-made task is one on which two domain experts working separately would give the same pass-or-fail call, and 20 to 50 simple tasks taken from real failures are a sufficient start. A seller that can show a buyer's version of that document has evidence for the third name.
These grading rules keep the three grades repeatable from one company to the next:
- Budget owner. Evidenced: a written request, order or contract naming the buying function and the budget it draws on. Asserted: the seller names a team or a title but holds no buyer document. Missing: the claim names only a type of company, such as AI labs.
- Task. Evidenced: the request names the job, the input records and the failure to be reduced. Asserted: a capability area such as agents, with no specific job. Missing: the claim describes volume or modality only.
- Acceptance process. Evidenced: the test the buyer will run, who runs it, the pass threshold and what happens on failure. Asserted: a quality claim made by the seller, such as label accuracy. Missing: no test stands between delivery and payment.
- Dating rule. Every grade carries the date of its newest supporting document, and a grade resting on documents older than 12 months drops one level until it is refreshed.
The demand claim audit: seven pitches scored on the three names
None of the seven claims names all three with public evidence. Two name a budget owner, three name a task and two name an acceptance process, and the claims that dominate pitch decks, scarcity and adoption, name almost nothing. The table scores each claim as public sources stood on 26 October 2026, with the evidence that sets each grade and the observation that would falsify the claim.
Grades are the writer's judgment applied to the cited documents. A company holding private buyer documents could score its own version of a claim higher, which is exactly what diligence should ask it to show. The figure condenses the grid: filled marks for named, half marks for partly named, empty marks for missing.
| Demand claim | Budget owner | Task | Acceptance process | Dated evidence; falsified if |
|---|---|---|---|---|
| Developers pay for non-public content | Named: Reddit's S-1 (2024) disclosed data licensing deals from January 2024 worth $203.0 million in aggregate, with two- to three-year terms | Partly named: OpenAI's data partnerships page asks for large datasets not easily accessible online, not for a task | Missing: no buyer's acceptance test is public | Reddit FY2024 10-K: substantially all licensing contract value from two partners. Falsified if renewals lapse and no new licensee appears |
| Post-training buys vendor data | Partly named: Meta's Llama 3.1 model card says it combined human-generated data from vendors with synthetic data | Named: safety data meant to mitigate potential risks | Partly named: the card describes LLM-based classifiers that select high-quality prompts and responses | Same card: over 25M synthetically generated fine-tuning examples. Falsified if vendor data drops out of later model cards |
| Agent builders buy real work | Missing: no public source names who pays for workflow records | Named: OpenAI's GDPval (2025) defines 1,320 tasks across 44 occupations | Named as a method: Anthropic's guidance sets tasks with success criteria and expert-agreed pass or fail verdicts | Falsified if evaluation teams keep building tasks in-house and no written request for outside workflow records appears |
| Services revenue proves demand | Named: Innodata's FY2025 10-K attributes about 58 percent of 2025 revenue and 48 percent of 2024 revenue to one customer | Partly named: AI data services, task not disclosed | Missing: acceptance terms not disclosed | Shows labor spend, not licensing. Falsified as a licensing signal if no licensed-records line ever appears beside the services |
| Regulation forces documented data | Partly named: the provider of a high-risk AI system carries the duty | Named: the system's intended purpose | Named in law: EU AI Act Article 10 quality and data governance criteria | Falsified if providers meet Article 10 with data they already hold and buy nothing new |
| Every adopter needs data | Missing: adoption surveys do not ask about data purchases | Partly named: Census found sales and marketing the most common AI function, at 52 percent of AI-using firms | Missing | Census working paper (April 2026): 18 percent of U.S. firms used AI. Fails as stated; testable only once a function and a buyer are named |
| Scarcity pushes prices up | Missing | Missing | Missing | Epoch AI (2024): about 300 trillion tokens, fully used between 2026 and 2032 on trend. Falsified if synthetic or multimodal data closes the gap, which Epoch's 2030 outlook calls surmountable |
Reading the audit: budgets and tests never sit in the same row
The audit's main finding is a split: where a budget owner is visible, the task and the acceptance test are not, and where the task and the test are visible, the budget owner is not. No row pairs a named budget owner with a named acceptance process. That gap, not the size of the market, is what an AI data investment demand thesis has to close.
The budget side is visible mainly in filings, and it is concentrated. In Reddit's FY2024 10-K, the licensing line rests on two partners, which to date account for substantially all of its contract value. Innodata's FY2025 10-K attributes about 58 percent of total 2025 revenue, and 48 percent of 2024 revenue, to a single customer. Both are real budgets that paid outsiders, and both sit in very few hands. Neither filing publishes the buyer's task or its acceptance test.
The task side is visible mainly in evaluation work, and its budgets are hidden. GDPval and Anthropic's evaluation guidance describe tasks and tests in detail, because model developers publish evaluations to show progress. They say nothing about whether, or from whom, anyone buys records to build them.
Two readings follow for enterprise AI data investment. A thesis resting on frontier developers alone is a bet on a handful of budget decisions with renewal risk attached. A thesis resting on agent builders, enterprise AI programs or regulated providers needs private evidence of budget owners, because public sources stop at the task and the test. Either can be right; only the first can be checked from public documents today.
How to measure a company's demand claims: the request ledger
Ask for a request ledger: one row per written request from a prospective buyer, with each of the three names graded evidenced, asserted or missing. A ledger makes demand countable, gives every share a denominator, and stops a few memorable conversations from standing in for a market.
At a glance
How to measure a company's demand claims: the request ledger
Count only written requests received through a stated set of channels in a stated window, with the rule for what counts as a request fixed before counting begins. Keep open requests open rather than scoring them as wins or losses, and report how many are open beside any conversion figure. Separate repeat requests from the same organization so that one active buyer does not look like many.
Three metrics come out of a ledger. The share of requests with all three names evidenced shows whether demand is specified well enough to fulfill. The number of distinct organizations behind those requests shows breadth. The share of accepted deliveries that began as fully specified requests shows whether specification predicts payment, which is the claim the whole thesis rests on.
SourceX has not published results from a ledger like this; the design is offered as a method, and the same fields apply to any company that says buyers want its data. Record at least these fields for every request:
Checklist · 7 items
- Request ID, date received and channel, with the counting rule that admitted it.
- Requesting organization and the buyer function behind it: pretraining, post-training, evaluation, an enterprise AI program or compliance.
- Budget owner grade, with the document that sets it: order, contract, written request or none.
- Task named in the request and the failure it should reduce, in the buyer's own words.
- Acceptance process grade: the test, who runs it, the threshold and the consequence of failing.
- Status on the reporting date: open, declined by the buyer, declined by the seller, delivered, accepted or paid, with a reason code for every decline.
- Repeat flag: whether the organization has requested before, and whether its earlier delivery was accepted.
0 of 7 ticked. Tick each item as you go.
Risks: four ways an AI data demand thesis fails
Four developments would weaken any AI data demand thesis, and each already has a public marker an investor can track. They belong ahead of the upside in any memo.
- Budgets stay concentrated. Two partners stand behind substantially all of Reddit's licensing contract value in its FY2024 10-K, and a single customer behind about 58 percent of Innodata's 2025 revenue in its FY2025 10-K. A market with a few buyers carries renewal, pricing and timing risk that no supply advantage removes.
- Synthetic data fills the task. Meta's Llama 3.1 model card reports fine-tuning on publicly available instruction datasets and over 25M synthetically generated examples. Wherever synthetic data passes a buyer's acceptance process, the budget for real records on that task shrinks.
- Adoption never becomes data purchasing. The Census working paper finds AI use concentrated in sales and marketing, strategy and IT functions. If firms keep buying tools rather than training or evaluating models on their own work, enterprise demand for licensed records stays small.
- Acceptance moves in-house. Anthropic's guidance tells teams to start evaluations from 20 to 50 tasks drawn from their own real failures. If buyers build evaluation sets from internal records, outside suppliers compete for a smaller and harder slice of the work.
Questions that test demand claims at any AI data company, SourceX included
Ask each question about a specific claim in the company's own materials, and ask for the document that answers it. The list applies to SourceX as much as to anyone.
SourceX is one clearly labeled application of this investment question. Its work is converting business records that companies select into licensed AI datasets, and it takes on each stage itself: sourcing and qualifying, rights review, contracting, preparation, delivery and payment. Its buyer side starts from a request rather than an inventory: datasets come from U.S. businesses once a buyer asks for them, and the request-data form collects scope, volume, language, format, timing and licensing needs, which is where a request's task and licensing terms first get written down.
No demand, request or commercial metrics from SourceX are public; company and commercial metrics are disclosed to investors directly. The company's Rule 506(c) raise is limited to accredited investors, and sourcex.si/investors is where materials are requested and where any question below can be put to it.
- For each market in your materials, which budget owner, task and acceptance process can you document, and how old is the newest document?
- How many written requests arrived in the last four quarters, through which channels, and what share had all three names evidenced?
- How many distinct organizations stand behind those requests, and what share came from the largest one?
- Which requests ended in an accepted delivery, and were they the fully specified ones?
- Which requests went to synthetic data, public data or a buyer's in-house team instead, and on what test?
- Which slide states a market size or an adoption figure without naming a buyer?
Methods note: how the demand claim audit was scored
The three-name test, the grading rules, the demand claim audit and the request ledger are editorial frameworks built for this article on 26 October 2026. The grades rest on public documents: Reddit's 2024 S-1 and FY2024 10-K, Innodata's FY2025 10-K, OpenAI's data partnerships page and GDPval announcement, Meta's Llama 3.1 model card, Anthropic's guidance on agent evaluations, two Epoch AI publications, an April 2026 Census Bureau working paper on AI use and Regulation (EU) 2024/1689. Each grade is the writer's judgment applied to those documents.
Not measured: private buyer requests, contract terms, acceptance thresholds and budget sizes at any company, SourceX included; demand from buyers that publish nothing; and whether specified requests actually convert to payment more often. The audit can show that public evidence does not support a claim. It cannot show that private evidence is absent.
Frequently asked questions
What is an AI data investment demand thesis?
It is the part of an investment case that explains who will pay for AI training or evaluation data, for which task and under what acceptance process. A strong one names all three for every market it claims and backs each with a dated document. A weak one points to the size of the data stock or the pace of AI adoption instead.
Does rising enterprise AI adoption mean companies will buy training data?
Not by itself. A Census Bureau working paper found 18 percent of U.S. firms used AI in a business function between November 2025 and January 2026, most often in sales and marketing. Those firms buy tools and services; a training-data purchase also needs a task and an acceptance test, which adoption surveys do not ask about.
Who controls AI training data budgets at model developers?
Public sources rarely say. Licensor filings such as Reddit's show that budgets exist and are concentrated, and Meta's Llama 3.1 model card mentions data from vendors, but the buying function is seldom named. That is why the budget owner is the name an investor most needs to see in private documents.
Is the scarcity of public text evidence of demand for licensed data?
No. Epoch AI's estimate of about 300 trillion tokens of usable public text, fully used between 2026 and 2032 on trend, describes supply. It names no buyer, task or test, and Epoch's later outlook judges the bottleneck surmountable, partly through synthetic data.
Can investors see SourceX's buyer request data?
Not publicly. Request, demand and commercial figures are unpublished, and SourceX reserves company and commercial metrics for investors, sharing them directly. Accredited investors can request materials through sourcex.si/investors and ask how a request ledger like the one on this page would be defined.
Sources
- Reddit's S-1 disclosed data licensing arrangements entered in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years. Source
- Reddit's FY2024 Form 10-K states that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
- Innodata's FY2025 Form 10-K states that one customer in its Digital Data Solutions segment generated approximately 58 percent and 48 percent of total revenues in 2025 and 2024. Source
- OpenAI's data partnerships page seeks large-scale datasets that are not already easily accessible online and says it is not seeking sensitive or personal information or information that belongs to a third party. Source
- OpenAI's GDPval covers 1,320 tasks across 44 occupations, built from the work of professionals with an average of 14 years of experience. Source
- Anthropic's agent evaluation guidance defines a task as a single test with defined inputs and success criteria, says a good task is one where two domain experts would reach the same pass/fail verdict, and suggests 20-50 tasks drawn from real failures to start. Source
- Epoch AI estimated the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens and projected full use between 2026 and 2032 on current trends. Source
- Epoch AI's 2030 outlook finds a data bottleneck difficult to fully rule out but surmountable, noting synthetic data can be generated in large quantities. Source
- A Census Bureau working paper (CES-WP-26-25, April 2026) found 18 percent of U.S. firms used AI in a business function from November 2025 to January 2026, 32 percent employment-weighted, most often in sales and marketing (52 percent of users), strategy and IT. Source
- Article 10 of Regulation (EU) 2024/1689 requires training, validation and testing data for high-risk AI systems to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete, under data governance practices that cover data origin. Source
- Meta's Llama 3.1 model card says fine-tuning data included publicly available instruction datasets and over 25M synthetically generated examples, and that safety data combined human-generated data from vendors with synthetic data selected by LLM-based classifiers. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.