AI data market
Next big thing in AI data? A call, watch or pass evidence checklist
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Call an AI data market the next big thing only when evidence clears five thresholds: several unrelated buyers pay, they renew, the data beats a cheaper substitute on the buyer's own test, rights-cleared supply can be added on standard terms, and each dataset earns more than it costs to clear and prepare. Filings through mid-2026 show real but concentrated, uneven demand, so most markets rate watch, not call.
Call to transcript
Next big thing AI data
Key takeaways
- A large data stock is not a market: Epoch AI puts public human text at about 300 trillion tokens, yet purchases concentrate in a few buyers.
- Breadth is where most theses stall: Reddit attributes substantially all licensing contract value to two partners, and one customer made up about 58 percent of Innodata's 2025 revenue.
- Durability is the next test: contracted licensing value runs off on a schedule unless renewed, and Shutterstock's data segment revenue roughly halved in the first quarter of 2026.
- Utility only counts against a credible substitute, and synthetic data already sits in production pipelines such as NVIDIA's Nemotron.
- Private business records rate open: OpenAI says it wants data that is not easily accessible online and grades models on real work, but no public price, renewal or margin data exists yet.
What evidence would justify calling an AI data market the next big thing?
Five kinds of evidence, all observable: demand from several unrelated buyers, demand that renews, measured utility against a cheaper substitute, rights-cleared supply that can be added on standard terms, and per-dataset economics that stay positive. When a market clears all five, call it. When demand is real but narrow or untested on renewal, watch it. When a filing contradicts demand or a substitute matches the data, pass for now.
Most next-big-thing arguments in AI data rest on supply. One version says enormous amounts of data exist; the other says the free supply of public text is running out. Both can be true, and neither shows that buyers pay repeatedly for a particular kind of data at prices that cover clearing the rights and preparing it. That gap separates a data category from a data market, and the market is what an investor actually underwrites.
The checklist below grades with public evidence dated through October 26, 2026: securities filings, lab publications, platform terms and research audits. It also marks where public evidence stops and diligence has to begin, which for the newest markets is almost immediately.
Five thresholds, each with the evidence that clears it and the evidence that breaks it
Each threshold has three readings: what counts as early evidence, what clears the bar, and what would falsify the claim. A market is called only when all five clear. It is watched when demand reaches at least the early reading and nothing is falsified. It is passed for now when a falsifier appears on demand or utility, and regraded when new filings arrive.
Two rules keep the grading honest. An announcement without disclosed terms counts as early evidence at most, because a partnership can be a pilot, a press release or a one-time transfer. And one falsifier outranks any amount of supportive evidence, because the point of a checklist is to find the reason a thesis fails before the market finds it.
| Threshold | Early evidence | Clears the bar | Falsified by |
|---|---|---|---|
| Demand breadth | One or two buyers carry most of the disclosed value | Several unrelated buyers on signed contracts, none dominant | The dominant buyer leaving or cutting spend with no replacement disclosed |
| Demand durability | First contracts signed, with value disclosed | Renewals or follow-on contracts after the first terms end | Contracted value running off unreplaced, or the data revenue line falling in consecutive periods |
| Utility against a substitute | Buyers saying they want the data | Measured lift on the buyer's own evaluation that beats a cheaper substitute at equal cost | A synthetic, public or in-house substitute matching the result |
| Rights-cleared supply | One licensor with clear rights | New suppliers cleared on standard terms without bespoke legal work | Each supplier needing custom terms, or platform terms closing the export route |
| Unit economics | Total revenue with no per-dataset view | Revenue per dataset above clearance, preparation and rework, with reuse or renewal | Revenue that depends on one-off delivery timing, or costs that grow with every deal |
Why a large data stock can coexist with weak purchasing demand
Because stock measures what exists and demand measures what someone pays for, and the two move independently. The strongest objection to any data-market thesis is that buyers already have more data than they can use, and cheaper ways to make more.
Start with the public stock. Epoch AI's 2024 study estimated that quality public human text, once adjusted for repetition, gives an effective stock of about 300 trillion tokens, with a 90 percent confidence interval of 100 trillion to 1,000 trillion, and projected that language model training would fully use it at some point between 2026 and 2032 on current trends. Scarcity of free text is real on that timeline. It does not follow that labs will buy any particular replacement.
Substitutes compete for the same budget. NVIDIA's Nemotron documentation says the models use synthetic data in post-training and that the pre-training corpus also includes synthetic and rephrased sources. OpenAI's GDPval benchmark took another route to real-work material: 1,320 tasks across 44 occupations, each based on a real work product and written by experienced professionals. A lab that can generate, rephrase or commission what it needs is a weaker buyer of licensed data, whatever the stock.
Business archives make the same point from the supply side. Many companies hold years of operating records, and the volume says nothing about whether any buyer has a task those records improve. A thesis that leads with terabytes has not yet reached the threshold that matters.
Who actually pays: what filings show about breadth and renewal
Filings show demand that is real, priced and concentrated, with renewal still the open question. The clearest public numbers come from three listed companies whose disclosures touch AI data, and each one tests a different threshold.
Reddit's S-1 disclosed that in January 2024 it entered into data licensing arrangements with an aggregate transaction price of $203.0 million, expected at least $66.4 million of it to be recognized in 2024, and said it delivers the content through continuous API access and quarterly data transfers. Its annual report for 2024 then stated: "To date, substantially all of the contract value associated with our licensing revenue is derived from two of our partners." At March 31, 2026, its remaining performance obligations were $120.6 million, primarily long-term content licensing contracts, with $91.6 million expected in the rest of 2026 and $29.0 million in 2027. That schedule works as a renewal clock: contracted value runs down unless new or renewed contracts replace it.
Shutterstock reports AI-related licensing inside its Data, Distribution, and Services segment, which includes access to its metadata for machine learning and generative AI training. Segment revenue was $21.0 million in the first quarter of 2026 against $39.7 million a year earlier. In its second-quarter report the company attributed the segment's decline mainly to its data offering and said revenue recognition there can vary from quarter to quarter with the delivery timing of metadata licenses.
Innodata, a listed data services company, shows the services version of the pattern. Total revenue was $251.7 million in 2025 against $170.5 million in 2024, and one customer in its data services segment generated about 58 percent of total revenue in 2025 and 48 percent in 2024. Spending grew while dependence on a single buyer rose.
Utility is the threshold most market theses skip
Utility means a measured improvement on a task the buyer cares about, against the cheapest credible alternative, and public evidence of it for licensed datasets is rare. Labs announce partnerships and publish benchmarks; they seldom publish the lift a specific licensed dataset produced. Until they do, utility claims are early evidence, not a cleared threshold.
The direction of lab measurement is still informative. GDPval grades models on deliverables built from real work products rather than exam questions, and OpenAI notes that it is limited to one-shot tasks that do not capture building context over several drafts. Evaluations anchored in real work raise the value of records that show how real work was done, and they also show that a lab can build such material itself.
When a company claims utility, ask for the design rather than the headline. A credible utility study has five elements:
- A buyer-owned evaluation set fixed before the data is delivered, with the task and metric named.
- Three arms: the base model, the base model plus the candidate data, and the base model plus the cheapest credible substitute, whether synthetic, public or commissioned.
- Equal compute and equal example or token budgets across arms, so the data, not the spend, explains the difference.
- Lift reported with intervals for each task slice, not one average across everything.
- A repeat after the next model release, to test whether the lift survives a stronger base model.
Unit economics: what must hold for each dataset
Each dataset has to earn more than it costs to clear, prepare, fix and deliver, and the margin has to come back through reuse or renewal rather than one-off timing. No market graded here discloses per-dataset economics, so this threshold is a diligence item everywhere.
Rights clearance is the cost most theses underprice. Operational records often sit in platforms whose terms limit how data leaves them: under Slack's API terms, an application offered outside the developer's organization cannot use API data to train a large language model, and bulk exports of message and file data need an additional agreement. Permission metadata is unreliable too. When the Data Provenance Initiative audited 1,800-plus text datasets (the work is published through the MIT Media Lab), it found popular hosting sites omitting licenses at rates above 70 percent and carrying errors at rates above 50 percent. Each gap is legal and operational work that grows with the number of suppliers unless terms are standardized.
Revenue shape matters as much as size. Recognition tied to delivery timing, as Shutterstock describes for metadata licenses, can make a market look larger in one quarter and smaller in the next without any change in underlying demand. A per-dataset ledger, with each closed dataset's revenue, supplier payment, preparation, rework and delivery cost, is the only way to tell a repeatable business from a run of large one-off deliveries.
The thesis map: four AI data markets graded on the same thresholds
On public evidence as of October 26, 2026, no market clears the call bar. Licensed platform text and managed data services rate watch, stock media metadata rates pass for now on durability, and private business operating records rate open, because the public record cannot grade them either way.
Open is a fourth verdict for a market whose demand evidence is not public at all. It is not a polite pass. It means the thesis can be tested only through diligence, which is where an early investor's edge would come from if the thesis is right, and where the risk sits if it is not.
| Market and public proxy | Demand evidence | Utility and substitutes | Supply and economics | Read, October 2026 |
|---|---|---|---|---|
| Licensed platform text, read through Reddit's filings | $203.0 million of arrangements from January 2024; substantially all contract value from two partners; $120.6 million contracted at March 2026, most due by 2027 | No public lift data; synthetic and rephrased text already used in production model pipelines such as Nemotron | One licensor controls supply; per-dataset margin not disclosed | Watch: real and priced, narrow, renewal untested |
| Stock media metadata, read through Shutterstock's filings | Data segment revenue $21.0 million in Q1 2026 against $39.7 million a year earlier; decline continued in Q2 | No public lift data | Recognition varies with the delivery timing of metadata licenses | Pass for now: segment revenue fell year over year in Q1 and Q2 2026, led by data |
| Managed data services, read through Innodata's filings | Revenue $251.7 million in 2025 against $170.5 million in 2024; one customer about 58 percent | Labs also commission expert work directly, as GDPval shows | Labor-based delivery, no record licensing to clear; project margins not disclosed | Watch: growing, with rising concentration |
| Private business operating records, no listed proxy | OpenAI states interest in datasets not easily accessible online; no contract values disclosed | OpenAI grades models on real work products in GDPval; no public lift data | Platform terms and unreliable license metadata raise clearance cost; no public economics | Open: only diligence can grade it |
What could make the private business records thesis wrong
Six things could keep private business records from becoming a market even if labs want them, and each has a public precedent in an adjacent market. Price these before the upside:
- Buyer concentration: if two or three labs are the only buyers, they set prices and terms, the pattern Reddit's and Innodata's filings show next door.
- Substitution: synthetic and rephrased data already sit in production training pipelines, as NVIDIA documents for Nemotron, and could cover more business tasks than expected.
- In-house creation: labs can commission professionals to build real-work tasks directly, as OpenAI did for GDPval, instead of licensing records.
- Uneven revenue: licensing recognized on delivery can swing from quarter to quarter, as Shutterstock describes for metadata licenses, which makes early growth hard to read.
- Rights friction: platform terms such as Slack's API restrictions, and license metadata that audits find often missing or wrong, can make each new supplier slow and costly to clear.
- Timing: Epoch AI's 2026 to 2032 range for using up public text leaves room for the scarcity argument to arrive later than a fund's holding period assumes.
How the checklist applies to SourceX, and the metrics to request
SourceX is one company building in the open row of the map, so the checklist applies to it as to any other. What it does is turn selected records from a business into datasets that business licenses for AI, taking each deal from sourcing and qualification through rights review and contracting to data preparation, delivery and payment. Buyers' datasets are sourced on request from U.S. businesses, with nothing held in stock, and the supplying company must approve the scope and terms and put its name to the agreement before any records change hands.
On-request sourcing is a design answer to the stock-versus-demand objection, not proof that the objection is beaten: whether written requests arrive broadly and repeatedly is exactly what the demand thresholds test. SourceX has not published demand, renewal or per-dataset figures. Its Rule 506(c) raise is from accredited investors, and those who ask for materials through sourcex.si/investors receive company and commercial metrics directly.
Whatever company you evaluate in this market, SourceX included, request these six metrics with their definitions:
Checklist · 6 items
- Distinct requesting buyers per period, and the largest buyer's share of all written requests in the same period.
- Repeat-request rate: buyers who submitted a second written request within a stated window, divided by buyers who submitted a first.
- Request-to-signature conversion: signed deals divided by qualified written requests, with requests still open reported as open rather than lost.
- Rights-clearance time from supplier qualification to signed scope, as a median with unfinished cases censored rather than dropped.
- A per-dataset ledger: revenue, supplier payment, preparation, rework and delivery cost for every closed dataset.
- Any buyer-reported utility result, with the evaluation and the substitute arm described.
0 of 6 ticked. Tick each item as you go.
Methods note: how the four markets were graded and what the grades leave out
The thresholds and the verdict rule are editorial frameworks defined in this article. Grades use only the cited public sources: Reddit's S-1 (2024), annual report for 2024 and quarterly report for March 31, 2026; Shutterstock's quarterly reports for March 31 and June 30, 2026; Innodata's annual report for 2025; OpenAI's data partnerships and GDPval pages; Epoch AI's 2024 data study; NVIDIA's Nemotron documentation; Slack's API terms; and the Data Provenance Initiative audit. All were checked on October 26, 2026.
Each listed company is read as a proxy for one market, not rated as a business or an investment, and one company's disclosures cannot represent a whole market. Contract values, remaining performance obligations and segment revenue are reported figures, not forecasts. Nothing here measures prices, margins, private transactions or any SourceX data, and the grades should be redone when new filings arrive. Most sources were verified through search excerpts on the publishers' own domains, so a reviewer should open each one before relying on a figure.
Frequently asked questions
What is the strongest public evidence that an AI data market is real?
Signed contracts with disclosed value from several unrelated buyers, followed by renewals. Reddit's filings show the first part, $203.0 million of arrangements from January 2024, alongside concentration in two partners, which is why licensed platform text rates watch rather than call.
Does running out of public text make data licensing the next big thing?
Not on its own. Epoch AI projects that the public text stock could be fully used between 2026 and 2032, but labs can fill gaps with synthetic, rephrased or commissioned data. Scarcity raises the question of what labs will pay for; only purchase and renewal evidence answers it.
Why grade private business records as open instead of watch?
Watch requires at least early public demand evidence, such as disclosed contracts. For private operating records, OpenAI, for one, has stated interest in data that is not easily accessible online, but no contract values, prices or renewals are public, so the thesis can only be graded through diligence.
Which metrics should an investor request from an AI data licensing company?
Buyer breadth with the largest buyer's share, repeat requests, request-to-signature conversion with open cases reported, rights-clearance time with unfinished cases censored, a per-dataset ledger and any buyer-reported utility result with its substitute arm. Each needs a stated denominator and period.
Has SourceX published demand or per-dataset figures?
No. None of those figures is public; investors who ask through sourcex.si/investors get company and commercial metrics from SourceX directly. Purchases of its securities are open only to verified accredited investors.
Sources
- 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 belonging to a third party. Source
- OpenAI's GDPval covers 1,320 tasks across 44 occupations, each based on a real work product written by experienced professionals, and is limited to one-shot tasks. Source
- Epoch AI estimated the effective stock of public human text at about 300 trillion tokens (90 percent CI 100 trillion to 1,000 trillion) and projected full use between 2026 and 2032. Source
- Reddit's S-1 disclosed January 2024 data licensing arrangements with an aggregate transaction price of $203.0 million, at least $66.4 million expected in 2024 revenue, delivered by continuous API access and quarterly data transfers. 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. Source
- Reddit's quarterly report for March 31, 2026 lists $120.6 million of remaining performance obligations, primarily long-term content licensing contracts, with $91.6 million expected in the rest of 2026 and $29.0 million in 2027. Source
- Shutterstock's quarterly report for March 31, 2026 shows Data, Distribution, and Services revenue of $21.0 million against $39.7 million a year earlier; the segment includes metadata access for ML and generative AI training. Source
- Shutterstock's quarterly report for June 30, 2026 attributes the segment's decline mainly to its data offering and says data revenue recognition can vary quarter to quarter with the delivery timing of metadata licenses. Source
- Innodata's annual report for 2025 shows total revenue of $251.7 million (2025) and $170.5 million (2024), and one data services customer at about 58 percent and 48 percent of total revenue. Source
- NVIDIA's Nemotron documentation says synthetic data is used in post-training and that the pre-training corpus also includes synthetic and rephrased sources. Source
- Slack's API terms bar applications offered outside the developer's organization from using API data to train a large language model and from bulk exporting message and file data without an additional agreement. Source
- The Data Provenance Initiative audit of more than 1,800 text datasets found license omission rates above 70 percent and error rates above 50 percent on popular hosting sites. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.