Skip to content

AI data market

AI data demand downside scenario: 4 futures, 8 dated signposts

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

The AI data demand downside scenario is real but uneven. Repeating data for up to about four epochs costs little, synthetic data now fills large parts of disclosed training mixes, and labs collect first-party usage data, all of which can shrink licensed budgets. Yet Meta's model cards show pretraining tokens rising from over 15 trillion to about 40 trillion in a year, with licensed data named in the mix. Track dated signposts, not one forecast.

License terms

AI data demand downside

This page covers AI data demand downside scenario: 4 futures, 8 dated signposts.Illustrative

Key takeaways

  • Four paths could shrink licensed data budgets: learning more from each token, synthetic data, reinforcement learning with distillation, and first-party usage data.
  • Exposure is uneven: generic annotation, public-like text and consumer conversation logs are most exposed; records with real outcomes, rare exceptions and private evaluation sets are least.
  • Meta's own disclosures cut both ways: pretraining tokens rose, disclosed fine-tuning moved from human annotation to synthetic examples, and licensed and product data entered the mix together.
  • Two questions produce four scenarios: how much new data a capability needs, and whether substitutes pass tests built from real work.
  • As of 26 October 2026 no scenario wins: pretraining signposts read Scarcity holds, post-training signposts read Generated volume, and only consumer conversation data shows Thin market conditions.

Could AI models need less licensed data than investors expect?

Yes, for some budget lines, and an investor sizing a data licensing thesis should know which ones before trusting a demand forecast. The scarcity case is well documented: Epoch AI puts the effective stock of quality public human text near 300 trillion tokens and projects that, if trends continue, models will fully use it between 2026 and 2032. That window is often read as proof that demand for licensed data must rise.

Epoch's projection is conditional on current trends continuing. If models learn more from each token, generate their own training data or collect it from their own products, the same window can arrive with smaller licensed budgets. This article separates those substitution and efficiency paths, builds four scenarios from them and scores eight dated signposts, so the forecast can be checked rather than believed.

Volume and value need separating too. Fewer tokens purchased can coexist with higher prices for the records that stay scarce, so the useful question in an AI data demand downside scenario is not whether demand falls but which purchases fall and which hold.

Four paths that could shrink licensed data budgets

Four mechanisms have public evidence behind them, and each shrinks a different line in a model developer's data budget. None of them removes the need for records a substitute cannot be checked against.

The repetition finding matters most for pretraining. Muennighoff and colleagues found that, with data constrained and compute fixed, training on repeated data for up to about four epochs changed loss very little compared with unique data, with returns falling toward zero after that. A developer that can reuse its corpus four times needs fewer new tokens.

Synthetic data and reinforcement learning matter most for post-training. NVIDIA's Nemotron documentation says its models use synthetic data in post-training, that the pre-training corpus includes synthetic and rephrased sources, and that synthetic data is validated through small-scale ablations. DeepSeek reports that DeepSeek-R1-Zero developed reasoning through large-scale reinforcement learning without supervised fine-tuning as a preliminary step, and that six smaller dense models were distilled from R1 using samples it generated.

First-party data is the quietest path. Anthropic's consumer terms update of August 28, 2025 lets Free, Pro and Max users choose whether their chats are used to train models, with retention extended to five years for those who allow it. Meta's Llama 4 model card says pretraining drew on information from Meta's products and services, including publicly shared Instagram and Facebook posts and people's interactions with Meta AI.

Four paths that could shrink licensed data budgets
PathPublic evidence, datedBudget line it shrinksWhat it does not replace
Learning more from each tokenMuennighoff et al., 2023: up to about 4 epochs of repeated data performed close to unique dataNew bulk text for pretrainingDomains the model has never seen
Synthetic dataLlama 3.1, July 2024: over 25 million synthetically generated fine-tuning examples; Nemotron: synthetic data in pre- and post-trainingAnnotation and instruction dataReal outcomes needed to check what was generated
Reinforcement learning and distillationDeepSeek-R1-Zero, 2025: reasoning trained without supervised fine-tuning; R1 distilled into six smaller modelsPurchased reasoning demonstrationsTasks with no automatic way to verify an answer
First-party usage dataAnthropic, August 2025: consumer chat training by user choice; Llama 4, April 2025: Meta product data in pretrainingLicensed conversational dataEnterprise systems of record

One model family, twelve months of disclosures

Meta's own model cards show volume rising, not falling, even as the mix shifted toward synthetic and first-party data. This is a small, verifiable evidence set: the official cards for three Llama releases, read from Meta's GitHub repositories on 26 October 2026.

Three readings follow. Pretraining volume went up: Llama 4 Scout's roughly 40 trillion tokens compares with the over 15 trillion stated for Llama 3, which cuts against this article's headline, although Llama 4's tokens are multimodal, so the comparison is indicative rather than like for like. Disclosed fine-tuning moved from human annotation toward synthetic examples between April and July 2024, which supports the headline. And the Llama 4 card is the first in this sample to name licensed data in the pretraining mix, in the same sentence as data from Meta's own products: licensed and first-party data arrived together.

Sampling frame and exclusions: one publisher was chosen because its three cards share a disclosure format, which makes them comparable. Intermediate Llama releases and every closed model without token disclosures are excluded, so the sample shows direction for one open-weight family, not the market.

One model family, twelve months of disclosures
DisclosureLlama 3 (April 18, 2024)Llama 3.1 (July 23, 2024)Llama 4 Scout and Maverick (April 5, 2025)
Pretraining tokens statedOver 15 trillionAbout 15 trillionAbout 40 trillion (Scout) and about 22 trillion (Maverick), multimodal
Pretraining sources statedPublicly available sourcesPublicly available sourcesPublicly available data, licensed data and information from Meta's products and services
Tuning data statedPublic instruction datasets and over 10 million human-annotated examplesPublic instruction datasets and over 25 million synthetically generated examplesFor safety data: human-generated data from vendors combined with synthetic data
Meta user or product dataNot included in pretraining or fine-tuning dataNot mentioned in the training data overviewIncluded: publicly shared Instagram and Facebook posts and interactions with Meta AI

Four scenarios for licensed data demand

Two questions generate four distinct futures: how much new data a capability needs as learning efficiency improves, and whether substitutes, meaning synthetic, first-party or public data, pass tests built from real work in place of private records. Each scenario names what would refute it, so a reader can tell which one the next disclosure supports.

The scenarios are not equally bad for a licensing business. Fewer, sharper purchases shrinks volume but rewards whoever holds rare, outcome-linked records. Generated volume moves bulk spending to generation while keeping seeds, verifiers and evaluation sets in demand. Thin market is the true downside: budgets shrink across lines, and the survivors sell evaluation, provenance and access to systems of record rather than training volume.

Four scenarios for licensed data demand
ScenarioEfficiency and substitutesLicensed budgetsWhat still sellsWould be refuted by
Scarcity holdsLittle efficiency gain; substitutes fail real testsGrow with model scaleBulk licensed text and specialized recordsToken counts in new model cards flattening
Fewer, sharper purchasesStrong efficiency gains; substitutes fail real testsVolume falls and spend concentratesRare cases, real outcomes and held-out evaluation setsSynthetic or self-generated data passing evaluations built from real work
Generated volumeLittle efficiency gain; substitutes pass for bulk dataBulk purchases shift to generationSeeds, verifiers and evaluation dataDisclosed synthetic shares falling in new releases
Thin marketStrong efficiency gains; substitutes pass widelyShrink the mostEvaluation, provenance and system-of-record accessBuyers returning to license again after model upgrades

Public forecast scorecard: eight signposts, read on 26 October 2026

The scorecard states what each signpost currently shows, which scenario it points toward and when to check it next, so the forecast can be scored in public. Each reading cites a dated source; the direction column is this article's judgment, made with the scoring rule in the methods note.

One row needs a caution. Reddit's S-1 says its January 2024 data licensing arrangements ran for two to three years, which on this article's reading puts the end of those initial terms between early 2026 and early 2027. Its annual report for 2024 adds that two partners have so far supplied substantially all of its licensing contract value, and that renewal is not assured: the agreements could lapse, or continue on terms less favorable to Reddit. Whether they were renewed was not verified here.

Tally as of 26 October 2026: two signposts point to Scarcity holds, two to Generated volume, two to Fewer, sharper purchases, one to Thin market for a single data type, and one is unresolved. No scenario wins. Pretraining signposts still read Scarcity holds, post-training signposts read Generated volume, and only consumer conversation data shows Thin market conditions. For an investor, that argues for underwriting licensing businesses built on non-public, outcome-linked records and evaluation, and for discounting plans that depend on generic annotation or public-like text. This is a dated reading of public evidence, not a forecast of any company's revenue.

Public forecast scorecard: eight signposts, read on 26 October 2026
SignpostReading and sourcePoints towardNext check
Disclosed pretraining tokensOver 15 trillion (Llama 3, April 2024) to about 40 trillion (Llama 4 Scout, April 2025)Scarcity holdsThe next major open-weight model card
Human versus synthetic tuning dataOver 10 million human-annotated examples (Llama 3) to over 25 million synthetic examples (Llama 3.1)Generated volumeThe next Llama or Nemotron disclosure
Synthetic data in pretrainingNemotron FAQ: quality-filtered and synthetic web crawl is about 49% of Ultra's Phase 1 tokensGenerated volumeThe next Nemotron technical report
Reasoning without human demonstrationsDeepSeek-R1-Zero trained by reinforcement learning without supervised fine-tuning (2025)Fewer, sharper purchasesWhether the method reaches tasks without automatic verifiers
First-party data collectionAnthropic consumer opt-in (August 28, 2025); Meta product data in Llama 4 pretrainingThin market, conversational data onlyThe next consumer terms or model card change
Public text exhaustionEpoch AI: public stock fully used between 2026 and 2032 if trends continueScarcity holdsLabs reporting data limits in model documentation
Buyer appetite for non-public dataOpenAI partners for non-public content and does not pursue paid deals for purely public informationFewer, sharper purchasesChanges to OpenAI's data partnership pages
Licensing renewalsReddit: January 2024 deals ran two to three years; 2024 annual report says they may not be renewedUnresolvedReddit's next periodic filing
A timeline from 2024 to 2027 placing dated signposts for licensed data demand, each marked as pointing toward more or less licensed data, with the Epoch public-text window and a today marker.
Public forecast scorecard, read on 26 October 2026. Sources: Meta model cards, Anthropic, Epoch AI, Reddit filings.

Which purchases are most exposed to the downside?

Generic annotation, public-like text and consumer conversation logs are most exposed; operational records with real outcomes, rare exceptions and held-out evaluation sets are least. The dividing line is whether a substitute can be checked against reality without the original records.

Evaluation is the clearest durable line. Anthropic's engineering guidance on evaluating agents calls human graders the gold standard used to calibrate model-based graders, while noting that they are expensive, slow and often need access to experts at scale. Tests that reflect real work, and the expert judgment to grade them, do not get cheaper because generation does.

Buyer statements point the same way. OpenAI says it partners to access non-public content such as archives and metadata and does not pursue paid partnerships for purely public information. A licensing business whose records are non-public, tied to outcomes and hard to recreate sits in the line the downside scenarios shrink last.

  • Most exposed: instruction and preference annotation that a strong model can generate and an automatic check can verify.
  • Exposed: text that resembles what is already public, where repetition and rephrasing stretch existing corpora.
  • Least exposed: records of real decisions and their outcomes, rare exceptions, and private evaluation sets that must stay unseen by the model being tested.

Ways this scenario analysis could be wrong

The scorecard can mislead in both directions, and these are the reasons to hold it loosely. Each one names the evidence that would expose it.

  • Disclosure bias. Model cards are written by the developer and disclose selectively; closed models disclose less, so the sample may not reflect frontier practice.
  • A one-family sample. Three Llama releases show direction for one publisher; other developers may be moving differently.
  • Substitute quality can move fast. If synthetic or self-generated data starts passing evaluations built from real work, Thin market gets closer quickly; if generated data needs more real seeds than expected, Scarcity holds strengthens.
  • Buyer concentration. A few developers set most demand, and Reddit's filings show how much licensing value can sit with two partners, so one strategy change can outweigh many signposts.
  • Efficiency can raise total demand. Cheaper training can mean more models trained for more tasks, which could lift demand for specialized records even as each model needs fewer; this is a hypothesis, not a measured finding.

What the downside means for a licensing thesis, SourceX included

The downside analysis changes what a licensing business should sell more than whether it should exist. SourceX helps companies license selected business records as datasets for AI, and it sources each buyer's dataset on request from U.S. businesses instead of holding stock, so there is no inventory of generic text for a price drop to strand. SourceX's published rule of thumb is that buyers want several years of steady history, usually at least two, and tens of thousands of records or more in a single system, though a smaller set can still qualify if the work it captures is rare. That favors the long-history, rare-work records this analysis expects to shrink last.

None of this proves demand for SourceX. SourceX has not published commercial metrics and reserves company and commercial figures for investors, who receive them directly. It is raising under Rule 506(c) from accredited investors; readers can request investor materials through sourcex.si/investors. Questions worth bringing to SourceX or any licensing company: what share of buyer requests are for training versus evaluation, how many buyers came back after a model upgrade, and which signposts in this scorecard the company tracks.

Methods note: sample, scoring rule and what was not measured

The paths table, scenario matrix, scorecard and exposure tiers are editorial frameworks built on 26 October 2026. Sources: Epoch AI's data-stock projection; Muennighoff et al. (2023); Meta's Llama 3, Llama 3.1 and Llama 4 model cards, DeepSeek's R1 repository and NVIDIA's Nemotron FAQ, all read from GitHub; Anthropic's consumer terms update and evaluation guidance; OpenAI's data page; and Reddit's S-1 and 2024 annual report, each verified on its publisher's domain.

Scoring rule: a signpost points to a scenario when its reading matches that scenario's row in the matrix; readings that match none are marked unresolved. Not measured: prices, budgets or volumes of licensed data at any developer, closed-model training mixes, the renewal outcomes of Reddit's agreements, and any SourceX data. The scorecard will be wrong in places, and each row's next-check event tells a reader where to look for the evidence that shows which.

Frequently asked questions

Will synthetic data replace licensed data?

For some lines it already has: Meta's Llama 3.1 card lists over 25 million synthetically generated fine-tuning examples. Synthetic data replaces purchases most easily where an automatic check can verify the output. Where checking needs real outcomes or expert judgment, real records remain the reference that synthetic data is tested against.

Are AI companies running out of training data?

Epoch AI projects that, if trends continue, models will fully use the stock of public human text between 2026 and 2032. Repetition, synthetic data and first-party data stretch that stock, so running out of public text does not translate one for one into licensed demand. Non-public business records sit outside that public-text estimate.

Which licensed data is least exposed to falling demand?

Records of real decisions with their outcomes, rare exceptions and private evaluation sets. Each is hard to generate convincingly, and evaluation sets must stay unseen by the models they test. Generic annotation and public-like text are the most exposed.

How often should an investor update an AI data demand scenario?

At each signpost's next check: new model cards, technical reports, consumer terms changes and licensing counterparties' periodic filings. Between those events the readings rarely change. Score the old reading against the new evidence rather than rewriting the scenario to fit it.

Does this scenario analysis say anything about SourceX's results?

No. It uses public evidence about the market and contains no SourceX metrics. SourceX shares company and commercial metrics with investors directly, investor materials are available on request through sourcex.si/investors, and any offering is made only to verified accredited investors.

Sources

  • Epoch AI estimates about 300 trillion tokens of quality public human text and projects full use between 2026 and 2032 if trends continue. Source
  • Muennighoff et al. (2023) found that up to about four epochs of repeated data changed loss very little compared with unique data under fixed compute. Source
  • Llama 3 (April 18, 2024) was pretrained on over 15 trillion tokens from public sources and fine-tuned with over 10 million human-annotated examples. Source
  • Llama 3.1 (July 23, 2024) was pretrained on about 15 trillion tokens and fine-tuned with over 25 million synthetically generated examples. Source
  • Llama 4 Scout and Maverick (April 5, 2025) were pretrained on about 40 and 22 trillion multimodal tokens from public, licensed and Meta product data. Source
  • NVIDIA's Nemotron FAQ says synthetic data is used in post-training and pre-training and validated through small-scale ablations. Source
  • DeepSeek-R1-Zero was trained through large-scale reinforcement learning without supervised fine-tuning, and six dense models were distilled from R1. Source
  • Anthropic's August 28, 2025 update lets Claude Free, Pro and Max users choose whether chats train models, with five-year retention if they opt in. Source
  • OpenAI partners for non-public content such as archives and metadata and does not pursue paid partnerships for purely public information. Source
  • Anthropic calls human graders the gold standard for calibrating model-based graders, and notes they are expensive, slow and need experts at scale. Source
  • Reddit's S-1 says its January 2024 data licensing arrangements had terms of two to three years. Source
  • Reddit's 2024 annual report says that, to date, substantially all licensing contract value comes from two partners and the agreements may not be renewed. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies