Skip to content

AI data market

AI lab data budget breakdown: six cost lines and which ones pay outsiders

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

No AI lab publishes its data budget by line item, so any breakdown is a reconstruction. Public evidence splits it into six lines: collection, licensing, curation, annotation, experiments and evaluation. Curation and experiments mostly buy a lab's own compute and staff; licensing and annotation pay outsiders. Epoch AI estimates only about 10 percent of OpenAI's 2024 R&D compute went to final training runs. Track the outside lines, and count paid commitments, not announcements.

Value drivers

AI lab data budget breakdown

This page covers AI lab data budget breakdown: six cost lines and which ones pay outsiders.Illustrative

Key takeaways

  • No lab discloses data spending by line, so investors have to reconstruct it from model cards, technical reports, seller filings and estimates.
  • Of six cost lines, licensing and annotation pay outsiders, evaluation and collection are mixed, and curation and experiments mostly stay inside the lab.
  • Headline training costs leave out the experiments that decide which data gets bought; DeepSeek's V3 report excludes ablations on data explicitly.
  • On public evidence only licensing reaches a paid commitment, and no cost line shows a repeat purchase yet.
  • Synthetic data is the strongest competing explanation: Meta's Llama 3.1 card counts over 25 million synthetic fine-tuning examples.

Can anyone see an AI lab's data budget?

Not directly. No major AI lab publishes its data spending by line item, so every AI lab data budget breakdown is a reconstruction from model cards, technical reports, the filings of companies that sell data to labs, disclosure laws and third-party estimates. The useful question for an investor is narrower: which parts of that budget pay outside suppliers, and what evidence shows money moving rather than interest.

The most cited public cost model shows why the data line is hard to find. Epoch AI's study of what it costs to develop frontier models attributes 47-67 percent of development cost to hardware, 29-49 percent to R&D staff, including equity, and 2-6 percent to energy. Those are the categories it measures, so whatever labs spend on data has to be found elsewhere: in staff time, in vendor contracts and in the compute used to filter, generate and test data.

This article reconstructs the budget as six cost lines, sorts them by who gets paid, ranks the public evidence for each on a six-rung demand-signal ladder, and publishes an interview design for the evidence that is still missing. It estimates no lab's spending, and it treats every announcement as a claim to be placed on the ladder rather than as proof of demand.

Six cost lines: what each one buys and who gets paid

A lab's data budget splits into collection, licensing, curation, annotation, experiments and evaluation. Curation and experiments mostly buy the lab's own compute and staff time; licensing and annotation pay outsiders; collection and evaluation are mixed. Only the outside lines form a market that a data supplier can sell into.

Collection is the cheapest line to buy into because most of it is not bought. Meta's Llama 3 model card, dated April 18, 2024, describes pretraining on over 15 trillion tokens from publicly available sources, which is mostly a crawling and processing cost borne inside the lab. Licensing is the opposite: it pays for material others cannot reach. OpenAI's data partnerships page, for instance, seeks large-scale datasets that the public cannot already find easily online.

Annotation and expert data are where outsiders are paid per task. Anthropic's Claude Opus 4.5 system card lists data provided by data-labeling services and paid contractors among Claude's training sources, alongside non-public data from third parties. Evaluation uses the same kind of skilled labor for a different purpose: OpenAI's GDPval paper describes 1,320 tasks across 44 occupations, built from the representative work of professionals with an average of 14 years of experience.

Read the table by its third column. A line that mostly pays the lab's own engineers and chips can grow quickly without a single new supplier contract, which is why capability news alone says little about outside data spending.

Six cost lines: what each one buys and who gets paid
Cost lineWhat it buysWho gets paidPublic evidence, datedSupplier opening
CollectionCrawling, archives and existing datasetsMostly the lab: crawler compute and staffLlama 3 model card (April 2024): over 15 trillion pretraining tokens from publicly available sourcesSmall: public data is collected, not bought
LicensingRights to non-public content and recordsOutside rights holdersReddit S-1 (2024): data licensing arrangements with $203.0 million aggregate contract valueDirect: the line a licensor sells into
CurationFiltering, deduplication, mixing and quality classifiersMostly the lab: compute and staffLlama 3.1 model card (July 2024): LLM-based classifiers select high-quality prompts and responsesIndirect: cleaner inputs save curation work
AnnotationDemonstrations, preference labels and expert writingLabeling vendors and paid contractorsLlama 3 model card (April 2024): over 10 million human-annotated examplesDirect, but exposed to synthetic substitutes
ExperimentsData ablations, mixture tests and scaling runsMostly the lab: R&D computeDeepSeek-V3 report (December 2024): headline cost excludes ablation experiments on dataIndirect: experiments decide which data gets bought
EvaluationTask construction, expert grading and test refreshMixed: outside experts plus internal gradersOpenAI GDPval: 1,320 tasks from professionals averaging 14 years of experienceDirect for held-out, expert-built tests

Why headline training costs hide most data work

Because the headline figure usually counts one final training run, while the work that decides which data to collect, buy and keep happens in experiments before it. An investor who reads a lab's data appetite from its training-cost headline is reading the smallest part of the bill.

DeepSeek's V3 repository says full training took 2.788 million H800 GPU hours. Its technical report adds that the stated costs include only the official training of DeepSeek-V3, excluding the costs of prior research and ablation experiments on architectures, algorithms or data. The data experiments are named and then left out.

Third-party estimates suggest the excluded part is the larger one. Epoch AI estimates that OpenAI spent about $5 billion on R&D compute in 2024 and that only about $500 million of it, roughly 10 percent, went to the final training runs behind released models, with the rest spent on experiments, derisking runs, basic research and unreleased models. Epoch builds the estimate partly on press reporting, so treat it as an order of magnitude, not an audited split.

Curation follows the same pattern. Meta's Llama 3.1 card says the company built many large language model classifiers to select high-quality prompts and responses, which means a lab's own models do filtering work that could otherwise be bought. For a supplier, the implication is practical: the buyer runs ablations on your sample before it signs, and the experiment, not the announcement, decides the purchase.

The demand-signal ladder: from announcement to repeat purchase

Rank every piece of evidence about lab data spending on six rungs, from a capability announcement at the bottom to a repeat purchase at the top. Only the top two rungs show money committed; the four below show intent, use or volume.

At a glance

The demand-signal ladder: from announcement to repeat purchase

The ladder exists because the strongest alternative explanation is real: capability progress and partnership news do not have to turn into commercial data spending. A lab can improve a model with its own synthetic data, more compute or better experiments and pay no outside supplier at all. Each rung therefore asks what the evidence proves and stops there.

  • Rung 1, capability news: a model release or benchmark gain. It shows a lab is working, not that it is buying.
  • Rung 2, stated need: a public call for data, such as OpenAI's request for large-scale datasets not already easily accessible online. It names demand without a price or a volume.
  • Rung 3, disclosed use: a model or system card names a data category, such as data-labeling services or paid contractors. It confirms the line exists at that lab.
  • Rung 4, disclosed volume: counts of examples, tasks or tokens, such as the over 10 million human-annotated examples in Llama 3's fine-tuning data. It sizes the work, not the spend.
  • Rung 5, paid commitment: a contract value and term on the record, usually in the seller's filing, such as the data licensing arrangements Reddit disclosed in its S-1.
  • Rung 6, repeat purchase: a renewal, refresh or expansion on stated terms. It is the only rung that shows a budget line recurs rather than spiking once.

The procurement evidence matrix: how high each line climbs

On public evidence, licensing is the only line with a paid commitment, annotation, evaluation and collection reach disclosed volume, and no line shows a repeat purchase. That empty top rung, not any headline, is where diligence on a company selling into lab budgets should start.

Licensing reaches rung 5 only because one seller is public. In its S-1, Reddit put the aggregate contract value of the data licensing arrangements it signed in January 2024 at $203.0 million, with terms running two to three years. Its annual report for 2024 adds that "to date, substantially all of the contract value associated with our licensing revenue is derived from two of our partners," and warns that those arrangements might lapse at renewal or come back on less favorable terms. The labs on the other side rarely disclose what they pay.

Disclosure law will move more lines to rung 3, not to rung 5. California's AB 2013, signed in September 2024, requires developers of generative AI systems to post a high-level summary of their training datasets that states, among other things, whether the datasets were purchased or licensed. It asks for a yes or no, not an amount, so investors will learn which lines exist at a lab sooner than they learn their size.

The figure places each cost line on its highest public rung as of the check date. The fourth column is the private evidence a company selling into that line should be able to show an investor under confidentiality.

The procurement evidence matrix: how high each line climbs
Cost lineHighest public rungWho discloses itEvidence to request in diligence
Licensing5, paid commitment (Reddit S-1, 2024)Sellers' filings, rarely labsSigned licenses with term, renewal clause and the top buyer's share of value
Annotation4, disclosed volume (Llama 3 card, 2024)Model cardsPurchase orders by quarter, split into human and synthetic examples
Evaluation4, disclosed volume (GDPval task count)Lab research papersPaid evaluation projects, refresh cadence and who owns the tasks
Collection4, disclosed volume (pretraining token counts)Model cardsPurchased datasets listed apart from crawled data
Curation3, disclosed use (classifier filtering)Model cardsHours or compute spent cleaning outside data before use
Experiments3, disclosed use (DeepSeek-V3 report)Technical reportsAblations that changed a purchase decision, with dates
A six-rung ladder from capability news to repeat purchase, with licensing at the paid-commitment rung, annotation, evaluation and collection at disclosed volume, curation and experiments at disclosed use, and the repeat-purchase rung empty.
Highest public evidence for each cost line, checked 2026-10-26. Placements reflect public disclosures, not spending.

How to interview budget owners without collecting hype

Interview the people who sign data purchase orders, place every answer on the ladder by dated cohort, and count requests, evaluations and paid commitments separately. SourceX has not run or published this study; what follows is the design, so an investor or a company can run it the same way and compare results.

At a glance

How to interview budget owners without collecting hype

Results are not yet published. Until they are, treat any claim about how lab data budgets split, including the six-line structure above, as a hypothesis with a stated test rather than a finding.

Checklist · 7 items

  1. Question: for each of the six cost lines, what did the team commit in writing in each quarter, to whom, and against what acceptance test?
  2. Population: independent budget owners at model developers and at companies that fine-tune models, such as data leads, post-training leads, evaluation leads and procurement staff. Exclude anyone with a commercial relationship to the interviewer, including SourceX counterparties.
  3. Sampling frame: a list of organizations fixed before outreach, with refusals and non-responses reported beside completed interviews.
  4. Variables: cost line, rung reached, quarter of first request, internal or external spend, and whether a synthetic or in-house alternative was tested first.
  5. Denominators: requests issued, then evaluations run on those requests, then paid commitments from those evaluations, each counted within its quarterly cohort so that later conversions are not credited to earlier quarters.
  6. Uncertainty: with fewer than 30 respondents, report counts and ranges rather than percentages, and attribute every quote to a role and a date, never to a named person or employer without written consent.
  7. Challenge case: report at least one team whose outside annotation spend fell after synthetic generation took over, next to the headline result.

0 of 7 ticked. Tick each item as you go.

Could labs improve models without buying more outside data?

Yes, and the public evidence for that path is as strong as the evidence for outside spending. Synthetic generation, distillation and engineering efficiency can all raise capability while the outside lines stay flat or shrink.

Meta's cards show the swap inside one model family. Llama 3's fine-tuning data included over 10 million human-annotated examples; the Llama 3.1 card, dated July 23, 2024, describes public instruction datasets plus over 25 million synthetically generated examples, and says the model collection supports using its outputs to improve other models through synthetic data generation and distillation. The same card's safety section still describes combining human-generated data from vendors with synthetic data, so the outside line narrowed rather than vanished.

Efficiency is the other substitute. DeepSeek's repository attributes V3's cost-effective training to its Multi-head Latent Attention and DeepSeekMoE architectures, a reminder that a capability jump can come from engineering rather than purchased inputs.

The test that separates the two explanations is easy to state and hard to fake. If outside data spending is growing, rung 5 and rung 6 evidence should multiply across several sellers and several buyers. If capability is coming mainly from internal substitutes, model cards will keep reporting larger synthetic counts while seller filings show concentration and renewal risk.

What could make the outside-spend thesis wrong

Four developments would shrink the outside lines even while total AI spending grows. A company selling into lab data budgets should be able to say which one it is most exposed to, and investors should ask before valuing the market.

  • Synthetic substitution spreads. Meta moved from counting human-annotated examples for Llama 3 to counting over 25 million synthetic ones for Llama 3.1; if evaluation follows, the expert-task line shrinks too.
  • Experiments stay internal. If most data work happens in a lab's own ablations, which Epoch AI's estimate of OpenAI's 2024 compute suggests, outside suppliers compete for a thin slice of the budget.
  • Buyers stay few. Reddit's annual report for 2024 shows a pair of partners accounting for substantially all of its licensing contract value, and a supplier facing a handful of buyers carries renewal and pricing risk whatever the size of the market.
  • Disclosure arrives without dollars. AB 2013 makes purchased or licensed status public, not amounts, so better transparency may still leave outside spending unverifiable from the outside.

Questions to ask any company selling into lab data budgets, SourceX included

Ask which cost line a company sells into and which rung its evidence reaches, using your definitions rather than its own. The same questions apply to SourceX.

SourceX sells into the licensing line. It helps companies turn the business records they choose into datasets licensed for AI, running the chain from sourcing and qualifying supply through rights review and contracting to data preparation, delivery and payment. It keeps no inventory of datasets; each one is sourced from U.S. businesses after an AI buyer asks for it.

No buyer, deal or commercial metrics for SourceX are public; it hands company and commercial metrics to investors directly. The company's raise, under Rule 506(c), is limited to accredited investors, and a request through the investors page brings materials and a channel for any of the questions below.

Checklist · 6 items

  1. Which of the six cost lines do your buyers pay from, and who inside the buyer signs?
  2. Over the last four quarters, how many written requests, paid evaluations and signed commitments did you receive, counted within each cohort?
  3. What share of committed value came from your largest buyer?
  4. Which requests were lost to synthetic data or to a buyer's internal team, and how do you know?
  5. Has any buyer renewed or expanded a purchase, and on what terms relative to the first?
  6. Which disclosure rule or model release would most change demand for the line you sell into?

0 of 6 ticked. Tick each item as you go.

Methods note: what the six-line lab budget breakdown rests on

The six-line breakdown, the demand-signal ladder, the procurement evidence matrix and the interview design were drawn up for this article on 2026-10-26 as editorial frameworks. The evidence behind them comes from Meta's Llama 3 and Llama 3.1 model cards and DeepSeek's V3 repository, read directly on GitHub, and from DeepSeek's V3 technical report, two Epoch AI analyses, OpenAI's data partnerships page and GDPval paper, Anthropic's Claude Opus 4.5 system card, Reddit's S-1 and annual report for 2024, and California's AB 2013, each checked that day.

Not measured: any lab's spending on any line, the human-to-synthetic split at labs other than in Meta's disclosures, prices, and SourceX's own demand. Epoch AI's figures are third-party estimates built partly on press reporting. Rung placements show the highest public evidence found, not the most a lab has actually done in private.

Frequently asked questions

How much do AI labs spend on data?

No lab publishes a figure, so any total quoted without a source is a guess. Public evidence gives pieces: Reddit disclosed data licensing arrangements with a $203.0 million aggregate contract value and two-to-three-year terms in its 2024 S-1, and Epoch AI estimates OpenAI spent about $5 billion on R&D compute in 2024. Neither is a lab data budget, and the second is mostly compute rather than data purchases.

Which parts of an AI lab's data budget go to outside companies?

Licensing and annotation pay outsiders most directly: rights holders for non-public content, and labeling vendors and paid contractors for examples and expert judgments. Evaluation pays outside experts for some test construction and grading, and collection occasionally buys existing datasets. Curation and experiments mostly buy a lab's own compute and staff time.

Is synthetic data replacing paid human annotation?

Partly, at least in Meta's disclosures. Llama 3's fine-tuning data included over 10 million human-annotated examples, while the Llama 3.1 card describes over 25 million synthetic examples and still mentions human data from vendors in safety work. How far the swap goes depends on whether a model can generate and check the examples a task needs.

What evidence shows a lab is actually paying for data?

A contract value and term on the record, a signed purchase order, or a renewal on stated terms. Model releases, benchmark gains and public calls for data show intent or use, not payment. On this article's ladder, only rungs 5 and 6 show money committed.

Does SourceX publish data on AI lab demand?

No. Its buyer, deal and commercial figures are unpublished, and company and commercial metrics are given to investors directly. Materials can be requested through sourcex.si/investors; SourceX makes any offering solely to verified accredited investors, through its offering documents.

Sources

  • Meta's Llama 3 model card (release April 18, 2024) reports pretraining on over 15 trillion tokens from publicly available sources and fine-tuning data that includes over 10 million human-annotated examples. Source
  • Meta's Llama 3.1 model card (July 23, 2024) reports over 25 million synthetically generated fine-tuning examples, LLM-based classifiers for selecting prompts and responses, human-generated data from vendors combined with synthetic data, and support for synthetic data generation and distillation. Source
  • DeepSeek's V3 repository states that full training required 2.788 million H800 GPU hours and attributes cost-effective training to its MLA and DeepSeekMoE architectures. Source
  • The DeepSeek-V3 technical report says its stated costs include only the official training run, excluding prior research and ablation experiments on architectures, algorithms or data. Source
  • Epoch AI's cost study attributes 47-67 percent of frontier model development cost to hardware, 29-49 percent to R&D staff including equity, and 2-6 percent to energy. Source
  • Epoch AI estimates OpenAI spent about $5 billion on R&D compute in 2024, of which about $500 million, roughly 10 percent, went to final training runs of released models, based partly on press reports. Source
  • OpenAI's data partnerships page seeks large-scale datasets that are not already easily accessible online to the public. Source
  • OpenAI's GDPval paper describes 1,320 tasks across 44 occupations, constructed from the representative work of industry professionals with an average of 14 years of experience. Source
  • Anthropic's Claude Opus 4.5 system card lists non-public data from third parties and data provided by data-labeling services and paid contractors among Claude's training data sources. Source
  • 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 annual report for 2024 says substantially all of the contract value associated with its licensing revenue comes from two partners and that the arrangements may not be renewed or may be renewed on less favorable terms. Source
  • California's AB 2013, signed in September 2024, requires developers of generative AI systems to post a high-level summary of training datasets, including whether the datasets were purchased or licensed. 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