Deal economics
Real data's synthetic seed value: why fewer records can be worth more
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Real data synthetic seed value comes from steering, not volume. NVIDIA's Nemotron-4 340B report (2024) describes about 20,000 human-annotated examples behind an alignment set more than 98 percent synthetic. A seed record can shape thousands of outputs, so its value depends on measured uplift over public or synthetic seeds and on whether the license prices generation. A per-record price captures neither.
Value drivers
Real data synthetic seed
Key takeaways
- A seed is worth the uplift it causes in generated data times the uses the license permits, not its value as one more corpus row.
- Generated volume is an operator setting, so seed reuse is easy to inflate; only uplift on untouched real tasks shows value.
- Public and synthetic seeds already handle topic scaffolds and style exemplars; procedural grounding, outcomes and judge calibration are where real records keep pricing power.
- A license that grants generation, output training, retention and reuse for a per-record fee has sold the multiplier at corpus prices.
- Seed use shrinks order sizes, so a real-data business needs pricing tied to uplift and generation rights rather than record counts.
What is real data worth when it seeds a generator instead of filling a corpus?
As a seed, a real record is worth the measured improvement it causes in generated data, multiplied by the uses the license allows. That is a different quantity from its worth as one more row in a training corpus. A corpus record is consumed once inside a mix; a seed record steers every output sampled from it, so its value scales with reuse and depends on what public or synthetic seeds could have done instead.
The distinction matters to anyone underwriting a real-data business, because seed use cuts both ways. It can shrink the number of records a buyer needs, which hurts suppliers paid by volume, while raising what each well-chosen record is worth, which helps suppliers who can show the uplift and price the generation rights. A synthetic data investment thesis that treats real data as either obsolete or irreplaceable skips this middle case, and the middle case is where contract value gets decided.
Settling the question takes two kinds of evidence. One is measured: quality gained on untouched real tasks per dollar spent acquiring seeds, with the generator and compute held fixed. The other is contractual: which generation uses the license permits, because a multiplier the supplier never priced belongs to the buyer. This article designs the measurement and maps the clauses; it reports no results of its own.
How one seed row turns into many generated records
In current pipelines, the number of generated records is a setting chosen by the operator, not a property of the seed file. The reference for NVIDIA's Nemotron synthetic data pipeline lists num_records as a required value and says that, when a seed dataset is supplied, the tool samples one seed row for each generated record and exposes that row's fields to the prompts. Ask for 10,000 records from a 500-row seed file and each row is drawn about 20 times on average.
Variety is multiplied on top of that reuse. The same documentation's customer-support tool-calling example gives every seed row five anchor fields, such as the issue and a policy hint, then adds two category columns, urgency and channel, so the generator crosses each seed with combinations of those values without anyone writing extra seed rows.
The pattern holds at frontier scale. NVIDIA's Nemotron training FAQ says the pre-training corpus includes synthetic counterparts of its higher-quality web crawl tiers, produced by rephrasing filtered web documents, so public pages serve as seeds for synthetic text. Microsoft's AgentInstruct paper (July 2024) went further, using only raw text documents and code files as seeds for a post-training set its abstract puts at 25 million pairs; the body describes about 22 million instructions.
Reuse alone proves nothing about value, because the operator can set it to any number. What a seed contributes is information the generator lacks. NVIDIA's planning guide for the same pipeline puts it plainly: the seed file and prompts are where domain knowledge enters, and without that knowledge a run produces generic output drawn from the model's own training data, however carefully its columns are specified.
Fewer records, more value per record: the seed paradox
Seed use concentrates value in a small number of records, which favors suppliers who can prove quality and squeezes suppliers who sell bulk. The clearest public example is NVIDIA's Nemotron-4 340B technical report (June 2024). Its alignment relied on roughly 20,000 human-annotated examples, about 10,000 for supervised fine-tuning and about 10,000 from the HelpSteer2 dataset for reward modeling and preference tuning, while its generation pipeline produced more than 98 percent of the fine-tuning and preference data.
Read that as a procurement fact. A frontier alignment run needed a five-digit count of human examples and generated the rest. A supplier pricing by the record would have filled a small order; a supplier whose examples set the reward model's standards, or anchored the generator in a real domain, would have supplied the part that the synthetic 98 percent depended on.
The paradox gives investors a sharper question than whether synthetic data replaces real data. Ask how a company's revenue behaves when buyers need far fewer records, each doing far more work. If contract value falls with record count, seed use is a headwind. If contract value follows measured uplift and priced generation rights, it is a tailwind, and the company should be able to show which.
The seed substitution matrix: which seed jobs public or synthetic inputs can do
A seed does one of six jobs, and public or synthetic inputs already do the first two well for most tasks. The matrix orders the jobs from those where public or synthetic seeds suffice to those where real records are required, and rates each on the license term a real seed needs and the pricing power that follows.
Held-out evaluation is deliberately missing from the list. A record used to seed generation should never also judge the result, so evaluation sets are bought and kept separately, and a supplier who offers the same records for both should be asked how it prevents leakage.
Read across each row before reading down; the investor case lives in the bottom four rows:
| Seed job | What the seed supplies | Public or synthetic substitute? | License term a real seed needs | Pricing power |
|---|---|---|---|---|
| Topic scaffold | The subjects, intents and task types to generate around | Usually yes: AgentInstruct drew its seeds from public text and code | Generation from seeds | Low |
| Style exemplar | What a ticket, invoice note or shift handoff looks like on the page | Often yes: templates and category columns add variety cheaply | Generation from seeds | Low |
| Judge calibration | Human preference or quality labels that train a reward model or grader | Partly: Nemotron-4 trained its reward model on about 10,000 human examples | Training judges, reuse in later models | High |
| Procedural grounding | Real steps, codes, exception paths and the order work happens in | Rarely: without domain input a generator falls back on generic output | Generation, plus keeping the outputs | Medium to high |
| Outcome truth | What happened after the decision: a refund, a callback, a write-off | No: outcomes are observed, not written | Training on outputs, reuse across model versions | High |
| Anchor in the mix | Real data kept beside synthetic generations to limit drift | No by definition; one 2024 study found accumulation avoided collapse | Direct training, retention after the term | Medium |
Permitted generation uses decide who keeps the multiplier
A supplier captures seed value only through the uses its license prices; whatever the license allows for free, or leaves unaddressed, accrues to the buyer. Training rights and generation rights are separate questions, and a grant covering one does not settle the other. The companion piece on synthetic derivatives of licensed records works through those clauses for a supplier; here the question is what each one is worth.
Consider a license that permits generation, training on the outputs, retention after the term and reuse in later models, all for a one-time per-record fee. It has sold the multiplier at the price of the corpus. That can be a sensible trade for a small, low-risk deal, but it should be a deliberate choice that shows up in contract value, not a default.
Use the comparison to read any data company's standard terms, SourceX's included:
| Generation use | What it allows | Risk if the license is silent | What a supplier can price |
|---|---|---|---|
| Generate from seeds | Prompting a generator with licensed records | Disputed coverage under a training-only grant | A generation fee or an output-volume band |
| Train on generated outputs | Training on synthetic records derived from the seeds | The buyer assumes yes once outputs exist | Bundled with generation, or priced per model |
| Keep outputs after the term | Retaining synthetic derivatives when the license ends | Outputs outlive the seeds and weaken renewal | A survival fee, or deletion with a buyout option |
| Reuse in later model generations | Regenerating or retraining on the same seeds for new models | Unclear whether a refresh counts as a new use | Per-generation or term-based refresh rights |
| Send seeds to a hosted generator | Passing records to a third-party model provider | A processor the supplier never approved | A named-processor list with audit rights |
| Share or sell outputs | Passing synthetic derivatives to other parties | Seed content can travel inside the outputs | Excluded, or priced as a sublicense |
A seed-only experiment that prices uplift per acquisition dollar
The experiment answers one question: how much better does generated data get when real records seed the generator, compared with the best cheaper seed, per dollar spent acquiring those records? Everything except the seed stays fixed: generator model and version, prompt templates, category columns, sampling settings, number of generated records, training recipe and training compute.
The defining arm keeps real records out of training entirely. In arm D, licensed records only prompt the generator and the student model never sees them. Comparing D with arm F, where the same records are trained on directly, separates seed value from corpus value, and arm E, which does both, shows whether the two add up or overlap. Record, for every arm, which generation uses its seed source permits, so the scores can be read next to the clause comparison above.
Evaluation uses real tasks the generator never saw, drawn from different companies and a later time window than the seeds, and graded on outcomes rather than on what a model claims. Anthropic's engineering guidance on agent evaluations illustrates the difference with a booking agent whose transcript announces success while the real outcome is whether a reservation exists in the environment's SQL database. Run several trials per task, report intervals across repeated training runs, and treat a gain smaller than the run-to-run spread as no gain.
The headline metric is uplift per seed-acquisition dollar: the score gain of arm D over the best of arms A to C, divided by what the seeds cost to acquire, with rights review and preparation included. Running D at three seed counts shows where the curve flattens. NVIDIA's FAQ says it validates synthetic data by training models on it in small-scale ablations, generally under 10 billion parameters, which keeps a design like this affordable. SourceX has not run this experiment and publishes no results from it; the arms are the design:
| Arm | Seeds given to the generator | Real records in training? | What it isolates |
|---|---|---|---|
| A. No seed | None: category columns and prompts only | No | What the generator knows on its own |
| B. Public seeds | Public documents matched to the task domain | No | What free grounding achieves |
| C. Synthetic seeds | Seed rows written by a second model | No | Whether seeds need to be real at all |
| D. Licensed seeds at three counts | Real business records at a small, medium and large count | No | Seed value and its diminishing returns |
| E. Seed plus corpus | The same licensed records | Yes, beside the generated set | Whether seed and corpus value add up |
| F. Corpus only | None | Yes, the licensed records alone | What the records are worth without generation |
The strongest objection: public and synthetic seeds may be enough
For many general skills, they already are. AgentInstruct's seeds came from public material, including KnowledgePile, AutoMathText, part of OpenStax and Apache-2.0 licensed source code, and the abstract reports that the resulting Orca-3, a fine-tuned Mistral-7B, outperformed models such as Llama-8B-instruct and GPT-3.5-turbo on the benchmarks it used. Nothing in that recipe needed a private record.
The public pool is large, too. Epoch AI's 2024 study put the effective stock of quality public human text at about 300 trillion tokens, with full use projected at some point between 2026 and 2032 on current trends. A buyer looking for topic scaffolds or style exemplars has plenty of public seed material and every reason to exhaust it before paying for private records.
The objection weakens where a seed must carry something no public document records: how a particular kind of business resolves a dispute, which exceptions recur, and what happened afterward. Those are the procedural, outcome and calibration rows of the matrix. The decision rule follows: count a company's seed revenue only where its records do one of those jobs, and treat scaffolds and exemplars as a commodity.
What could make the seed thesis wrong
Five developments would cut the value of real seeds, and each can be tracked. A company that cannot say which one it watches most closely has not stress-tested its own story.
- Seed orders could stay small. If alignment pipelines keep needing only five-digit counts of human examples, as Nemotron-4 340B did, seed deals stay modest even when each record is valuable, and revenue depends on pricing generation rights rather than volume.
- Outputs could remove the need to renew. Once a buyer holds a large synthetic set derived from licensed seeds, its reason to renew shrinks unless the license limits retention or reuse in later models.
- Generators could absorb the domain. As models train on more rephrased and synthetic text, the knowledge a seed supplies may already sit inside the generator, closing the gap between arm D and arms A to C.
- The anchor evidence is narrower than it sounds. Gerstgrasser and colleagues (2024) found that keeping the original real data alongside successive synthetic generations avoided model collapse in their experiments, but their formal result covers a sequence of linear models, and nothing in it shows that private business records are the anchor required.
- Demand signals are plans, not purchases. micro1 announced in October 2026 a commitment to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data, which it says it uses to build reinforcement learning environments. That is a stated plan, not money spent, and it says nothing about whose records will be bought.
Questions to put to any real-data supplier about seed use, SourceX included
Ask where a company's records sit in the substitution matrix and how its terms price generation, then hold SourceX to the same standard. SourceX works on the supply side: a business picks the records it is willing to license, and SourceX carries them to AI developers by qualifying the source, reviewing rights, negotiating the contract, preparing and delivering the data and handling payment. The supplying company signs off on scope and terms and executes the agreement before any data reaches SourceX, which can then use, sublicense and distribute the de-identified dataset within what that agreement allows.
No seed-uplift results, generation terms or contract figures from SourceX are public, and none appear here. The company discusses company and commercial metrics with investors directly, and its current raise is open only to accredited investors under Rule 506(c). Bring these questions to that conversation, or to any other supplier:
- Which matrix rows do your delivered datasets serve, and how do you learn whether a buyer used them as seeds or as a corpus?
- Do your standard terms separate generation, training on outputs, retention after the term and reuse in later models, and are any of them priced separately?
- Have you measured uplift per seed-acquisition dollar against public and synthetic seeds with the generator held fixed, and on which held-out tasks?
- How does contract value change when a buyer needs fewer records for the same task?
- Which seeds have passed through a third-party generator, and who approved that processor?
- What happens to synthetic derivatives when a license ends?
Methods note: how the seed matrix and experiment were built
The substitution matrix, clause comparison and experiment design were written for this article on 2026-10-27 as editorial tools. They draw on NVIDIA's Nemotron synthetic data documentation and training FAQ (read from the public repository on 2026-10-27), the Nemotron-4 340B technical report (2024), Microsoft's AgentInstruct paper (2024), Gerstgrasser and colleagues on accumulating real and synthetic data (2024), Epoch AI's 2024 data-stock study, Anthropic's guidance on agent evaluations (November 2025) and micro1's October 2026 announcement.
Not measured: uplift per seed-acquisition dollar for any real dataset, how often buyers use licensed records as seeds rather than training data, the generation terms in any actual license and the prices attached to them. The pricing-power column is a judgment from the evidence cited, not a market measurement, and the 10,000-from-500 reuse figure is arithmetic on an example, not a reported pipeline.
Frequently asked questions
What is seed data in synthetic data generation?
Seed data is the set of real or hand-written examples a generator conditions on while producing synthetic records. In NVIDIA's Nemotron pipeline the tool samples one seed row per generated record, so a small seed file can steer a much larger output. The seeds carry the domain knowledge the generator would otherwise lack.
Is real data more valuable as a seed or as training data?
It depends on the job the record does. Topic scaffolds and style exemplars are easy to replace with public or synthetic seeds, while procedural grounding, observed outcomes and human judgment labels are not. The only reliable answer for a given task comes from comparing a seed-only arm with a corpus-only arm at a fixed generator and compute budget.
How many real examples does a synthetic pipeline need?
There is no general count. NVIDIA's Nemotron-4 340B report describes roughly 20,000 human-annotated examples behind an alignment set that was more than 98 percent synthetic, but that is one frontier pipeline, not a rule. Measure it for your task by running licensed seeds at several counts and watching where gains flatten.
Can a buyer generate synthetic data from licensed records without asking?
Only if the license grants it. A right to train on records does not automatically cover prompting a generator with them, keeping the outputs or reusing them in later models, so each use should be named and priced. This is general information, not legal advice; counsel should review the specific agreement.
Does SourceX publish seed-uplift or licensing figures?
No. SourceX has published no seed-uplift results, generation terms or contract figures, and it shares company and commercial metrics with investors directly. Investors can request materials through sourcex.si/investors, and any offering is made only to verified accredited investors.
Sources
- NVIDIA's Nemotron synthetic data configuration reference lists num_records as required and says that, with a seed dataset, the tool samples one seed row per generated record and exposes its fields to prompts. Source
- NVIDIA's tool-calling data guide gives each seed row five anchor fields and adds urgency and channel category columns to add variety without more seed rows. Source
- NVIDIA's planning guide says seed files and prompts are where domain knowledge enters, and a run without it produces generic output from the model's own training data. Source
- NVIDIA's Nemotron training FAQ says the pre-training corpus includes synthetic counterparts of higher-quality crawl tiers made by rephrasing filtered web documents, and that synthetic data is validated through small-scale ablations, generally on models under 10B parameters. Source
- NVIDIA's Nemotron-4 340B technical report (June 2024) says alignment relied on about 20,000 human-annotated examples (10,000 for fine-tuning, 10,000 from HelpSteer2) while over 98 percent of fine-tuning and preference data was synthetic. Source
- Microsoft's AgentInstruct paper (July 2024) uses only raw text documents and code files as seeds, reports a 25-million-pair post-training set (about 22 million instructions in the body) and says Orca-3, a fine-tuned Mistral-7B, outperformed Llama-8B-instruct and GPT-3.5-turbo. Source
- Gerstgrasser and colleagues (2024) found that accumulating synthetic generations alongside the original real data avoided model collapse in their experiments, with a formal bound proven for a sequence of linear models. Source
- Epoch AI (2024) estimated the effective stock of quality public human text at about 300 trillion tokens, fully used between 2026 and 2032 on trend. Source
- Anthropic's engineering guidance on agent evaluations (November 2025) runs multiple trials per task and distinguishes a transcript's claim from the outcome, using a flight-booking agent whose real outcome is a reservation in the environment's SQL database. Source
- micro1 announced in October 2026 a commitment to spend $1 billion over 12 months acquiring and licensing enterprise operational data, which it says it uses to build reinforcement learning environments; a stated plan, not spending. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.