Deal economics
Cheaper synthetic generation won't zero out data pricing: a ceiling test
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Cheaper synthetic generation lowers data pricing only by the generation share of a buyer's cheapest substitute route. Stanford's AI Index 2025 found the price of GPT-3.5-level output fell more than 280-fold in about 18 months, but seeds, checks and access to buyers did not fall with it. After a drop that large, a dataset's price ceiling falls by roughly that share, and value migrates to scarce inputs, validation and distribution.
Value drivers
Cheaper synthetic generation
Key takeaways
- A buyer's price ceiling for a dataset is the cost of its cheapest other route to the same acceptance result, and generation is only one line in that route.
- When generation gets k times cheaper, the ceiling falls by the generation share times (1 minus 1/k); after a 280-fold drop that is almost exactly the generation share.
- Seeds, validation and access to the buyer's pipeline did not get cheaper, so the value leaving the generation line moves to them.
- Price dispersion should widen: generic instruction data heads toward generation cost, while outcome-linked records, sealed tests and fresh periods hold their ceiling.
- A matched-budget mixture experiment repriced at lower generation costs would settle the question; SourceX has published no such results.
What happens to data pricing when generation gets cheaper?
Data prices fall by less than a straight-line forecast from token prices suggests, and unevenly. A buyer will pay for a dataset up to the cost of its cheapest other route to the same result on its own acceptance test; call that the ceiling. Synthetic generation is one line in that route, beside the seed material it needs, the checking that makes its output usable and the work of getting it into the buyer's pipeline on terms the buyer can accept.
So the ceiling moves only with generation's share of the route. If generation is a share g of the substitute's cost and becomes k times cheaper, the ceiling falls by g times (1 minus 1/k). For a 280-fold drop the bracket equals 0.996, so the ceiling falls by almost exactly the generation share and then stops. A dataset whose best substitute is mostly generation loses most of its price; one whose substitute is mostly seeds, checks and access loses little.
That is the thesis this article tests. Cheaper generation does not push data prices to zero; it moves value off the generation line and onto three others, scarce inputs, validation and distribution. Which of the three captures it depends on the data product, and the substitution matrix sorts six of them.
The collapse forecast, and the arithmetic it leaves out
The forecast that cheap generation will collapse data prices gets the trend right. Stanford's AI Index 2025 reports that the price of a model scoring 64.8 on MMLU, its marker for GPT-3.5-level performance, fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold reduction in about 18 months.
At a glance
The collapse forecast, and the arithmetic it leaves out
The same chapter cites Epoch AI's finding that, depending on the task, inference prices for a fixed level of performance have fallen anywhere from 9 to 900 times per year. Generation costs are not falling at one rate, so routes that lean on the slower-falling tasks keep more of their ceiling for longer, and the spread between data products widens even as the average falls.
What the forecast leaves out is the rest of the route. Three illustrative routes make the arithmetic concrete; the generation shares are assumptions chosen to show the rule, not measurements:
- Route A, generic instruction pairs: generation is 70 percent of the substitute's cost. After a 280-fold drop the ceiling falls by about 69.8 percent.
- Route B, labeled operational records: generation is 30 percent, with checking and seed licensing making up the rest. The ceiling falls by about 29.9 percent.
- Route C, outcome-linked case histories: generation is 10 percent, because the expensive part is records of what actually happened. The ceiling falls by just under 10 percent.
Scarce inputs: what a generator still has to be fed
The first place value moves is the material a generator cannot make for itself: seeds that steer it, real records that anchor it and facts about events that happened after a model was trained. Cheaper generation raises demand for these, because each extra generated example is only as good as the inputs it was conditioned on.
Microsoft's Phi-4 technical report (December 2024) shows the pattern in a disclosed recipe. Its synthetic data ran to about 400 billion unweighted tokens across 50 broad types of synthetic dataset, yet the authors describe curating and filtering organic sources, including web content, licensed books and code repositories, to extract the seeds for that pipeline. The generator supplied the volume; the seeds set what it could produce.
Research on repeated training points the same way. Gerstgrasser and colleagues (2024) found that replacing real data with each generation's synthetic output tends toward model collapse, while accumulating synthetic data alongside the original real data avoids it. If that holds at scale, real data keeps a standing role as the anchor of a mixture, which is a complement relationship rather than substitution.
At least one buyer now prices scarce inputs in public. In October 2026 micro1 announced a commitment to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data to build reinforcement learning environments. That is a stated plan, not money spent, and it says nothing about whose data will be bought; it does show at least one buyer putting a number on inputs that generation cannot fake.
Validation: the gate cheap generation cannot open by itself
The second destination is checking. A generated example is worth paying for only after something has confirmed it is right, and the cost of that confirmation does not fall with token prices unless the check itself is automatic.
Where an automatic check exists, generation goes a long way. The Tülu 3 post-training report from the Allen Institute for AI (2024) introduced reinforcement learning with verifiable rewards, a method that only gets a reward when the model's completions are verified to be correct, and aimed it at skills with checkable answers such as mathematics and precise instruction following. For those skills, cheap generation plus a cheap verifier can stand in for bought data.
Most business tasks have no such verifier. Whether a refund was appropriate, a dispatch plan sound or a contract clause acceptable is settled by what happened next or by an expert, so real outcomes and sealed test sets become the reference every check leans on. Value moves to whoever owns that reference: the author of the verifier, the holder of the outcome records or the custodian of the held-out set.
Distribution: who controls the route into a buyer's pipeline
The third destination is distribution: control over how data reaches a buyer and on what terms. When the data itself is cheap to produce, the channel and its terms claim a larger share of what is left.
Data marketplaces show the mechanics. AWS documentation says every AWS Data Exchange product is subscription-based, with prices and durations set in the provider's offer terms, and that once a subscription expires the subscriber can no longer view or export the data sets. AWS Marketplace charges providers tiered fulfillment fees on revenue it collects for new subscriptions. Price attaches to access over time and to the channel, not to a count of records.
Generator licenses are a quieter lever. Meta's Llama 3.1 community license requires anyone who uses Llama outputs to create, train or improve an AI model that is distributed to put Llama at the beginning of that model's name, and it requires licensees whose products had more than 700 million monthly active users when Llama 3.1 was released to request a separate license, which Meta may grant at its discretion. Synthetic data can carry obligations from the model that wrote it, so whoever writes those terms holds part of the value of the datasets built downstream.
For a data seller this layer cuts both ways. A seller with direct buyer relationships, and terms buyers accept, captures it; a seller that reaches buyers only through someone else's channel pays it away.
The substitution matrix: six data products after a steep fall in generation cost
Sort data products by how much of their best substitute is generation, and the price outlook follows. The matrix rates six products a business-data seller could offer. The generation share is an editorial band, not a measurement, and the last column splits six points of remaining value across scarce inputs, validation and distribution; the triangle chart plots those splits.
Ratings are editorial judgments made on 27 October 2026 from the evidence cited in this article:
| Data product | Best synthetic route today | Generation share of that route | Ceiling after a steep drop | Remaining value, scarce / validation / distribution (of 6) |
|---|---|---|---|---|
| Generic instruction and chat pairs | Substitutes: generate, then filter | High | Falls toward generation cost | 0 / 1 / 5 |
| Domain reference text: manuals, policies, product documents | Partial: rewrite from public or licensed seeds | Medium | Falls substantially | 3 / 0 / 3 |
| Labeled operational records: tickets, invoices, work orders with codes | Partial: generate, then check the labels | Medium to low | Falls modestly | 2 / 3 / 1 |
| Outcome-linked case histories | Complement: real records seed and anchor | Low | Holds | 4 / 1 / 1 |
| Sealed evaluation sets from real work | None by design: synthetic items cannot certify themselves | Very low | Holds, or rises as more models need testing | 1 / 5 / 0 |
| Fresh-period feeds: new months of records | Complement: no generator knows next quarter | Low | Holds while freshness matters | 4 / 0 / 2 |
A controlled-mixture plan that measures the ceiling directly
The ceiling can be measured instead of argued, with a matched-budget experiment that logs every cost line and scores every arm on real cases none of them has seen. Run it per data product, because the answer differs from row to row of the matrix.
Hold the total budget fixed across arms, counting generation, checking, seed licensing and integration work, and score every arm on the same untouched evaluation set drawn from the target task. The output is a parity cost: the lowest-cost arm without bought task records whose score cannot be distinguished statistically from the real-only arm. Then reprice the logged generation lines at one-tenth and one-hundredth of their cost, without retraining, and recompute parity. The change in parity cost is the measured generation sensitivity of that product's ceiling.
SourceX has published no results from this design; it is a protocol for an investor to run, or to ask a seller to run, and its arms are:
| Arm | Training mix at matched budget | Costs logged | What it isolates |
|---|---|---|---|
| R: real only | Licensed task records | Licensing, preparation, integration | The score every other arm must match |
| S: synthetic from public seeds | Generated from public material, then filtered | Generation, checking, filtering | Whether generation alone reaches parity |
| S+: synthetic from licensed seeds | Generated from a small licensed seed set | Generation, checking, seed license | What seeds are worth compared with volume |
| H: hybrid | Half real, half synthetic by cost | Every line | Complement or substitute |
| V: synthetic with extra checking | As S, with a larger review budget | Generation, expanded checking | Whether validation, not inputs, is the gap |
| Holdout: sealed real cases | Never used for training or tuning | Collection and custody | The single score every arm is judged on |
What could make the ceiling thesis wrong
The thesis rests on one claim: seeds, checks and access stay expensive while generation gets cheap. Two risks sit outside that claim and are not measured here, buyer concentration, where a few buyers can set the price for every seller, and rights, since a ceiling means nothing for data that cannot be licensed for the intended use.
Five developments would undercut the claim itself, and each has a public marker to watch:
- Verifiers spread. Verifiable rewards began with skills that have checkable answers; if reliable automatic checks reach business judgments, the validation line shrinks and ceilings fall with generation.
- Buyers build their own seeds. Phi-4 drew seeds from web content, licensed books and code; if public and already-licensed material proves enough for most business tasks, the scarce-input premium narrows to what is truly private.
- Distribution takes the margin. Marketplaces and generator licenses can claim the value that leaves the generation line, leaving little for the seller of the underlying records.
- Collapse evidence fails to generalize. The accumulation result comes from controlled experiments; if mixtures with little real data hold up at frontier scale, the anchor role of real data weakens.
- Stated demand does not convert. A published spending plan is not spending; if buyer plans stall, scarce inputs lose the price support this thesis expects.
Questions that test a data seller's pricing story, SourceX included
Good questions place a seller's revenue on the matrix and test how it holds when generation gets cheaper again. For SourceX the structural facts are public: it takes businesses that offer their operational records to AI developers through qualification, rights review, contracting, preparation, delivery and payment, and it sources datasets against buyer requests instead of holding stock. Its commercial figures are not public; SourceX shares company and commercial metrics with investors directly, and materials can be requested at sourcex.si/investors.
The questions apply to any seller of training or evaluation data, SourceX among them:
- Which data products make up revenue, and what is the best synthetic route to each?
- What share of each product's substitute cost is generation, and how was that share estimated?
- Has any buyer run a matched-budget comparison against the seller's data, and what did parity cost?
- How are prices set: per record, per term of access, per seed or holdout license, or per evaluation run?
- Which channel reaches buyers, and how much of the value does it take?
- Which products depend on a generator's license terms, and what do those terms require downstream?
Methods note: how the ceiling rule, matrix and plan were built
The ceiling rule, the substitution matrix, the value splits and the experiment plan are editorial tools written for this article on 27 October 2026. The rule is arithmetic on an assumed cost structure. The generation shares and value splits are judgments drawn from the sources cited: Stanford's AI Index 2025 and the Epoch AI estimate it cites for price trends, the Phi-4 report and Gerstgrasser and colleagues for the role of real inputs, the Tülu 3 report for verifiable rewards, AWS Data Exchange documentation and the Llama 3.1 license for distribution terms, and micro1's October 2026 announcement for stated demand.
Not measured: actual generation shares for any real data product, transaction prices for licensed data, buyer willingness to pay, and any SourceX revenue, cost or acceptance figure. Routes A to C and every matrix rating are illustrative. A reader who disagrees with a rating can change it and redraw the triangle; the rule itself does not depend on any rating.
Frequently asked questions
If generation keeps getting cheaper, does licensed business data keep any price?
Yes, unless its best substitute is mostly generation. Cheaper generation lowers a dataset's price ceiling by roughly the share of its best substitute that is generation. Data whose substitute depends mainly on seeds, outcome records, checks or access loses little, while generic instruction data can fall toward the cost of generating it.
What is a data price ceiling?
It is the most a buyer should pay for a dataset: the cost of the cheapest other route to the same result on the buyer's own acceptance test. That route can be synthetic generation with checking, a public source, a different seller or the buyer's own collection, and every cost line in it counts.
Which kinds of training data lose the most value as generation gets cheaper?
Data whose substitute is mostly generation: generic instruction and chat pairs, and text that can be rewritten from public material. Outcome-linked histories, sealed evaluation sets and new periods of records hold value longer, because a generator cannot produce the facts they contain.
Does SourceX publish pricing or parity results?
No. SourceX has not published pricing, parity or acceptance results, and this article reports none. It shares company and commercial metrics with investors directly; materials can be requested at sourcex.si/investors, where an investor can also ask how any such metric is defined.
How can an investor check a seller's claim that its data cannot be synthesized?
Ask for a matched-budget comparison scored on sealed real cases, with every cost line logged, and for the parity cost of the best synthetic arm. A claim backed only by examples of bad synthetic output is weaker than one backed by a failed parity test.
Sources
- Stanford's AI Index 2025 reports the price per million tokens for GPT-3.5-level performance (64.8 on MMLU) fell from $20.00 in November 2022 to $0.07 by October 2024, more than 280-fold, and cites Epoch AI that prices fell 9 to 900 times per year depending on the task. Source
- The Phi-4 technical report describes curating organic sources, including web content, licensed books and code repositories, to extract seeds for synthetic data, and about 400 billion unweighted synthetic tokens across 50 dataset types. Source
- Gerstgrasser and colleagues (2024) found that replacing real data with synthetic data across generations tends toward model collapse, while accumulating synthetic data alongside real data avoids it. Source
- The Tülu 3 report introduces reinforcement learning with verifiable rewards, which only gives a reward when model completions are verified correct, targeting skills such as mathematics and precise instruction following. Source
- AWS Data Exchange pricing states that AWS Marketplace charges tiered fulfillment fees for revenue collections made by AWS for new subscriptions to data products. Source
- AWS documentation says all AWS Data Exchange products are subscription-based under the provider's offer terms, and that once a subscription expires the data sets can no longer be viewed or exported. Source
- The Llama 3.1 community license requires models trained with Llama outputs and distributed to include Llama at the beginning of their names, and requires licensees above 700 million monthly active users to request a license from Meta. Source
- micro1 announced in October 2026 a commitment to spend $1 billion over the next 12 months on acquiring and licensing enterprise operational data for reinforcement learning environments, a stated plan. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.