AI data market
Open models and enterprise data licensing demand: a six-effect ledger
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Open model releases lower the cost of a capable base model and widen the pool of teams that adapt models, but they also ship free substitutes: licenses that permit distillation and openly published post-training data. Enterprise data licensing demand therefore falls for generic examples and rises for domain records and evaluation sets. Count it by independent budget owners and paid commitments, not downloads.
License terms
Open models enterprise data
Key takeaways
- Open-weight licenses moved from barring the use of outputs to improve other models (Llama 2) to explicitly allowing distillation (Llama 3.1, DeepSeek-R1's MIT license).
- That legal shift lowers demand for generic instruction data, which open recipes such as Ai2's Tülu 3 already publish.
- The same releases widen the potential buyer base: more teams can adapt a capable model, and each new checkpoint forces another round of testing.
- Downloads are adoption, not spending; Meta's billion Llama downloads say nothing about purchased data.
- The thesis holds only if written requests and paid commitments from distinct budget owners rise by dated cohort.
Do open model releases raise or lower enterprise data licensing demand?
Both, in different places. An open release makes a capable base model cheap to obtain and run, which brings new organizations into model adaptation, and adaptation consumes domain records and evaluation sets. The same release usually makes generic training data cheaper as well, through licenses that permit training on model outputs and openly published post-training recipes, so outside demand shrinks for anything a public model or dataset can already supply.
For an investor weighing open models and enterprise data licensing demand, the useful question is not whether open releases are good for data suppliers but which supplier categories they help and which they hollow out. This article builds a six-effect ledger, ties each effect to dated public evidence and sets out the procurement evidence that would reveal the net effect, which no public source has yet measured.
What an open release now hands every adapter
A modern open release hands over weights, a license that increasingly permits training on the model's outputs, and often a hardware footprint small enough for one company to run. The license change matters most for data demand, and it happened in roughly a year.
Meta's Llama 2 license, dated July 18, 2023, barred using the model or its outputs to improve any other large language model. The Llama 3.1 license, dated July 23, 2024, instead lets developers use outputs to create, train or fine-tune another model, on condition that a distributed model's name begins with Llama, and the Llama 3.1 model card says the license supports synthetic data generation and distillation. Both licenses keep a separate clause requiring companies with more than 700 million monthly active users to request a license from Meta.
Other releases went further. DeepSeek's repository releases DeepSeek-R1 under the MIT license and states that it allows commercial use and derivative works, including distillation for training other language models. OpenAI's gpt-oss models ship under Apache 2.0 and are described as fully fine-tunable; the smaller gpt-oss-20b, with 21 billion parameters of which 3.6 billion are active, runs within 16GB of memory.
The open-release demand ledger: three effects down, three up
Each open release pushes demand for licensed business records down through substitutes and up through new adapters and new tests, and the two forces land on different products. Generic instruction and preference data sit almost entirely on the downside; domain records with outcomes and evaluation sets sit mostly on the upside.
The ledger names six effects, the direction each pushes outside data demand, the dated evidence and the purchasing observation that would confirm it. Directions are editorial judgments drawn from the cited evidence; magnitudes are unmeasured.
| Effect of an open release | Push on licensed business data | Dated evidence | What would confirm it |
|---|---|---|---|
| Permissive output terms | Down for generic examples | Llama 2 license (2023) barred improving other models; Llama 3.1 card (July 2024) supports distillation; DeepSeek-R1 under MIT (2025) | Adapters report that teacher-generated data replaced a planned purchase |
| Open post-training recipes | Down for generic instruction and preference data | Ai2's Tülu 3 reproduction trains on published mixtures such as tulu-3-sft-mixture | Requests that name only general chat skills decline by cohort |
| Weights that run in-house | Down where the adapter's own records suffice | gpt-oss-20b runs within 16GB of memory under Apache 2.0 | Pilots close using internal records only |
| More teams adapting models | Up in the number of potential buyers | Meta reported more than 1 billion Llama downloads (March 2025) | Written requests from distinct, independent budget owners rise |
| Frequent checkpoints to compare | Up for evaluation sets | AI Index: the open-closed gap narrowed in its 2025 edition and reopened in its 2026 edition | Test-set purchases recur at model swaps |
| Domain gaps open data cannot fill | Up for outcome-linked domain records | Public mixtures are public by design and hold no private operating history (editorial) | Paid commitments that specify domain decisions and their outcomes |
Who becomes a data buyer when the base model costs nothing to license?
Mostly the adapter itself, using its own records; an outside purchase follows only when those records run thin. Open weights let a company tune inside its own environment, which removes the need to send records to a model vendor and much of the privacy friction of adaptation. That expands adaptation faster than it expands buying.
Adoption is large but it is not spending. Meta said in March 2025 that Llama had been downloaded more than one billion times. A download can be a test, a mirror or an automated job, and none of those shows a data purchase, which is why downloads sit at rung zero of the ladder later in this article.
Closed providers are also changing the terms of custom models. OpenAI's fine-tuning best-practices page, checked on 2026-10-26, says the company is winding down its fine-tuning platform and no longer accepts new users, which leaves teams that want their own tuned model with open weights or another provider. If that pattern spreads, open-model adapters become the main population that a supplier of tuning data sells to, and that population is many independent teams rather than a few labs.
That breadth is the investable part of the thesis, and it would be a change from the record so far. Reddit's Form 10-K for fiscal 2025 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." A buyer base made of many adapters would be a structural shift away from that pattern, and it is testable.
Why every new open checkpoint triggers a round of testing
Because the best available model keeps changing, each adapter has to re-run its own tests before swapping a base model, and those tests need real tasks. Stanford HAI's 2025 AI Index reported that open-weight models narrowed their gap with closed models from 8 percent to 1.7 percent on some benchmarks within a single year. Its 2026 edition reports that the gap reopened during 2025, with the top closed model 3.3 percent ahead of the top open model as of March 2026, up from 0.5 percent in August 2024, on that edition's measure.
A gap that moves both ways in two years is the condition under which evaluation demand recurs. Anthropic's engineering guidance on agent evaluations suggests starting with 20 to 50 simple tasks drawn from real failures and keeping regression suites near a 100 percent pass rate. For an adapter choosing among open checkpoints, those tasks come from its own operations or from licensed records of comparable work, and they are re-run on the release calendar of someone else's model.
Free substitutes that arrive with the weights
Open releases increasingly ship the data recipe along with the model, which commoditizes general post-training data. Ai2's open-instruct repository documents how to reproduce its Tülu 3 models with training commands that load published datasets by name, a supervised mixture and preference mixtures among them, so anyone can repeat general instruction tuning without buying examples.
Developers also generate their own. The Llama 3.1 model card says its fine-tuning data included more than 25 million synthetically generated examples, and DeepSeek's distilled models were fine-tuned on samples generated by DeepSeek-R1. With licenses now permitting that use, a supplier of generic demonstrations competes with a free teacher.
What the substitutes do not contain is private operating history: how a particular type of claim, dispute or work order was resolved, and what happened afterward. That is where the upside rows of the ledger point, and why a supplier's mix of generic and domain-specific requests is one of the first things to check.
Counting demand from open-model adapters without counting downloads
Count independent budget owners moving up a dated ladder, not activity on model hubs. Every rung above zero needs a document; downloads, forks and publicly posted fine-tunes are all rung-zero signals.
SourceX has published no results from this design. It is the evidence an investor should request from any company claiming that open releases drive its demand, and it has two parts: the interview protocol, then the ladder it fills.
- Sample independent budget owners who adapted an open-weight model for a production task in a stated window, excluding model vendors and consultants paid on the outcome.
- Record each adaptation as one case: the base model, any checkpoint swaps during the window, the data used and where it came from (own records, generated, open dataset or licensed).
- Give each case a commitment class, request, evaluation or paid commitment, and confirm everything above a request with a dated document.
- Report by quarterly cohort against the full denominator, including cases that never asked for outside data.
- Separate evaluation purchases from training purchases, and note whether each followed a checkpoint swap.
- Publish nonresponse and intervals; with small cohorts, report counts rather than percentages.
| Rung | Signal from an open-model adapter | Proof document | Open-model trap |
|---|---|---|---|
| 0. Adoption | Downloads, forks and posted fine-tunes | Hub statistics | Automated pulls and mirrors inflate the count |
| 1. In-house pilot | A team tunes an open base model on a named task | Pilot plan with base model and test set | Tuning on own records only never reaches an outside supplier |
| 2. Checkpoint comparison | Two or more open bases tested on the same held-out tasks | Dated evaluation report | Public leaderboards used instead of the adapter's own tasks |
| 3. Written outside request | Scope, volume, rights and acceptance test specified | Dated request from the budget owner | Requests for generic chat data that open mixtures already cover |
| 4. Paid commitment | Signed license or order for data | Contract or purchase order | Spending on GPUs or hosting counted as data demand |
| 5. Repeat at swap | A new purchase when the base model changes | Second order or renewal | Re-running an old test set at no cost |
Risks that could break the open-model demand thesis
Announcements and capability gains may not turn into data spending, and several forces could keep it flat. Most of them have a public marker an investor can watch.
- Teachers get good enough. Licenses that allow distillation, combined with stronger open teachers, could fill most domain gaps with generated data; DeepSeek's distilled models show how far teacher samples alone can take a small open model.
- Adapters stay in-house. Running weights inside the firewall makes a company's own records the default input, and many adaptations may never need outside data.
- Buying stays concentrated. Reddit's fiscal 2025 Form 10-K traces substantially all of its licensing contract value to two partners, and a wider adapter base has not yet appeared in public filings.
- License terms can tighten. Llama 2 barred training other models on its outputs and Llama 3.1 permits it, which shows that output terms are set release by release and could narrow again.
- Open checkpoints may consolidate. If one open family dominates, re-testing becomes less frequent and the evaluation upside shrinks; the AI Index's reopening gap shows the ranking can move either way.
Questions for a supplier selling into open-model adapters, SourceX included
Ask a supplier to split its demand by ledger row and ladder rung, backed by documents. The same questions apply to SourceX.
SourceX does not hold datasets in stock: each one is sourced from U.S. businesses after a buyer request that states scope, volume, language, format, timing and licensing needs, so its demand evidence begins at the written-request rung. From there it sources and vets suppliers, reviews their rights, draws up the contract, prepares and delivers the data and handles payment, and a supplying company releases no data until it has approved the scope and terms and signed. Request and deal metrics are not published. SourceX reports company and commercial metrics to investors directly, not on its site, and its current raise is from accredited investors under Rule 506(c), with materials on request through its investors page.
- What share of written requests in the last four quarters came from teams adapting open-weight models, and which base models did they name?
- How many of those requests asked for data that an open post-training mixture already covers, and what happened to them?
- What share of paid commitments were evaluation sets, and how many recurred after a checkpoint swap?
- How many distinct, independent budget owners stand behind your paid commitments, and what share of value comes from the two largest?
- Which license change by an open model releaser would hurt your demand most, and how quickly would you see it in requests?
Methods note: sources and limits of the six-effect ledger
All three tools here, the ledger, the ladder and the interview design, are editorial frameworks drawn up for this article on 2026-10-26 from Meta's Llama 2 and Llama 3.1 licenses and the Llama 3.1 model card, the DeepSeek-R1 repository, OpenAI's gpt-oss repository and fine-tuning documentation, Ai2's open-instruct documentation, Meta's March 2025 download announcement, Stanford HAI's 2025 and 2026 AI Index reports, Anthropic's evaluation guidance and Reddit's Form 10-K for fiscal 2025. Ledger directions are judgments from that evidence.
Not measured: the size of any ledger effect, the share of open-model adaptations that buy outside data, how adapters split spending between evaluation and training, and how license terms change purchasing. No public source reports these figures, and SourceX has not published its own.
Frequently asked questions
Do open-weight models reduce the need to license training data?
For generic training data, largely yes: licenses such as DeepSeek-R1's MIT terms allow distillation, and open recipes publish general post-training mixtures. For domain records with real outcomes, and for evaluation sets drawn from real work, open models can increase the number of teams that need them.
Can a company train its own model on outputs from Llama or DeepSeek-R1?
Under the published terms of the versions reviewed here, generally yes: the Llama 3.1 license permits using outputs to train other models, with a naming condition for distributed models, and DeepSeek-R1's MIT license explicitly allows distillation. The Llama 2 license did not allow it. Terms vary by release, so counsel should read the license for the exact model in use.
Why does a new open model release create evaluation work?
Each release can change which model is best for a task, and public rankings have moved both ways: the AI Index reported the open-closed gap narrowing in its 2025 edition and reopening in its 2026 edition. A team that swaps base models has to re-run its own held-out tasks first, and those tasks come from real work.
Are open model downloads a good proxy for data demand?
No. Downloads measure adoption and include tests, mirrors and automated jobs. Meta's report of more than one billion Llama downloads says nothing about data purchases; written requests and paid commitments from independent budget owners are the evidence that counts.
How should an investor ask SourceX about open-model demand?
Ask for requests and paid commitments split by ledger row and ladder rung, with the number of distinct budget owners behind them. SourceX has not published these metrics; it gives company and commercial metrics to investors in direct conversations, sends materials on request through the SourceX investors page and limits any offering to verified accredited investors.
Sources
- Meta's Llama 2 license (version release date July 18, 2023) barred using the Llama Materials or their outputs to improve any other large language model, and required companies above 700 million monthly active users to request a license. Source
- The Llama 3.1 license (version release date July 23, 2024) permits using outputs to create, train or fine-tune another AI model, requiring a distributed model's name to begin with Llama, and keeps the 700 million monthly active user clause. Source
- The Llama 3.1 model card (released July 23, 2024) says the license supports using outputs to improve other models, including synthetic data generation and distillation, and that fine-tuning data included over 25M synthetic examples. Source
- DeepSeek-R1 is released under the MIT license, supports commercial use and allows derivative works including distillation for training other LLMs; its distilled models were fine-tuned on samples generated by DeepSeek-R1. Source
- OpenAI's gpt-oss repository: Apache 2.0 license; fully fine-tunable; gpt-oss-20b has 21B parameters with 3.6B active and runs within 16GB of memory. Source
- Ai2's open-instruct documentation gives commands to reproduce the Tülu 3 models using published datasets such as allenai/tulu-3-sft-mixture and Tülu 3 preference mixtures. Source
- Meta announced on March 18, 2025 that Llama had been downloaded more than one billion times. Source
- OpenAI's fine-tuning best-practices page says OpenAI is winding down its fine-tuning platform and that it is no longer open to new users (checked 2026-10-26). Source
- Stanford HAI's 2025 AI Index: open-weight models narrowed the gap with closed models from 8 percent to 1.7 percent on some benchmarks in a single year. Source
- AI Index 2026, technical performance chapter: the open-closed gap reopened in 2025; as of March 2026 the top closed model led the top open model by 3.3 percent, up from 0.5 percent in August 2024. Source
- Anthropic's guidance on agent evaluations suggests starting with 20 to 50 simple tasks drawn from real failures, with regression suites kept near a 100 percent pass rate. Source
- Reddit's Form 10-K for fiscal 2025 states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.