AI data market
AI data supplier concentration risk: a backup plan for one-supplier models
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Inventory every model asset that relies on the supplier, from refresh data and eval sets to label definitions and rights after termination. For each, estimate how long a substitute would take to qualify and what you keep if the relationship ends, then accept the dependency with a written contingency, pilot a standby source, or negotiate continuity terms. Recheck at each renewal and major release.
System export
AI data supplier
Key takeaways
- Plan per dependent asset, not per supplier: refresh data, eval sets, schemas, rights, documentation and delivery fail differently.
- Compare the time to qualify a substitute with how long you can run on data you are allowed to keep.
- Move evaluation data to an independent source before adding training suppliers.
- A standby source becomes a backup only after a pilot proves schema mapping, model parity, rights, documentation and delivery.
- Retention after termination and notice periods often reduce concentration risk more cheaply than a second supplier.
How should an AI buyer plan for dependence on one data supplier?
Start from what would break, not from how many suppliers you have. List every model asset that relies on the supplier, estimate how long each would take to replace and what you could keep if the relationship ended, then decide per asset whether to accept the dependency with a written contingency, qualify a standby source, or renegotiate continuity terms. Revisit the answer at every renewal and before every major model release.
Concentration on one supplier is often the right call. A unique set of records, a supplier with clean rights documentation, or a narrow fine-tuning job may have no realistic second source. The risk is not the single supplier itself. It is discovering, at the moment a delivery slips or a renewal is declined, that nobody knows which models, evaluations and retraining plans depended on it.
This page gives a worksheet for that inventory and a decision rule for each line. Choosing how many suppliers to use from the start, and the mechanics of switching once you have decided to, are covered in the linked procurement guides.
Map what breaks, not who you buy from
A dependency is any asset whose next version needs something only this supplier provides. For AI data that usually means more than the training records. Six kinds of dependency show up repeatedly, and each fails differently.
At a glance
Map what breaks, not who you buy from
- Refresh supply: scheduled deliveries that keep a production model current. Fails when a delivery slips or the source system behind it migrates.
- Evaluation sets: held-out data used to decide whether a model ships. Fails quietly when the eval set comes from the same supplier as the training data and shares its formats and labeling choices.
- Schema and label definitions: field meanings, status vocabularies and outcome codes that downstream code assumes. Fails when a substitute supplier uses different definitions.
- Rights scope: what you may keep, retrain on and redistribute after the term. Fails when the agreement ends and retained copies or derived assets fall outside what it allows.
- Documentation: provenance, collection and preprocessing records you rely on for your own disclosures. Fails when the supplier stops providing them or the upstream holder changes.
- Delivery path: storage, access and transfer arrangements built around one counterparty. Fails when the supplier's storage or access model changes.
When a single supplier needs a written contingency
Write a contingency when a supplier's failure would stop a high-impact model from being retrained, evaluated or defended within your release cycle. NIST's AI Risk Management Framework expresses the same expectation in its GOVERN 6.2 outcome: contingency processes in place to handle failures or incidents in third-party data or AI systems deemed to be high-risk. NIST's generative AI profile adds an action to document value-chain risks so as to identify over-reliance on third-party data and identify fallbacks.
Regulated sectors give useful templates even when they do not apply to you. The EU's Digital Operational Resilience Act asks financial entities to assess, before contracting, whether an ICT provider is not easily substitutable, and to keep documented, tested exit strategies with alternative solutions and transition plans for services that support critical functions. The 2023 U.S. interagency guidance for banking organizations treats termination as a stage of the third-party life cycle and lists factors such as moving an activity to another third party or in-house, and the return or destruction of data. In September 2026 the agencies proposed new guidance that, once finalized, would replace it, so check which text is in force before quoting it.
Borrow the structure, not the legal duties. An AI lab is not a bank, but the questions those frameworks force, namely who else could provide this, how long the move would take and what happens to the data, are the right questions for a data supplier too.
Two signals usually mean the contingency is overdue. The first is a model whose release notes cite gains that nobody can attribute to anything except the latest delivery from one supplier. The second is a renewal discussion in which your team cannot say, within a few weeks, how long it would take to replace that supplier. Either one means the dependency has grown faster than the documentation around it.
The supplier concentration and substitution worksheet
Fill in one row per dependency, not per supplier. The values below are illustrative and show the level of detail that makes the worksheet usable in a review.
| Input | What to record | Illustrative entry | Why it matters |
|---|---|---|---|
| Dependent asset | Model, eval set or pipeline and its owner | Support assistant v3 fine-tuning set, applied ML lead | Names who feels the failure |
| Share from this supplier | Percent of records or tokens in the asset | 80% of fine-tuning records | Shows how much would need replacing |
| Substitutability | Unique, substitutable with effort, or commodity | Substitutable with effort | Sets the default posture |
| Time to qualify a substitute | Weeks from first contact to accepted pilot | 10 to 16 weeks | Compared with the release cycle |
| Retained after termination | What the agreement lets you keep and retrain on | Delivered copies kept; no new refreshes | Decides whether models survive an exit |
| Schema portability | Whether your canonical schema maps to other sources | Canonical schema exists; outcome codes are supplier-specific | Drives mapping effort |
| Eval independence | Whether eval data comes from another source | No, same supplier | Hidden coupling to fix first |
| Upstream holder | Who originally holds the records | Two company sources behind the supplier | Two contracts may share one source |
| Posture and trigger | Accept, standby source or renegotiate; what event triggers action | Standby source; trigger is a missed refresh or a narrowed scope | Turns the row into a decision |
Turning the worksheet into a decision for each row
Compare the time to qualify a substitute with the time you can run on what you already hold. If the agreement lets you keep delivered data and the model can go one or two release cycles without a refresh, accepting the dependency with a written contingency is usually enough. If qualification takes longer than you could operate without the supplier, you need either a standby source or stronger continuity terms.
Fix hidden coupling before adding suppliers. An evaluation set from the same supplier as the training data is the cheapest dependency to remove and often the most consequential, because it makes a supplier problem look like a model improvement. Move the eval set to an independent source first.
Measure concentration with a simple number so it can be tracked. The U.S. Department of Justice computes the Herfindahl-Hirschman Index as the sum of squared market shares, approaching zero with many equal participants and reaching 10,000 when a single firm holds everything. The same arithmetic on supplier shares of an asset, computed by original holder as well as by contracting supplier, shows whether diversification is real. Treat the number as a monitoring signal, not a threshold borrowed from merger review.
What a substitute pilot must prove before you depend on it
A standby source is only a backup once a pilot shows the model can use it. Run the pilot on a small, evaluation-only license and decide in advance what counts as success. Five things should be established.
At a glance
What a substitute pilot must prove before you depend on it
Checklist · 5 items
- Schema mapping: every field your pipeline reads maps to the substitute's data, with differences in vocabularies, time zones and status codes written down.
- Model parity: a model trained or tuned with the substitute's sample performs within an agreed margin on your independent eval set, reported with its uncertainty.
- Rights equivalence: the substitute's license covers the same uses, retention and derived assets as the primary, or the gap is documented and accepted.
- Documentation: the substitute provides the provenance, composition, collection and preprocessing information you need, in the style of a datasheet for datasets.
- Delivery: one test delivery arrives through the agreed path with a manifest, and both sides verify it.
0 of 5 ticked. Tick each item as you go.
Continuity terms worth negotiating with the primary supplier
Many concentration problems are cheaper to solve in the contract than with a second supplier. The terms below make an exit survivable without changing what you buy.
| Term | What it secures | Question to ask | Fallback if refused |
|---|---|---|---|
| Retention after termination | Delivered data and derived models remain usable | May we keep and retrain on data delivered before the end date? | Prioritize a standby source |
| Notice period | Time to qualify a substitute | How much notice before non-renewal or scope changes? | Shorten the release dependency |
| Change notice | Warning before upstream sources or schemas change | Will you tell us before a source system or holder changes? | Add schema checks on every delivery |
| Transition assistance | Documentation and mapping help during exit | Will you hand over field definitions and history? | Build your canonical schema now |
| Data return or deletion | Clear obligations for copies at the end | Which copies must be deleted, and do models count? | Get counsel to settle it before signing |
| Documentation delivery | Provenance records you rely on for disclosures | Do datasheets come with every delivery? | Treat missing documentation as a defect |
Worked scenario (illustrative): a support assistant with one supplier behind most of its data
Illustrative scenario, invented to show the method; it does not describe a real company or supplier. An AI team ships a customer support assistant. One supplier provides 80% of its fine-tuning records and all of its evaluation set, with quarterly refreshes. The team releases a new model version every 12 weeks.
The worksheet shows three rows that matter. Fine-tuning supply is substitutable with effort, with an estimated 10 to 16 weeks to qualify another source. The agreement allows delivered data to be kept after termination, so the current model survives an exit but stops improving. The eval set is the weak point: it comes from the same supplier, so any change in that supplier's data would also shift the yardstick.
The team's decisions follow from the rows. It sources a new eval set from a different holder first. It negotiates a 90-day non-renewal notice and a change notice for upstream source changes, which covers most of the qualification window. It runs a small evaluation-only pilot with one standby source and records the schema gaps. It does not split production volume, because the pilot showed outcome codes that would take weeks to harmonize.
The plan names its trigger: a missed refresh or a narrowed scope moves the standby source from pilot to order. That one sentence is what turns a worksheet into a backup plan.
Which assumption changes the answer, and how to show uncertainty
Three assumptions usually decide the outcome, so state them as ranges and show what changes at each end. Time to qualify a substitute is the first: a range of 6 to 20 weeks can flip a row from accept to standby. Retention after termination is the second: if delivered data must be deleted at the end of the term, every model trained on it inherits the dependency. The third is whether the upstream holder is really different, because two suppliers can resell the same source.
Present the worksheet with a best case and a worst case for each of these, and mark which rows change posture between them. Reviewers from research, security and legal can then argue about the assumption that matters instead of the whole plan. Where an answer depends on how a contract clause is read, mark it for counsel rather than guessing.
Refresh the worksheet when the supplier's offering, contract, delivery path or acceptance criteria change, and whenever a new model or eval set starts depending on the same source.
Adding a second source through on-request sourcing
A standby source does not have to be another catalog. When a buyer asks SourceX for a backup source, nothing comes out of an inventory: records are sought at U.S. businesses against the scope, volume, language, format, timing and licensing terms the buyer's request lays out. The worksheet's schema, rights and documentation rows are the natural content of that request.
Frequently asked questions
Is relying on one AI data supplier always a risk to fix?
No. A unique dataset or a supplier with clean rights may have no realistic substitute. The risk to fix is an undocumented dependency: write down what breaks, how long replacement would take, and what you can keep if the relationship ends.
What should a backup data supplier pilot prove?
That your pipeline can read the substitute's fields, that a model using its sample performs within an agreed margin on an independent eval set, and that its license and documentation match your needs. One verified test delivery should complete the pilot.
Should the evaluation set come from a different supplier?
Usually yes. An eval set from the same supplier as the training data shares its formats and labeling choices, so it can hide problems in that supplier's data. Moving the eval set to an independent source is often the first fix.
How do you measure supplier concentration in a training set?
Compute each supplier's share of records or tokens in the asset and sum the squared shares, the same arithmetic as the Herfindahl-Hirschman Index. Repeat it by original data holder, because two contracts can trace back to one source.
Which contract terms reduce single-supplier risk?
Retention of delivered data after termination, a notice period long enough to qualify a substitute, notice of upstream source changes, transition help with field definitions, and clear rules on what must be deleted at the end. Have counsel review how deletion applies to trained models.
Sources
- NIST's AI Risk Management Framework includes the outcome that contingency processes are in place for failures or incidents in high-risk third-party data or AI systems (GOVERN 6.2). Source
- NIST's generative AI profile suggests documenting value-chain risks to identify over-reliance on third-party data and to identify fallbacks. Source
- The EU Digital Operational Resilience Act asks financial entities to assess providers that are not easily substitutable and to keep tested exit strategies with alternative solutions and transition plans. Source
- The 2023 U.S. interagency guidance treats termination as a life-cycle stage, including moving an activity to another third party or in-house and returning or destroying data. Source
- The Department of Justice computes the Herfindahl-Hirschman Index as the sum of squared market shares, reaching 10,000 when one firm holds the whole market. Source
- Datasheets for datasets document a dataset's motivation, composition, collection process, preprocessing, uses, distribution and maintenance. Source
- In September 2026 the federal banking agencies proposed new third-party risk management guidance that, once finalized, would replace the 2023 guidance. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.