Skip to content

Deal economics

Enterprise data licensing as recurring infrastructure: who holds the SLA?

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Enterprise data licensing becomes recurring infrastructure only when the licensor sells commitments that outlast the copy: scheduled flow, corrections, measured availability, access that ends when payment stops, and documentation refreshed for each model release. The clearest public precedent, Equifax's The Work Number, shows where that value lands: with the operator that carries the duties and charges per transaction, while contributing employers are generally not charged. Price the commitments, not the platform label.

License terms

Enterprise data licensing

This page covers enterprise data licensing as recurring infrastructure: who holds the SLA.Illustrative

Key takeaways

  • A dataset copy creates no renewal trigger; recurring revenue needs a commitment that keeps the vendor working after delivery.
  • Amazon's EC2 SLA shows the five parts of a real service commitment: a defined metric, a threshold, a remedy, exclusions and a claim procedure.
  • In The Work Number, about 188 million active employment records at December 31, 2024, the operator earns the recurring revenue while contributing employers are generally not charged.
  • Revocable hosted access is the strongest renewal trigger but fits evaluation and retrieval better than training, because a trained model keeps what it learned.
  • Ask for revenue and labor per delivered period per contracted scope; recurring revenue with flat per-period labor is a services business with a platform name.

Can enterprise data licensing become infrastructure as a service?

Yes, but only for licensors willing to sell obligations rather than copies. A dataset license, however large, is a delivery: once the buyer holds the records and a trained model, the next payment needs a new reason. Infrastructure revenue recurs because the vendor keeps doing something the buyer cannot do for itself after the contract ends.

So the investor question is not whether a company calls its product a platform. It is which service commitments the company signs and which event makes the customer pay again. This article names five such commitments, tests them against the clearest public precedents, and maps who owns rights, performs labor, supplies tooling and captures recurring value at each step, with a metric for each that can be compared across companies of different sizes.

Two neighboring questions are left to other pages: how to draft a refresh service level clause by clause, and how acquirers treat one-time versus recurring license revenue in a quality of earnings review.

What does a service commitment contain? Read a cloud SLA

A service commitment has five parts that can be read straight off a cloud contract: a metric defined by formula, a threshold, a remedy, exclusions and a claim procedure. Amazon's EC2 service level agreement, last updated May 25, 2022, is a compact example of all five.

For EC2 instances running across two or more availability zones in a region, AWS commits to commercially reasonable efforts toward a monthly uptime percentage of at least 99.99%, defined as 100% minus the share of minutes in the month the service was unavailable. Missing it earns service credits of 10%, 30% or 100% of the monthly EC2 bill depending on how far uptime fell. Credits are the customer's sole and exclusive remedy, outages outside AWS's reasonable control are excluded, and the customer must claim by the end of the second billing cycle with its own request logs as evidence.

Translate each part to data and the gap is plain. None of the five comes with a dataset copy. A supplier whose main business is freight or insurance has no reason to promise them, so someone else, usually an intermediary with its own pipeline, has to carry them before a license can be sold as a service.

What does a service commitment contain? Read a cloud SLA
SLA elementIn the EC2 agreementData licensing equivalentWho can realistically carry it
MetricMonthly uptime percentage, defined by formulaShare of scheduled deliveries on time and complete against a source control totalThe party that runs extraction and preparation
ThresholdAt least 99.99% for multi-zone deploymentsA delivery and correction window the pipeline can meet every periodAn operator with spare capacity, not the source company
RemedyService credits against future bills, sole remedyCredits or a free re-delivery when a period is missedWhoever is paid per period
ExclusionsForce majeure and the customer's own actionsSource-system outages, migrations and rights withdrawalsWritten jointly with the supplier
Claim evidenceCustomer's request logs within two billing cyclesDelivery manifests and acceptance logs both sides can readBoth parties, from shared logs

Payroll records already run as infrastructure: what The Work Number shows

The clearest public example of enterprise operating records running as recurring infrastructure is not an AI deal. Equifax's FY2024 Form 10-K says The Work Number held about 188 million active and 734 million total (active and historic) employment records at December 31, 2024.

The economics sit with the operator. Equifax's FY2025 10-K describes employers sending payroll data electronically so that Equifax can answer verification requests on their behalf, and the FY2024 filing says contributor turnover has been low because Equifax generally does not charge employers to add their data. The contributors' reward is verification work taken off their hands; the filing describes charging per transaction for the verifications themselves.

The duties travel with the revenue. In its FY2024 regulatory discussion, Equifax describes what the Fair Credit Reporting Act asks of consumer reporting agencies: reasonable procedures to assure maximum possible accuracy, and a reasonable reinvestigation when a consumer disputes information. Standing accuracy and correction duties of that kind are exactly what a one-time dataset copy never carries.

For an AI data investor the precedent cuts both ways. Recurring value from enterprise records is possible at very large scale. In this example, though, the recurring revenue is earned by the party that runs the pipeline, answers disputes and serves users every day, while the filings describe the contributing companies' benefit as verification handling and no charge for adding data.

Five commitments that make a license recur, and the renewal trigger each creates

A license recurs when at least one commitment creates a renewal trigger: an event after which the buyer's existing copy is no longer enough. The five below run from the easiest for an enterprise supplier to sign to the hardest, and the figure draws them as a ladder.

The fourth rung is the one AI training resists. A model trained on a copy keeps what it learned, so revocable access fits evaluation sets, retrieval corpora and agent environments better than pretraining. That is where an infrastructure thesis for enterprise records is most plausible, and where an investor should look for it first.

The five commitments, each with the trigger it creates:

  • Scheduled flow. The licensor delivers new records each period from live systems. Trigger: a period closes and the buyer's copy is out of date. Labor: extraction, re-qualification and preparation every cycle.
  • Correction and withdrawal. The licensor fixes errors, removes records whose rights lapse and tells every buyer. Trigger: a dispute, a deletion request or a change in a supplier's contracts. Labor: a standing dispute process, like the reinvestigation duty in consumer reporting.
  • Measured availability. The licensor runs an access point with a defined metric, threshold and credits, as in a cloud SLA. Trigger: none on its own; it raises switching costs rather than forcing a renewal.
  • Access that ends with payment. The buyer queries hosted data instead of taking a copy. Google's BigQuery sharing works this way: subscribing creates a linked dataset rather than a copy, the publisher can remove a subscriber, and an unlinked dataset can no longer be queried. Trigger: the contract end date itself.
  • Documentation per release. The licensor keeps provenance and processing records current for each model the buyer ships. The EU AI Act requires providers of general-purpose AI models to draw up and keep up to date technical documentation covering the type and provenance of training data and curation methods. Trigger: each new model version trained on the data.
A five-step ladder rising from a one-time copy with no renewal trigger: scheduled flow (a period closes), correction and withdrawal (a dispute or deletion), measured availability (switching cost), access that ends with payment (contract end) and documentation per release (a new model version).
The commitment ladder, with the renewal trigger each rung creates. Editorial framework, 27 October 2026; not a measurement.

Who owns rights, does the work and keeps recurring value at each rung

The comparison normalizes each model to the unit that actually recurs, so revenue can be compared per contracted scope rather than per headline contract. Roles follow the public documents cited above; the contributory row describes the Work Number pattern, and the documentation row is a composite of what the EU AI Act asks model providers to maintain.

Read the last column as the request an investor makes in diligence. Each metric has a denominator the company cannot choose after the fact: accepted datasets, delivered periods, subscriber-months, active records or supported model releases.

Who owns rights, does the work and keeps recurring value at each rung
ModelRights holder and recurring laborToolingRecurring value goes toScope-normalized metric to request
One-time dataset licenseSource company owns originals; licensor prepares onceLicensor or its processing partnerNobody; each sale needs new supplyRevenue per accepted dataset; share of suppliers who license a second dataset
Scheduled-delivery licenseSource company; licensor re-qualifies and prepares each periodLicensor's pipelineLicensor and supplier, while deliveries continueRevenue per delivered period per contracted scope; on-time complete deliveries
Hosted, revocable accessPublisher keeps the data; buyer never holds a copyCloud platform, such as BigQuery sharingPublisher, while access stays revocableRevenue per active subscriber-month, net of storage costs
Contributory networkContributors send records; operator runs accuracy and disputesOperatorOperator, per transactionTransactions per active record; contributor retention
Documentation and evaluation serviceLicensor maintains provenance and test sets per buyer releaseLicensorLicensor, per model releaseRevenue and refresh hours per supported model release

Is the platform label hiding a services business?

Often, and the commitments explain why. Every rung above the first adds labor: re-qualifying each period's records, answering disputes, maintaining test sets, rewriting documentation for a new model. A company can report recurring revenue that carries recurring cost almost one for one, which looks like infrastructure on a revenue chart and like a consulting firm on a margin chart.

Three tests separate the two. Does serving a renewal cost less than winning the original sale? Does labor per delivered period fall as more buyers take the same feed? Would the commitment survive the founding team leaving, because it lives in contracts and systems rather than in people's heads? A company that cannot answer the first two from its own cohort data runs a services business with a platform name. That can still be a good business; it is simply valued on different evidence.

Applied to SourceX: the rungs its published terms occupy

SourceX serves here as a labeled application of the framework, read from its own public pages on 27 October 2026. Its how-it-works page says SourceX usually pays a one-time fee for the dataset, that some agreements set recurring deliveries, each paid as the agreement states, and that there are no royalties or usage fees.

Two further terms on the same page keep it at the base of the ladder, with the first rung available as an option. Once the fee is paid, SourceX's license to the de-identified dataset is permanent and survives termination, and models already trained on it are not retrained or deleted. Exclusivity for AI training binds the supplier, so the same records cannot be licensed elsewhere for training, while SourceX's own right to distribute or sublicense the de-identified dataset is unaffected. In ladder terms, that is a copy business with a scheduled-delivery option, so the investor question is how often the option is used and what each later delivery costs.

Those figures are not published. SourceX's investor page says it shares company and commercial metrics with investors directly, and a reader evaluating the company can ask there. A pointed first question from this article: of the datasets delivered, how many sit under agreements with scheduled deliveries, and what did each later delivery cost to qualify and prepare compared with the first?

What could keep enterprise data licensing transactional

Each of the developments below would push enterprise data licensing back toward one-time transactions, and a company should be able to say which one it is most exposed to:

  • Buyers keep what they learn. Once a model is trained, a copy's main value is absorbed, so recurrence depends on new periods or new tasks rather than on access; permanent training licenses such as SourceX's make that explicit.
  • Hosted access may never fit training. If buyers insist on copies for training runs, the revocable-access rung stays confined to evaluation and retrieval work.
  • Standing duties are expensive. The accuracy and dispute obligations Equifax describes under the Fair Credit Reporting Act show what correction duties cost once they become law rather than a contract option.
  • Operators take the margin. In the contributory precedent the operator earns per transaction while contributors are generally not charged, so enterprise suppliers may end up with service benefits rather than recurring license income.
  • Model providers may build documentation in-house. The EU AI Act puts the keep-up-to-date documentation duty on general-purpose model providers, who may staff it themselves rather than buy it.
  • Labels can outrun contracts. A deal where recurring deliveries are available but never scheduled is a one-time sale with an option attached.

Methods note: how the commitment ladder and comparison were built

The five commitments, the ladder and the comparison table were drawn up for this article on 27 October 2026 as editorial tools, not as measurements. They rest on Amazon's EC2 service level agreement (last updated May 25, 2022, read in full through AWS's documentation reader), Equifax's FY2024 and FY2025 Form 10-K filings, Google Cloud's BigQuery sharing documentation, the EU AI Act as published on EUR-Lex, and SourceX's how-it-works and investor pages, fetched directly on that date.

Not measured: how many enterprise data deals carry each commitment, renewal rates, labor per delivery or margins at any company, SourceX included. We found no public source that reports these for enterprise data licensed to AI developers, so the scope-normalized metrics above are the questions to ask, not results.

Frequently asked questions

What separates a data license from data infrastructure?

A license transfers a copy and a set of permitted uses; the buyer can stop paying and keep what it received. Infrastructure keeps the vendor working after delivery, through scheduled flow, corrections, measured availability, revocable access or documentation that must be refreshed, so stopping payment costs the buyer something it still needs.

Does a delivery schedule make data licensing revenue recurring?

Only if deliveries are actually scheduled and the buyer needs each new period. A clause that makes later deliveries possible without committing either side is an option, and revenue from it should be counted per delivered period, not assumed.

Why might AI training resist an as-a-service model?

A trained model keeps what it learned from a copy, so revoking access after training removes little value. Evaluation sets, retrieval corpora and agent environments are better candidates for revocable, recurring access because they are used continuously.

Which single metric best tests an infrastructure claim in data?

Revenue per delivered period per contracted scope, paired with labor per delivered period. If revenue recurs while labor per period stays flat as buyers are added, the business is closer to services than to infrastructure.

Where does SourceX sit on the commitment ladder today?

By its published terms, SourceX usually pays a one-time fee per dataset, with recurring deliveries where an agreement sets them, and holds a permanent license once paid. That places it at the base of the ladder with scheduled flow as an option; how often recurring deliveries occur is not public.

Sources

  • Amazon's EC2 SLA (last updated May 25, 2022) defines monthly uptime, a 99.99% region-level commitment, 10%, 30% and 100% service credits as the sole remedy, exclusions and a claim procedure with request logs. Source
  • Equifax's FY2024 10-K reports The Work Number's employment records at December 31, 2024, low contributor turnover because employers are generally not charged to add data, and FCRA accuracy and reinvestigation duties. Source
  • Equifax's FY2025 10-K describes employers sending payroll data electronically so Equifax can answer verification requests, with verification generally charged per transaction. Source
  • Google Cloud's BigQuery sharing documentation describes linked datasets instead of copies, publisher removal of subscribers and unlinked datasets that can no longer be queried. Source
  • EU AI Act Article 53(1)(a) and Annex XI require general-purpose AI model providers to draw up and keep up to date technical documentation, including training data type, provenance and curation methods. Source
  • SourceX's how-it-works page states its usual one-time fee, optional recurring deliveries, no royalties, a permanent license once paid and supplier-side exclusivity. Source
  • SourceX's investor page says it shares company and commercial metrics with investors directly. 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