Skip to content

Deal economics

Data inventory or on-demand sourcing? The fit-rate hurdle that decides

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Holding data inventory beats on-demand sourcing only when the share of held datasets that find a buyer clears a hurdle: build-ahead cost times the on-demand completion rate, divided by licenses per dataset. Non-exclusive catalogs that resell clear it at low fit rates; exclusive, single-sale licenses push it toward certainty. Investors should ask for fit rates by cohort, resale counts, capital committed before signature and on-demand completion rates.

Checklist

Data inventory on demand

This page covers data inventory or on-demand sourcing? The fit-rate hurdle that decides.Illustrative

Key takeaways

  • Inventory and on-demand sourcing differ in when money is committed relative to a signed buyer, not in how fast a sample can be shown.
  • Five sourcing models sit behind the two labels: speculative catalogs, consigned catalogs, byproduct archives, commissioned collection and on-demand sourcing of existing records.
  • Holding stock is cheaper per accepted license only when the fit rate exceeds the completion rate times the build-ahead cost divided by licenses per dataset.
  • Exclusivity sets the resale count; in the illustrative table a single-sale dataset needs a fit rate above 70 percent where a four-sale dataset needs 17.5 percent.
  • Byproduct archives carry no AI-specific capital at risk, and IAS 38 keeps self-created customer lists and publishing titles off the balance sheet.

Inventory or on demand: which sourcing model should an AI data investor back?

Neither model wins in general; the arithmetic of three variables decides. Holding data inventory pays when the datasets bought or built ahead of demand fit enough requests, enough times, to cover the money committed before any buyer signed. Sourcing on demand pays when buyer specifications diverge, licenses sell once, and the time a buyer waits costs less than unsold stock.

Investors comparing AI data businesses tend to hear both pitches as moats. The inventory pitch says a ready catalog delivers in days while rivals start from scratch. The on-demand pitch says nothing sits unsold, because every dataset is sourced against a named requirement. Each claim can be true, and each hides a cost the other avoids: unsold stock in one case, lost or delayed orders in the other.

This article turns the choice into something testable. It maps five sourcing models by who funds the records and when, shows where each model's capital at risk appears in the accounts, and builds a fit-rate hurdle: the share of held datasets that must find a buyer before holding stock is cheaper per accepted license than sourcing to order. All model values are illustrative; the formula and the questions are what carry over to real diligence.

Five ways a dataset reaches the shelf, and who paid for it

The label inventory covers at least three different economics, and on demand covers two. What separates them is who paid to create the records and whether that money was spent before a buyer committed.

A speculative catalog is the textbook inventory. Appen's off-the-shelf dataset page presents its catalog as data that can be licensed and used right away without a custom collection, under a standard license it describes as perpetual and non-exclusive with commercial training rights, so the vendor funds collection and recovers it across many licensees. A consigned catalog shifts that cost: Getty Images reports in its 10-K for 2025 that it pays royalties to its content contributors, a group that includes content partners, a royalty structure in which, for contributed content, the people who made it carry the production cost and are paid as it licenses.

A byproduct archive is inventory nobody built for AI. Informa's 2024 annual report describes the AI partnership revenue at its Taylor & Francis publishing business as non-recurring data access revenue, earned from a publishing archive built for the core business rather than for AI. On the on-demand side, commissioned collection produces new data to a buyer's specification after an order, while on-demand sourcing licenses records that already exist inside operating businesses once a buyer asks. The map below puts the five side by side:

Five ways a dataset reaches the shelf, and who paid for it
Sourcing modelWho pays before a buyer signsLicenses per assetNormalized metric to requestPublic illustration
Speculative catalogThe vendor: collection, rights and preparationMany, when licenses are non-exclusiveLicenses sold per held dataset, by vintageAppen off-the-shelf datasets, perpetual non-exclusive standard license
Consigned catalogContributors fund production; the platform funds curationManyRoyalty-adjusted contribution per licensed itemGetty Images, royalties paid to content contributors
Byproduct archiveNobody for AI: records were created for the core businessFew, buyer by buyerRecurring versus one-off AI revenueInforma, AI partnership revenue described as non-recurring
Commissioned collectionLittle: labor starts after the orderUsually one, the commissioning buyerLabor hours per accepted datasetCustom collection offered beside catalogs, as Appen's guidance describes
On-demand sourcing of existing recordsSourcing effort after a request; records already existOne first license, more only as agreements allowFill rate and capital committed before signatureSourceX: sourced on request, not held in stock

Where capital at risk shows up in the accounts

Capital at risk appears in different places depending on who created the records, and the accounts can hide the most important kind. Records an operating business produced for its own work usually never reach its balance sheet as an asset; money a data company spends acquiring or preparing datasets before any order is cash already gone, whether it is capitalized or expensed.

International accounting standards make the first point explicit for several kinds of self-created assets. IAS 38 paragraph 63 states that "Internally generated brands, mastheads, publishing titles, customer lists and items similar in substance shall not be recognised as intangible assets", and paragraph 64 explains that spending on them cannot be distinguished from the cost of developing the business as a whole. A publisher's archive or a field service company's job history can therefore be valuable and invisible at the same time: when records were never booked as an asset, there is no carrying amount to impair.

That asymmetry matters for diligence. A byproduct archive carries no AI-specific capital at risk, which is why archive licensing can look highly profitable in the year it happens. A speculative catalog carries capital at risk from the day collection starts, and an investor should find out how the company accounts for it and how it decides that held datasets are still worth their cost. An on-demand model commits cash mainly after a request, so its risk shows up as sourcing effort spent on requests that never close.

Speed only pays when the held dataset fits the request

A catalog's speed advantage exists only for the requests it already fits. Appen's own guidance frames the choice that way: off-the-shelf data for broad coverage of common categories or extra volume, custom collection for specific demographics, specialist domains or a controlled protocol, and a combination when a baseline set leaves gaps.

In the hurdle model, speed enters through the on-demand completion rate, a, the share of sourcing attempts that end in an accepted delivery. Buyers who cannot wait, sources that cannot be found and deliveries that fail acceptance all lower it. A low completion rate is the honest version of the inventory pitch: when on-demand sourcing often fails or arrives too late, stock becomes worth holding even at a modest fit rate.

The reverse also holds. When buyers in a domain specify fields, history depth, consent terms and acceptance tests that differ from one request to the next, held datasets rarely match without rework, and the speed advantage becomes a re-collection project with a catalog label. Speed should be measured on requests closed from stock as delivered, not on how fast a sample can be shown.

The fit-rate hurdle: when holding stock beats sourcing to order

Holding inventory is cheaper per accepted license exactly when the fit rate f exceeds a times c divided by k. Here c is the cost of building or acquiring a dataset ahead of demand relative to sourcing the same scope after a request, k is the number of licenses each fitting dataset sells, and a is the on-demand completion rate. The inventory cost per accepted license is c divided by f times k; the on-demand cost is 1 divided by a.

Illustrative worked example: if building ahead costs 0.7 of sourcing to order, each fitting dataset sells once, 40 percent of held datasets fit a request and 80 percent of on-demand attempts complete, inventory costs 1.75 per accepted license against 1.25 on demand, so sourcing to order wins. Let each fitting dataset sell three times and inventory falls to about 0.58, and stock wins comfortably.

The table shows the hurdle fit rate for four resale counts under three illustrative cost and completion settings, rounded half up to one decimal place. The figure shades the cheaper route for c of 0.7 and a of 1.0 across four fit rates. Every value is an assumption to replace with a company's own cohort data:

The fit-rate hurdle: when holding stock beats sourcing to order
Licenses per fitting dataset (k)c 1.0, a 1.0c 0.7, a 1.0c 0.7, a 0.8
1100.0%70.0%56.0%
250.0%35.0%28.0%
333.3%23.3%18.7%
425.0%17.5%14.0%
A grid of licenses per dataset, one to four, against fit rates of 20, 40, 60 and 80 percent, each cell marked as the cheaper route, inventory or on demand, with the hurdle fit rate 70, 35, 23.3 and 17.5 percent for each row.
Illustrative: cheaper route per accepted license when building ahead costs 0.7 of sourcing to order and every on-demand attempt completes.

Exclusive licenses cut the resale count to one

Exclusivity sets k, and in the table k moves the hurdle more than either other input. A dataset that can be licensed only once needs a fit rate above 70 percent in the middle column before stock beats sourcing to order; the same dataset licensed non-exclusively to four buyers needs 17.5 percent.

That logic explains why a catalog vendor would default to non-exclusive terms. Appen's description of its standard off-the-shelf license as perpetual and non-exclusive is the inventory model protecting its own resale count. Buyers who want exclusivity for training push the other way, and a catalog that grants exclusive licenses to win deals is quietly converting itself into an on-demand business with a stock of sunk costs.

Exclusivity can also sit on a different party. SourceX's published terms ask the supplying business not to license the selected data to anyone else for AI training, while stating that SourceX's right to distribute or sublicense the de-identified dataset is unaffected. In hurdle terms the supplier's k is one by design and the intermediary's k depends on each agreement, a structure an investor should read clause by clause rather than infer from a label.

Catalog or showroom? Testing for a hidden collection service

It can be either, and the test that tells them apart is cheap. The strongest alternative explanation for a catalog business that grows is that it is mostly a collection service: buyers see the catalog, ask for a variant, and the company re-collects to specification. Revenue then tracks labor hours, not licenses per held dataset, and the inventory economics in the pitch never occur.

Ask what share of revenue came from licenses of held datasets delivered unchanged, what share from modified catalog datasets, and what share from new collection, each with gross margin. A rising unmodified share supports the inventory story; a rising custom share means the catalog works as a showroom.

The same test runs in reverse on an on-demand company. If a sourced dataset is later licensed again under the original agreement, the on-demand business is accumulating inventory as a byproduct of orders, with no speculative capital behind it. That is the attractive corner of the map, and it should show up as second and later licenses per dataset, dated by cohort.

What would make the hurdle argument fail

The hurdle is a cost comparison, so anything that changes revenue, carrying cost or rights can overturn its verdict. These are the ways it most often misleads:

  • Price is left out. Buyers may pay more for immediate delivery or for exclusivity, which can favor either route; replace the cost-only hurdle with contribution per accepted license when prices are known.
  • Held data ages. Refresh, re-validation and storage are carrying costs the formula ignores, and records that describe old workflows can lose fit even while sitting unsold.
  • Archive windfalls repeat poorly. Informa's own description of its 2024 AI partnership revenue as non-recurring is a reminder that a byproduct archive can sell once at scale and then face a lower base.
  • Consignment moves risk, it does not remove it. A royalty model protects the platform's capital but depends on contributors continuing to supply, which a falling license volume can undermine.
  • Fit rates are not public. No company in this article publishes the share of held datasets that found a buyer, so every value in the hurdle table is illustrative and the verdict for any real business depends on cohort data obtained in diligence.
  • Rights terms can change k overnight. A buyer market that standardizes on exclusive training licenses would raise the hurdle for every catalog; one that accepts non-exclusive terms lowers it.

Running the hurdle on SourceX and on any data company

SourceX sits in the on-demand-sourcing row by its own description. Its request page says datasets are sourced on request, not held in stock; its how-it-works page says the supplying business keeps ownership of the original data while SourceX owns the de-identified dataset under the signed agreement, subject to the business's underlying intellectual property rights, and may distribute or sublicense it. The bet is that sourcing existing records after a request keeps c low, and that agreements leave room for k above one.

SourceX has not published fill rates, completion rates, capital committed before signature or second-license counts, and this article does not estimate them. Its investor page says the company shares company and commercial metrics with investors directly. The questions below apply to any data company and are the ones that turn the hurdle into evidence:

  • Fit rate: of datasets acquired or prepared before a buyer signed, what share found at least one accepted license, by quarter of acquisition, with unsold datasets kept in the denominator?
  • Resale count: licenses per fitting dataset, split by exclusive and non-exclusive terms, and how many datasets earned a second license.
  • Capital committed before signature: cash spent on collection, rights and preparation for datasets with no signed buyer at quarter end, and how impairment or write-off is decided.
  • Completion rate on demand: share of buyer requests that ended in an accepted delivery, with abandoned and open requests reported separately, not dropped.
  • Revenue mix: unmodified catalog licenses, modified catalog datasets and new collection, each with gross margin.
  • Rights: for each held dataset, the clause that allows another license, and whether any supplier exclusivity limits it.

Inputs, formula and blind spots behind the hurdle table

The hurdle comes from a short deterministic script, hurdle_table.py, that implements the formula below; a spreadsheet reproduces every value. It compares cost per accepted license: inventory c divided by f times k, on demand 1 divided by a, with inventory cheaper only when f is strictly greater than a times c divided by k. Hurdles are computed as exact fractions, capped at 100 percent and rounded half up to one decimal place. The figure grid uses c of 0.7 and a of 1.0, fit rates of 20, 40, 60 and 80 percent and resale counts of one to four; no cell is a tie. No random numbers are used.

Public evidence was checked on 27 October 2026: Appen's off-the-shelf dataset page, Getty Images' 10-K for 2025, Informa's 2024 annual report and the IFRS text of IAS 38 through searches limited to each publisher's domain, and SourceX's request, how-it-works and investor pages by direct fetch. The examples illustrate models and are not a ranking or a measure of any company's performance.

Not measured: real fit rates, completion rates, prices, carrying costs or license counts for any company, SourceX included. The model treats every on-demand attempt as costing the same whether or not it completes, ignores timing of cash, and assumes one dataset scope equals one comparable unit across models. The next evidence needed is cohort data on held datasets and on requests, which only companies can supply.

Frequently asked questions

What is the difference between data inventory and on-demand data sourcing?

Inventory means datasets are acquired or prepared before a buyer commits and then licensed from stock. On-demand sourcing starts after a buyer states a requirement, either by commissioning new collection or by licensing records that already exist in operating businesses. The difference is when money is committed relative to a signed buyer.

When does holding AI training data inventory make sense?

When held datasets fit a high share of requests and can be licensed several times. In the illustrative hurdle, a dataset that sells to four buyers beats sourcing to order above a 17.5 percent fit rate, while one that sells once needs 70 percent under the same costs.

Why do catalog dataset licenses tend to be non-exclusive?

Because the inventory model recovers its upfront cost across licensees. An exclusive license caps the resale count at one, which raises the fit rate a catalog needs to cover what it spent before any buyer signed.

Is a data company with a catalog always an inventory business?

No. If most revenue comes from modifying catalog datasets or collecting new data to specification, revenue tracks labor rather than licenses per held dataset. Ask for revenue split into unmodified licenses, modified datasets and new collection, each with margin.

Which metrics show whether an on-demand data company is working?

The completion rate on requests with abandoned and open requests reported, capital committed before signature, and second and later licenses per sourced dataset where agreements allow them. Each should be reported by cohort with a stated denominator.

Sources

  • Appen presents off-the-shelf datasets as licensable without custom collection, under a standard license it calls perpetual and non-exclusive, and advises when to choose catalog or custom data. Source
  • Getty Images reports paying royalties to its content contributors, including content partners, in its 10-K for 2025. Source
  • Informa describes the 2024 AI partnership revenue at Taylor & Francis as non-recurring data access revenue. Source
  • IAS 38 paragraphs 63 and 64 bar recognizing internally generated customer lists, publishing titles and similar items as intangible assets. Source
  • SourceX states that datasets are sourced on request, not held in stock. Source
  • SourceX's terms on ownership of original and de-identified data, supplier exclusivity for AI training and its right to sublicense. Source
  • SourceX 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