Skip to content

Deal economics

Dataset inventory procurement risk: what to build before acceptance

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Dataset inventory procurement risk is the preparation work a data company pays for before any buyer has accepted the result. Building ahead looks fast, but it strands hours whenever the next buyer's specification differs, and specifications do differ: OpenAI's data partnerships page rules out personal and third-party information. Fund portable work early, buyer-specific work only after a signed order, and ask any data company for its stranded share by cohort.

Checklist

Dataset inventory

This page covers dataset inventory procurement risk: what to build before acceptance.Illustrative

Key takeaways

  • Inventory risk in AI data is specification risk: work done before a buyer fixes its requirements in writing may not fit them.
  • Source mapping and rights review travel between buyers; buyer-specific labels, splits and formats almost never do.
  • In this article's illustrative ten-candidate model, building ahead costs 1,120 hours per accepted dataset against 375 for staged qualification, and saves 80 hours of work after signature.
  • On hours alone a speculative build at best breaks even, and only when every candidate sells and no buyer asks for changes; what it really buys is speed.
  • A shelf built for one category inherits that category's buyer concentration, and public filings show how concentrated AI data buyers can be.

What is dataset inventory procurement risk, and what evidence settles it?

Dataset inventory procurement risk is the chance that preparation work a data company funds before a buyer accepts the result never becomes an accepted, paid delivery. Inventory here means datasets prepared and held for sale, not the catalog of systems and records a company keeps about itself; the guide to building a business data inventory covers that second meaning.

The investor's decision is narrow: is the company spending ahead of its buyers, and if so, is it buying speed it can sell or stocking a shelf that ages? A pitch about demand cannot answer that. Three records can: the buyer stage each preparation hour was spent against, the share of hours that went to candidates that never sold, measured by cohort, and the time from signature to delivery that the early spending was meant to buy.

The rest of this article builds those measures. It defines the buyer stages that justify spending, sorts six preparation layers by when each becomes buyer-backed, compares a speculative and a staged build on ten illustrative candidates, and then tests the main argument for building ahead anyway: a large buyer or mandate that seems to make demand certain.

Specification lock: which buyer stages justify which spending?

Spending is justified by how much of the buyer's specification is fixed in writing, not by how keen the buyer sounds. Each stage below locks more of what the dataset must be, and each lowers the chance that work done now will be redone or thrown away.

Stage one is category interest: a public statement, a mandate or a conversation saying a buyer wants a kind of data. Stage two is a written request naming the task, volume, format and timing. Stage three is a written acceptance test, the check a delivery must pass. Stage four is an accepted sample. Stage five is a signed order or license that carries the scope and the acceptance criteria. Stage six is acceptance of the full delivery.

Acceptance is the stage that matters for revenue, and accounting treats it as a real event. The FASB's revenue standard, ASU 2014-09, lists customer acceptance among the indicators that control of an asset has passed to the customer, and it points companies to dedicated guidance on acceptance clauses when they decide when that transfer happens. Commercial law orders the same steps for goods: under UCC section 2-606, a buyer accepts after a reasonable opportunity to inspect, by signaling that the goods conform or by failing to reject them effectively. A data license is usually not a sale of goods, so the code is an analogy here, but its order of inspection first and acceptance second is the order buyers follow.

Until stage six the seller carries the inventory. The useful question for each preparation layer is therefore which stage makes it safe to fund.

The build-ahead ledger: six preparation layers and when each is backed

Fund a layer ahead of the buyer only when the work travels to the next buyer or costs little; fund everything buyer-specific after a signed order. The ledger applies that rule to six layers of preparation found in most business-data deals.

Two properties decide each row. Portability asks whether a second buyer could use the work unchanged. Perishability asks what makes the work lose value while it waits: a new model generation, a changed specification, staler records or a lapsed approval. Work that is both portable and slow to perish is the only work that belongs on a shelf.

The ledger gives the stage at which each layer becomes buyer-backed and a verdict on building it early:

The build-ahead ledger: six preparation layers and when each is backed
Preparation layerBuyer-backed fromUsable by the next buyer?Loses value whenBuild-ahead verdict
Source mapping: systems, years, volumes and ownersStage one: category interestYes, almost unchangedSystems are migrated or records are purgedBuild ahead; it is cheap and every buyer asks for it
Rights review and the supplier's approval of scopeStage two: a written request in the categoryOften, if the license allows more than one buyerContracts, notices or the approved scope changeBuild ahead only for scopes several buyers could license
Sample extraction and de-identificationStage two, in the request's formatPartly; sample rules differ by buyerThe buyer changes its format or exclusion rulesBuild on a written request
Full extraction, linking and de-identificationStage five: a signed order with acceptance criteriaPartly; joins and exclusions follow one taskThe task, time window or exclusions changeBuild on a signed order
Buyer-specific labels, splits and formatsStage five: a signed orderRarelyThe buyer's model, rubric or evaluation program changesNever build ahead
Documentation: datasheet, provenance and acceptance evidenceGeneric part at stage two; buyer part at stage fiveThe generic part, yesProvenance cannot be shown for a sourceBuild the generic part ahead; finish it per order

Speculative or staged: ten candidate datasets, two build paths (Illustrative)

On the illustrative inputs below, building ahead costs about three times as many hours per accepted dataset as staged qualification and saves 80 hours of work after signature. Every input is an editable assumption, not a measurement of any company.

Take ten candidate datasets in one category. Layer hours are 10 for source mapping, 30 for rights review, 20 for a sample, 120 for full preparation, 60 for buyer-specific work and 20 for documentation. Five candidates receive a written request, three pass their sample and two reach a signed order and acceptance. In the speculative path the company prepares all ten through full preparation and generic documentation before any request arrives, and each of the two buyers then asks for rework on half of the full preparation.

The staged path spends 750 hours for two accepted datasets: 100 mapping all ten, 250 on rights and samples for the five requests and 400 finishing the two signed orders. The speculative path spends 2,000 hours before any request and 240 after signature, 2,240 in all. Its 1,600 hours on the eight candidates that never sold include 320 hours of mapping and rights review that a later buyer might reuse. Per accepted dataset, that is 800 stranded hours and 320 used hours in the speculative path, against 115 stranded and 260 used in the staged one.

Two results hold beyond these numbers. With the same hours per layer, the staged path can never spend more than the speculative one; they tie only when every candidate sells and no buyer asks for changes, at 2,600 hours each in this model. And the speculative path's one real product is speed, less work left once a buyer signs, so building ahead is a bet that buyers pay for that speed and that their specifications stay put.

The comparison, including the stranded share an investor should ask a company to report by cohort:

Speculative or staged: ten candidate datasets, two build paths (Illustrative)
MeasureSpeculative buildStaged build
Hours spent before any written request2,000100
Hours on candidates that never sold1,600, or 71 percent of all hours230, or 31 percent of all hours
Total hours for two accepted datasets2,240750
Hours per accepted dataset1,120375
Work left after signature, per dataset120 hours of rework and buyer-specific work200 hours of full preparation, buyer-specific work and documentation
What the path is betting onSpecifications hold and buyers pay for speedBuyers will wait for about 80 more hours of work
Two stacked bars of hours per accepted dataset: the speculative build at 1,120 hours, 800 of them stranded, and the staged build at 375 hours, 115 of them stranded.
Illustrative: hours per accepted dataset from the article's ten-candidate model; every input is an assumption, not a measurement.

Why a shelf built for one category inherits its buyers' concentration

A dataset built ahead for a category is exposed to every buyer in that category at once, so its risk is only as diversified as the set of buyers whose specifications it already meets. In a category with two serious buyers, a shelf of ten datasets is two bets, not ten.

Public filings show how few buyers can account for AI data revenue. Reddit's quarterly report for the period ended March 31, 2026 says that, to date, "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." Innodata's annual report for 2025 says one customer in its Digital Data Solutions segment generated approximately 58% of total revenues in 2025, up from 48% in 2024, and no other customer reached 10%. Neither company stocks datasets in this article's sense, but both show the buyer counts a category-level shelf would face.

Specifications concentrate the risk further. OpenAI's data partnerships page says the company is not looking for datasets containing sensitive or personal information, nor for information a third party owns. A dataset prepared before an exclusion like that was known may contain exactly the material a large buyer refuses, and the cost of taking it out falls on the inventory, not on the buyer.

Provenance is the third correlated exposure. In an audit covering more than 1,800 text datasets, the Data Provenance Initiative found licenses missing at rates above 70% on widely used dataset hosting sites. Inventory assembled before its rights were documented carries that gap into every sale, and each buyer's diligence finds it again. Three checks before counting a shelf as diversified:

  • Count the independent buyers whose written specifications each dataset already satisfies, not the buyers active in the category.
  • Check each dataset against every known buyer exclusion, such as personal data, third-party material and restricted sources.
  • Confirm documented rights for every source before preparation starts, not when a buyer asks.

Could a large buyer or mandate justify building ahead?

Sometimes, but a mandate is a reason to qualify candidates faster, not a purchase of any particular dataset. The strongest case for building ahead is that a large, credible buyer has announced demand, and a supplier that waits for written specifications will arrive late.

Announcements of this kind are real. In an October 2026 blog post, micro1 says it plans to put $1 billion into acquiring and licensing operational data from enterprises over the next 12 months, describing that data as material for reinforcement learning environments that mirror real workflows. That is a stated plan with a stated horizon. It does not say which datasets, in what format or judged by which test, and it is not an order for any particular supplier's records.

The same question has shadowed AI infrastructure. Sequoia Capital's 2023 essay AI's $200B Question asked how much of the GPU build-out reflected real end-customer demand and how much was capacity built in anticipation of it. A dataset shelf is a small version of that bet with one difference that argues for caution: a GPU can serve any model, while a prepared dataset serves the specification it was built to.

A large buyer can also create dependence without buying anything. A supplier that builds to one buyer's draft specification has spent money only that buyer can repay, which hands the buyer leverage on price and scope at the moment the supplier most needs the order. Building ahead can still be the better call when these conditions hold:

  • Buyers have said in writing that delivery time decides the award, and the speed bought is worth more than the hours at risk.
  • The work is portable: at least two independent buyers' written specifications already accept it.
  • Waiting would lose the records, for example a system due for retirement, and preserving them costs less than preparing them.
  • The supplier's approval or license allows the dataset to be offered to more than one buyer.

What could make the staged-build case wrong

The staged rule can fail in five ways, and an investor should test each against the company's own records rather than treat the rule as settled.

  • Speed is worth more than the model assumes. If buyers award work to whoever can deliver inside a training or evaluation window, staged qualification loses deals, and a speculative shelf's stranded hours may be a fair price.
  • Batch economies exist. Preparing ten datasets together may cost less per dataset than ten separate runs, which fixed hours per layer ignore.
  • Supply perishes while the company waits. A supplier's willingness, a key employee or the records themselves can disappear between the first conversation and a signed order.
  • Specifications converge. If buyers adopt shared formats and acceptance tests, more preparation becomes portable and the ledger's verdicts loosen.
  • The illustrative inputs are wrong. Rework rates, conversion and layer hours vary by category; a company's own cohorts should replace every number in the two-path model.

Questions to ask about inventory built ahead of acceptance, SourceX included

Ask for records, not reassurance. Each question below has an answer a company either keeps or does not, and a missing record is itself an answer.

SourceX's public description puts it at the staged end of this range: it keeps no stock of datasets for AI buyers, sourcing from U.S. businesses only when a request comes in, and no data is delivered until the supplying company has signed the agreement on a scope and terms it approved. That design limits stranded preparation by construction; the trade-off to examine is the time from a buyer's request to delivery, which SourceX has not published. Its conversion and stranded-share figures are not public either; SourceX shares company and commercial metrics with investors directly through its investor page.

Questions for any company preparing datasets for AI buyers:

Checklist · 6 items

  1. What share of preparation hours in each quarterly cohort was spent before a written request, and what share before a signed order?
  2. What share of hours went to candidates that never reached acceptance, and how much of that work was later reused?
  3. For each dataset on hand, how many independent buyers' written specifications does it already satisfy?
  4. Which buyer exclusions, such as personal data or third-party material, has each dataset been checked against?
  5. How much work remains after signature, and what is that speed worth to the buyers who signed?
  6. Does each supplier agreement allow the dataset to be offered to more than one buyer?

0 of 6 ticked. Tick each item as you go.

Methods note: how the ledger and the two-path model were built

The specification-lock stages, the build-ahead ledger and the two-path comparison are editorial frameworks built for this article on 2026-10-26. Stages are ordered by how much of a buyer's specification is fixed in writing. Ledger verdicts apply one rule, fund early only what is portable or cheap, to six preparation layers common in business-data deals; they are judgments, not measurements.

Every number in the two-path comparison is an illustrative input or arithmetic on those inputs: the layer hours, the funnel from ten candidates to five requests, three passing samples and two acceptances, and rework on half of full preparation for each signed buyer. No company's hours, conversion or rework rates were measured, and the model leaves out batch economies, waiting costs and the money value of speed.

Public evidence comes from the cited filings, standards and announcements, checked on 2026-10-26. Not measured: stranded shares or conversion rates at any data company, how often buyer specifications change after a written request, what buyers pay for speed, and any SourceX buyer, supplier or preparation data. SourceX has not published results for these measures.

Frequently asked questions

Is it ever smart to prepare AI datasets before a buyer signs?

Yes, for work any buyer would use, such as mapping which records exist and reviewing whether they can be licensed. Buyer-specific preparation, such as labels, splits and formats, should wait for a signed order unless buyers have said in writing that delivery speed decides the award. The test is portability: could a second buyer use the work unchanged?

What is a stranded share in a data business?

It is the share of preparation hours spent on candidate datasets that never reached an accepted delivery, measured for a cohort of candidates that entered in the same period. Report reused hours separately and keep open candidates in the denominator. A rising stranded share means the company is building ahead of its buyers.

How does buyer acceptance affect revenue for a data company?

Acceptance is often the event that turns delivered work into revenue. The FASB's revenue standard lists customer acceptance among the indicators that control has passed to the customer and gives specific guidance on acceptance clauses. Until a buyer accepts, the seller carries the cost of the work.

Does a large buyer's public data budget reduce inventory risk?

Not by itself. A public plan, such as micro1's stated plan to spend $1 billion over 12 months acquiring enterprise operational data, signals category interest but does not say which datasets, in what format or against which acceptance test. It justifies qualifying candidates faster, not preparing them in full.

How is dataset inventory different from a business data inventory?

A business data inventory is a catalog of the systems and records a company holds. Dataset inventory, in this article, means datasets already prepared and held for sale before a buyer accepts them. The first is cheap and useful to any buyer; the second ties up work that only a matching buyer can repay.

Sources

  • Reddit's Form 10-Q for the quarter ended March 31, 2026 says that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
  • Innodata's Form 10-K for 2025 says one Digital Data Solutions customer generated approximately 58% and 48% of total revenues in 2025 and 2024, with no other customer at 10% or more. Source
  • OpenAI's data partnerships page says it is not seeking datasets with sensitive or personal information, or information that belongs to a third party. Source
  • The Data Provenance Initiative audit of 1,800+ text datasets reports license omission of 70%+ on widely used dataset hosting sites. Source
  • ASU 2014-09 lists customer acceptance among the indicators of transfer of control and directs entities to paragraphs 606-10-55-85 through 55-88 to evaluate acceptance clauses. Source
  • UCC section 2-606 provides that a buyer accepts goods after a reasonable opportunity to inspect, by signifying that they conform or by failing to make an effective rejection. Source
  • micro1's October 2026 blog post states a plan to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data for reinforcement learning environments. Source
  • Sequoia Capital's 2023 essay AI's $200B Question asks how much of the GPU build-out reflects real end-customer demand versus capacity built in anticipation of future demand. 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