Deal economics
Latent dataset commercial value: three tests before it counts
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Yes. A dataset can be useful to a model yet unsaleable today, because a sale needs three conditions at once: technical usefulness, a lawful path to license and keep the records, and a buyer whose use pays for the work now. Latent commercial value is an option on the missing condition. Count it at its chance of clearing, net of decay and carrying cost, never at headline value.
Value drivers
Latent dataset commercial
Key takeaways
- IFRS 13's highest and best use test maps onto data: physically possible (useful), legally permissible (licensable and keepable) and financially feasible (a buyer now).
- IFRS 5 sets the bar for saleable today: available for immediate sale in its present condition, with a sale highly probable.
- Usefulness decays with time at a rate that depends on the task, so latent value shrinks while it waits.
- Personal data with no live purpose cannot be kept just in case a market appears; that condition is fatal, not fixable.
- In the illustrative worksheet, value today is 93.0 units against a 260-unit headline, 35.8 percent.
Can a dataset be valuable without being saleable today?
It can. Usefulness to a model is one condition of a sale, not the whole of it: the records must also be lawful to license and to keep, and some buyer's use has to pay for the work this quarter. When any of the three is missing, the value is latent, which makes it real but turns it into an option rather than inventory.
Accounting has a precise test for the opposite state. IFRS 5, the standard on assets held for sale, lets an asset carry that label only if it is "available for immediate sale in its present condition subject only to terms that are usual and customary", and only if the sale is highly probable: management committed to a plan, a buyer actively sought, a price reasonable against current fair value and completion expected within a year (paraphrased). Most business archives fail at least one clause. That does not make them worthless; it makes their value conditional.
The decision here is how much of a data holder's claimed value an investor or strategic buyer should count today, and what evidence would move the rest into the count. A useful answer names the condition that blocks each dataset, the evidence that would clear it, the chance and timing of clearing, what the wait costs and what the records lose while they wait.
Possible, permissible, feasible: three tests from fair value practice
Fair value practice already splits value into the three conditions a dataset needs. IFRS 13 says the highest and best use of a non-financial asset "takes into account the use of the asset that is physically possible, legally permissible and financially feasible", and that a feasible use is one that generates enough income or cash, after the costs of converting the asset to that use, to produce the return market participants would require (paraphrased).
Translated to data, physically possible means technical usefulness: the records pass a buyer's acceptance test for a named task. Legally permissible means transaction conditions: the holder may license the records for that use and may keep them long enough to do so. Financially feasible means market timing: a buyer exists now whose use pays for preparation, review and delivery. A dataset that passes all three is saleable; one that fails any of them holds latent value at best.
IFRS 13 adds a discipline that matters for investors: highest and best use is judged from the perspective of market participants, even if the holder intends something else (paraphrased). A founder's conviction that an archive will matter one day does not satisfy the third test. A buyer's written request naming a task does.
Technical usefulness has a clock
Usefulness is not a fixed property of records; it decays as the world moves away from the period they describe, at a rate that depends on the task. Latent value therefore shrinks while it waits, and any estimate has to say how fast.
Language model research measures the effect directly. Luu and colleagues built eight tasks across social media, science papers, news and reviews, each spanning five years or more, and reported "stronger effects of temporal misalignment on task performance than have been previously reported" (NAACL 2022, arXiv 2111.07408). They also found that continued pretraining on newer text recovered only a small part of the loss compared with task-specific finetuning on data from the target period (paraphrased). For a holder, the lesson cuts both ways: recent labeled records in a period a buyer cares about are hard to replace, and old ones lose ground every year they sit.
Decay differs by record type. Pricing quotes for a discontinued product line lose relevance fast; engineering change histories for long-lived equipment age slowly; procedures that have not changed may hardly age at all. The worksheet below gives each dataset its own annual decay rate rather than one rate for the archive.
Transaction conditions: the fixable and the fatal
Some transaction conditions can be fixed at a known cost; others mean the records cannot be held for a future sale at all. Telling the two apart is the most valuable step in a latent-value review, because a fixable condition is an investment decision and a fatal one is a deletion schedule.
Fixable conditions include customer contracts that are silent on outside use, missing assignments from contractors, absent documentation and exports that lose links between records. Each has a route: narrowing the scope, asking the counterparty, recovering paper or rebuilding the export, and each route has a cost and a chance of working.
The fatal kind often involves personal data kept with no live purpose. The UK Information Commissioner's Office guidance on storage limitation says organizations must be able to justify how long they keep personal data and should "periodically review the data you hold, and erase or anonymise it when you no longer need it." Its guidance on the research provisions adds, in paraphrase, that the general rule is that personal data cannot be held indefinitely just in case it proves useful later. A hoped-for future sale is not a purpose that keeps an expired record alive. Anonymising can create a different, smaller dataset with value of its own, a route this site's guide on keeping de-identified records after a retention period covers; the original records still leave the latent column.
A prior license can be fatal too. A dataset already licensed exclusively for AI training cannot be offered again for that purpose during the term, and on SourceX's own terms records already licensed for AI training are not eligible.
No buyer yet is a timing problem, not a verdict
A dataset with no buyer today fails the feasibility test today, not forever. The right treatment is to hold it as an option with a review date and a named trigger, while keeping the cost of holding it below what the option is worth.
Archives built for one purpose do acquire licensing value when buyers arrive. Stack Overflow announced on May 6, 2024 that OpenAI would use its OverflowAPI product as part of a partnership between the two companies (paraphrased from the announcement), so a question-and-answer archive built up over years for developers became the basis of an API arrangement with an AI developer. The records did not change; the market around them did.
What turns a timing problem into a sale is a buyer whose use clears the feasibility test: a written request that names a task, an acceptance test the records can pass and a budget that covers preparation. Signals worth tracking include buyer requests for the record type, model capabilities that newly depend on such records and rules that raise the value of documented sources. None of them is a forecast, and the worksheet treats each latent dataset's chance of finding a buyer as an input to be argued, not a fact.
The latent value worksheet: five datasets, one rule
The worksheet values each dataset today as the chance its blocking condition clears, times its value once cleared, minus the cost of carrying it until then, and sets value today to zero where the records may not lawfully be kept. Value once cleared is the value if saleable today, reduced by its annual decay for each year of waiting.
Five invented datasets, one per outcome, show how the rule sorts an archive. Across them the headline value, counting every dataset as if it were saleable now, is 260 value units (U); value today is 93.0 U, or 35.8 percent of the headline. A value unit is an editorial device, not a currency. The worksheet, with each dataset's inputs (Illustrative):
| Dataset (Illustrative) | Test it fails | Inputs | Value today (headline) | Decision |
|---|---|---|---|---|
| A. Work orders, rights clear, buyer request on file | None | Saleable now | 50 U (50 U) | License it now |
| B. Support tickets, customer contracts silent on outside use | Legally permissible, fixable | 1 year to fix, 10% decay a year, 0.6 chance, 4 U to carry and fix | 28.4 U (60 U) | Fix the condition now |
| C. Engineering change histories, no buyer task yet | Financially feasible | 3 years to a buyer, 5% decay a year, 0.3 chance, 6 U to carry | 14.6 U (80 U) | Hold as an option with a review date |
| D. Call recordings past their stated retention period | Legally permissible: may not be kept | No lawful purpose left to keep them | 0 U (40 U) | Delete or anonymise on schedule |
| E. Quotes for a discontinued product line | Physically possible: usefulness decayed | 2 years, 30% decay a year, 0.1 chance, 3 U to carry | 0 U (30 U) | Stop carrying it |
Is the headline value a fair stand-in for latent value?
No. A headline that counts every dataset as saleable sets every probability to one, every decay rate to zero and every carrying cost to nothing, and in the worksheet it overstates value today almost threefold: 260 U against 93.0 U.
The usual sources of a headline make it worse. Asking prices and published estimate ranges record what holders hope for, not what buyers paid, and for a latent dataset they price a sale that cannot happen yet. Stacking a hoped-for price on a hoped-for buyer compounds two unrealized assumptions. The worksheet's value if saleable today should come from executed terms for comparable rights, or from a buyer's acceptance test, and its probability from evidence about the specific blocking condition.
The strongest form of the alternative is that latent datasets are cheap to hold, so the headline is harmless optimism. Dataset E shows why it is not: at a 0.1 chance and 30 percent annual decay, two years of carrying cost exceed the expected value, so every year it stays on the books it destroys value. Dataset D shows the other limit, where the law, not the arithmetic, sets the answer.
What could make a latent-value estimate wrong
Six things can move a latent-value estimate, in either direction, and each can be checked.
- The probabilities are judgments. A 0.3 chance of finding a buyer within three years is an argument, not a measurement; show the value at the low and high ends of the plausible range.
- Decay can be misjudged. Some record types age faster than expected when a workflow or product changes; others hold value for a decade. Ask a buyer which period its task needs before setting a rate.
- Markets can arrive early or never. A new capability can create demand for records nobody wanted, and a substitute such as synthetic data can remove demand that seemed certain. Revisit dataset C at every review date.
- Rules change. Privacy and contract terms can tighten what may be kept or licensed, or a counterparty can grant a permission that was missing. Treat the legal inputs as dated.
- Carrying cost is easy to understate. Storage is the small part; access control, security exposure and periodic rights review grow with the archive, and records kept past their purpose add risk as well as cost.
- Deletion is irreversible. Stopping the carry of dataset E ends its option for good, so record the decision and its reason, and keep the inventory entry that describes what was removed.
Questions for any company claiming latent data value, SourceX included
Ask any company that claims latent data value to show the worksheet behind the claim, and put the same questions to SourceX.
SourceX's request-data page says datasets are sourced on request, not held in stock, so latent supply in its market sits with the companies that hold the records, not on SourceX's books. Its how-it-works page lists what a supplier confirms before a deal, among them its right to license the records, lawful collection of any personal data in them and the absence of any earlier license of the records for AI training, and its FAQ says licenses are exclusive for AI training, with the term set in each deal's Dataset Deal Confirmation. Those are transaction conditions in this article's sense: a dataset that fails one is not saleable through SourceX today, whatever its usefulness.
SourceX publishes no figures on how much supply passes or fails those conditions, and its investors page reserves company and commercial metrics for investors. Its fundraising is a Rule 506(c) offering open only to accredited investors, and requests for materials go through that page. Questions that apply to any company in this market:
- For each dataset you count, which of the three tests does it pass today, and what document shows it?
- For each blocked dataset, what is the condition, who can clear it, what does clearing cost and by when?
- What annual decay rate do you assume for each record type, and what buyer evidence supports it?
- Which records are you keeping only because they might be licensed one day, and what lawful purpose covers them?
- How much of your stated value would remain if every probability, decay rate and carrying cost were set to your own best estimate?
- Does any part of the latent figure come from asking prices or estimate ranges, and is it labeled that way?
Methods note: the worksheet rule, its inputs and its limits
The three-test mapping, the worksheet and its decision rule were built for this article on 2026-10-28 and computed by a short Python script (xs0216_latent.py, kept with its output in the writer's working folder). It uses exact fractions, rounds each printed figure half up once at output to one decimal place and uses no randomness. Every input is an editorial example in value units.
Rules as the script applies them. Value once cleared = value if saleable today x (1 - annual decay) raised to the number of years until the condition clears. Value today = 0 where the records may not lawfully be kept; otherwise the larger of zero and (chance of clearing x value once cleared - carrying cost). The decision is the first match of: may not be kept, delete or anonymise; no failing test, license now; chance x value once cleared - carrying cost at or below zero, stop carrying; failing the legal test, fix the condition; failing the feasibility test, hold as an option with a review date; failing the usefulness test, re-test usefulness. No dataset in the worksheet reaches the last rule, because dataset E's arithmetic stops the carry first. Headline value is the sum of the values if saleable today, and the 35.8 percent share is the sum of values today divided by that headline.
Public sources, each checked on 2026-10-28 through a search restricted to its publisher's domain: IFRS 13 and IFRS 5 on the IFRS Foundation's site, the ICO's storage limitation and research provisions guidance, Luu and colleagues on arXiv and Stack Overflow's May 2024 announcement; passages without quotation marks are paraphrases. SourceX's how-it-works, FAQ, request-data and investors pages were fetched directly. Not measured: the share of any real archive in each state, actual decay rates, probabilities or carrying costs, and any SourceX figure.
Frequently asked questions
What is latent dataset commercial value?
It is value a dataset would have if a missing condition were met: a fixable rights gap, a buyer who does not exist yet, or usefulness that still needs proving. It should be carried as an option, at the chance the condition clears, net of decay and carrying cost.
Should a company keep old records in case AI buyers want them later?
Only where it has a lawful reason to keep them anyway and the expected value exceeds the cost of holding them. Personal data with no remaining purpose generally cannot be kept just in case it becomes useful, according to the UK ICO's guidance, so it belongs on a deletion or anonymisation schedule.
How is latent value different from asking price?
An asking price is what a holder hopes a buyer will pay for a sale that could happen now. Latent value prices a sale that cannot happen yet, so it starts from the value if saleable and then discounts for the chance, the wait, decay and carrying cost.
Can a latent dataset become saleable quickly?
Yes, when the blocking condition is fixable or a buyer arrives with a named task. A dataset blocked only by silent customer contracts can move fast if the scope is narrowed, while one waiting for demand moves only when a buyer appears.
Does SourceX buy datasets to hold until demand appears?
No. Its request-data page says datasets are sourced on request, not held in stock, and its public pages give no figures on supply or prices. Company and commercial metrics are shared with investors directly.
Sources
- IFRS 13: highest and best use takes into account uses that are physically possible, legally permissible and financially feasible, judged from market participants' perspective. Source
- IFRS 5: held for sale requires availability for immediate sale in present condition, subject only to usual and customary terms, and a highly probable sale. Source
- ICO storage limitation guidance: justify retention periods, review data periodically and erase or anonymise it when no longer needed. Source
- ICO research provisions guidance: the general rule is that personal data cannot be held indefinitely just in case it may be useful in future. Source
- Luu et al. (NAACL 2022): across eight tasks spanning five years or more, temporal misalignment hurt task performance more than previously reported. Source
- Stack Overflow announced on May 6, 2024 that OpenAI would use its OverflowAPI product as part of a partnership. Source
- SourceX's how-it-works page lists what suppliers confirm, including rights to license, lawful collection of personal data and no prior license for AI training. Source
- SourceX's FAQ says licenses are exclusive for AI training, with the term set in the Dataset Deal Confirmation, and data cannot be licensed again for AI training once licensed. Source
- SourceX's request-data page says datasets are sourced on request, not held in stock. Source
- SourceX's investors page says company and commercial metrics are shared 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.