AI data market
Four AI data business downside risks that time will not repair
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Most AI data business downside risks are temporary: a slipped deal or a paused buyer delays cash but leaves the asset intact. Four events can destroy value for good: losing the rights or supply behind the records, a legal or regulatory bar on a data category, a substitute that passes the buyer's acceptance test, and concentration that ties the asset to one buyer. Check each against documents, not forecasts.
License terms
AI data business downside
Key takeaways
- Ask of every risk whether the asset is whole when the event ends; if not, model a cut to its size or life, not a delay.
- In the illustrative model, losing a full year of cash costs 17.3% of value, while rights that end after year four cost 43.1%.
- Rights failures can arrive through suppliers, as the Ross opinion's account of memos built from Westlaw headnotes shows.
- Legal bars that target a category, or order deletion of derived products, close markets instead of delaying them.
- Concentration and substitutes impair for good only when the records would not sell elsewhere or a cheaper source passes the buyer's own test.
Which AI data business downside risks are permanent?
Only the ones that remove something time, money or effort cannot restore: the right to use the records, the supply of new records, the legal permission to sell a category, or the buyer's need for it. A deal that slips, a buyer that pauses or a review that runs long hurts this year's cash, and the asset is still there when the episode ends.
Accounting draws the same line for intangible assets. IAS 38 says economic factors set the period over which benefits will be received, "while legal factors may restrict the period over which the entity controls access to these benefits," and that the useful life "is the shorter of the periods determined by these factors." Renewal periods count only when there is evidence the entity can renew without significant cost. A data licensing book is a bundle of such rights, so losing one does not dent the book's value for a season; it shortens its life.
That gives the test this memo applies to each risk: when the event ends, is the asset whole? If it is, model a delay. If it is not, model a cut to the asset's size or its life, and look for the documents that show how likely the cut is.
How much more does a permanent loss cost than a temporary one?
Usually far more, even when it looks smaller on the day. In an illustrative ten-year licensing book discounted at 15% a year, losing all of year one's cash costs 17.3% of value, losing 30% of cash for good costs 30.0%, and rights that end after year four cost 43.1%. The dramatic shock is the cheaper one.
Yearly cash is held at 100 index points, so the percentages describe shape rather than scale. A longer horizon shrinks the temporary losses, leaves a proportional cut where it is and deepens the cases where rights or a buyer disappear, which is why diligence time belongs on the four permanent events rather than on next quarter's slippage. The five illustrative cases:
| Case (illustrative) | Kind | Yearly cash, index points | Share of value lost |
|---|---|---|---|
| No cash in year 1, full recovery after | Temporary | 0, then 100 a year | 17.3% |
| Half cash in years 1 and 2 | Temporary | 50, 50, then 100 a year | 16.2% |
| 30% of cash lost for good | Permanent | 70 a year | 30.0% |
| Rights end after year 4 | Permanent | 100 for four years, then 0 | 43.1% |
| Buyer with half the book leaves after year 2 | Permanent | 100, 100, then 50 a year | 33.8% |
Rights and supply: when the chain of title breaks
A rights failure is permanent when it reaches records already delivered or the source of future ones, and it can arrive through a supplier. In Thomson Reuters v. Ross Intelligence, the court's February 2025 opinion records that LegalEase "sold Ross roughly 25,000 Bulk Memos, which Ross used to train its AI search tool," memos that "in turn were built from Westlaw headnotes." The court found copying of 2,243 headnotes and weighed the first fair-use factor for Thomson Reuters.
Supply loss is the slower version. A business that depends on suppliers renewing on similar terms holds rights with end dates, and the IAS 38 renewal test applies: count future years only where there is evidence of renewal without significant cost. A supplier that stops renewing is a temporary problem if others hold equivalent records, and a permanent one if it held the only usable history.
Evidence to check: the chain of title from each supplier back to the systems the records came from, warranties about third-party content, termination and renewal clauses, and the number of independent suppliers whose records would satisfy the same buyer requirement.
Legal or regulatory exclusion: when a data category closes
A legal bar is permanent when it targets the category rather than one company's paperwork. The FTC's final order against Avast in June 2024 bans it from selling or licensing web browsing data for advertising, after its subsidiary Jumpshot sold browsing information from 2014 to 2020 to clients including advertising, marketing and data analytics companies and data brokers. The order also requires deleting the data passed to Jumpshot along with products or algorithms derived from it.
Rules can close a category for everyone at once. The European Commission lists untargeted scraping of facial images from the internet or CCTV footage to create or expand facial recognition databases among the AI Act's prohibited practices, which have applied since 2 February 2025. A business whose asset was such a database faced not a delay but an EU market that no longer exists.
Evidence to check: the legal basis for collecting and licensing each category, the consent or notice records behind it, enforcement against comparable sellers, and whether a deletion obligation would reach derived products as well as raw records. Deletion that reaches models built from the data is the clearest sign a loss would not be temporary.
A substitute that passes the buyer's own acceptance test
Substitution is permanent when a cheaper source passes the buyer's own acceptance test for the task, and temporary when it passes only a weaker one. Generic text already has a compute-priced substitute: Hugging Face's Cosmopedia repository describes a dataset of synthetic textbooks, blog posts, stories and WikiHow-style articles with over 30 million files and 25 billion tokens, generated in more than 10,000 H100 GPU hours.
Records of real business decisions are harder to generate, because their value lies in realized outcomes, linked events and the constraints that actually bound people at the time. Harder is not a moat an investor can bank on, though. The useful question is narrow: for the tasks a company's buyers pay for, has anyone compared a model trained with the licensed records against one trained on the substitute, using the buyer's test?
Evidence to check: the acceptance tests in buyer contracts, any with-and-without comparisons the company or its buyers have run, which fields a generator cannot reproduce, and whether buyers' requirements are drifting toward fields synthetic data can fill.
Concentration: when one buyer's exit becomes an impairment
Concentration turns a buyer's departure into permanent impairment when the asset was shaped for that buyer: built to its specification, licensed to it exclusively or priced on its use alone. In the illustrative model, a buyer holding half the book and leaving after year two costs 33.8% of value, more than a permanent 30% cut, because the case assumes no replacement buyer appears.
Public companies show the level of disclosure to ask a private one to match. IFRS 8 requires an entity to disclose when revenue from a single external customer reaches 10% or more of its revenues, with the total from each such customer, and it treats a group of entities under common control as one customer. Two buyers under one parent are one exposure.
Evidence to check: revenue and contract value by customer group under that common-control rule, contract end dates and renewal terms, exclusivity grants, and the share of the dataset that would still sell if the largest group left. A concentrated book is fragile, but if the records would sell elsewhere the loss is a delay rather than an impairment.
The permanent impairment register
The register puts the four tests on one page, with a fifth row for the condition that converts temporary losses into permanent ones. Each row names the look-alike that should not be treated as permanent, so a committee spends its questions where they change the answer:
| Event | Why it can be permanent | Temporary look-alike | Evidence an investor can check | Early signal |
|---|---|---|---|---|
| Rights or supply lost | Delivered records, or the only source of new ones, become unusable | One supplier delays renewal while others hold equivalent records | Chain of title, third-party content warranties, renewal and termination clauses, independent supplier count | A supplier seeks to narrow scope, or a claim names upstream content |
| Legal or regulatory bar | The rule targets the category, and deletion can reach derived products | An inquiry into one company's notices that a fix can cure | Legal basis per category, consent records, enforcement against peers, deletion scope | Proposed rules or orders aimed at comparable data sellers |
| Buyer substitute | A cheaper source passes the buyer's own acceptance test | A substitute that passes only a weaker test | Acceptance tests, with-and-without comparisons, fields a generator cannot reproduce | Buyers drop fields or ask for shorter history |
| Concentration | The asset was built, licensed or priced for one buyer | A large buyer pauses while the records still sell elsewhere | Revenue by customer group, end dates, exclusivity, resale share | A renewal passes its date with no new terms |
| Runway (the converter) | A delay outlasts the cash and forces a sale or shutdown | A delay the balance sheet can carry | Months of cash against the longest plausible delay | Fundraising timed to one deal closing |
Could a temporary risk turn permanent, or a permanent one be cured?
Both happen, and the strongest objection to a four-event list is that permanence depends on the response. A delay turns permanent if it outlasts the cash, which is why the register carries runway as a converter. A failure that looks final can also be cured: LAION pulled its LAION-5B image-link dataset after a Stanford Internet Observatory report of 19 December 2023 found links to illegal content, said it was temporarily taking the datasets down to make sure they were safe before republishing, and released a revised Re-LAION-5B on 30 August 2024.
The defect sat in specific links that could be found and removed while the rest of the asset stood. Put the same question to any rights or legal problem: can the affected records be identified and removed, with what remains still selling? If so, it is a costly repair. If the defect runs through the whole asset, it is an impairment.
Correlation is the other caution. One court ruling on upstream content can be a rights failure, a reason for a buyer to try a substitute and a trigger for a concentrated buyer to leave, all in the same quarter, so the rows of the register should be read together rather than added up.
How SourceX's published pages map onto the four events
SourceX's public pages describe process steps that bear on the first two events and publish nothing measurable about any of the four, so this section maps them instead of scoring them. Its published process covers finding and qualifying records, reviewing rights, writing the contract, preparing the data, delivering it and paying the supplying company, and no data moves until the supplying company has accepted the scope and terms and signed.
On supply and concentration, the pages say datasets are sourced from U.S. businesses when a buyer asks rather than kept in stock, and that exclusivity and term are set in each signed agreement, whose terms are not published. An investor would ask how those agreements handle termination, renewal and third-party content, what acceptance tests buyers would apply to the datasets, and how any revenue would group under a common-control customer rule.
SourceX shares company and commercial figures with investors directly rather than in articles, so none appear here and nothing above predicts how any of these events would play out for it. It is raising from accredited investors under Rule 506(c), with materials available through its investors page.
Methods note: the cash model, its sources and its gaps
The four-event test and the register were written for this article on 27 October 2026. The sensitivity model holds a licensing book at 100 index points of cash a year for ten years, discounts at 15% a year with end-of-year timing and reports each case's lost share of present value, computed in exact fractions and rounded once, half up. On a 20-year horizon the year-one loss falls to 13.9%, the proportional cut stays at 30.0%, rights ending after year four cost 54.4% and the departing half-book buyer 37.0%.
Sources: IAS 38 and IFRS 8 on ifrs.org, the Thomson Reuters v. Ross opinion on the Delaware district court's site, the FTC's Avast announcement, the European Commission's AI Act page, LAION's two notices and the Cosmopedia repository, each checked on its own domain on the date above. Not measured: the probability of any event for any company, recovery rates, the correlation between events, and any SourceX figure, none of which is public.
Frequently asked questions
What is the most serious downside risk for an AI data business?
Losing the rights behind its records, because that shortens the asset's life instead of delaying its cash. A legal bar on the whole category comes close. Which one matters most for a given company depends on how many independent suppliers and buyers stand behind its datasets.
Is customer concentration a permanent risk?
Not by itself. It becomes permanent when the dataset was built, licensed or priced for the departing buyer and would not sell elsewhere. Ask what share of the dataset would still sell if the largest customer group left.
Could a synthetic corpus stand in for licensed operating records?
For some generic text it already competes, as large open synthetic corpora show. For tasks that depend on realized outcomes and real constraints, the test is the buyer's own acceptance test run with and without the licensed records. Without that comparison, neither side of the argument is evidence.
How can regulators cause a permanent loss for a data company?
By banning a category of use or ordering deletion. The FTC's Avast order bars selling or licensing browsing data for advertising and reaches products derived from the data, and the EU AI Act prohibits building facial recognition databases through untargeted scraping. Each removes a market rather than delaying it.
How should an investor model a temporary loss?
As a delay to cash with the asset intact, then check the company's runway against the longest plausible delay. A delay the balance sheet cannot carry becomes a permanent loss through a forced sale or shutdown.
Sources
- IAS 38 sets an intangible asset's useful life as the shorter of its economic and legal periods and counts renewals only with evidence they come without significant cost. Source
- IFRS 8 requires disclosure when one external customer provides 10% or more of revenues and treats entities under common control as a single customer. Source
- The February 2025 Thomson Reuters v. Ross opinion records that Ross trained on Bulk Memos bought from LegalEase that were built from Westlaw headnotes, finds copying of 2,243 headnotes and weighs factor one for Thomson Reuters. Source
- The FTC's June 2024 final order bans Avast from selling or licensing web browsing data for advertising and requires deleting the data passed to Jumpshot and products or algorithms derived from it. Source
- The European Commission lists untargeted scraping of facial images to build facial recognition databases among AI Act prohibited practices that apply from 2 February 2025. Source
- LAION's safety notice says it temporarily took its datasets down to make them safe before republishing. Source
- LAION released Re-LAION-5B on 30 August 2024 after the December 2023 Stanford Internet Observatory findings. Source
- The Cosmopedia repository describes an open synthetic dataset of over 30 million files and 25 billion tokens generated in more than 10,000 H100 GPU hours. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.