Skip to content

AI data market

Public filings and AI data spending: a visibility map for 8 deal parts

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Public filings show AI data spending mostly from the seller's side and only in pieces: licensing revenue when a public seller reports it as a line, fixed-price backlog, and deals named without prices. Buyers' data costs sit inside research and development or cost of revenues, usage-priced licenses can be left out of backlog, and private sellers file nothing. Treat a filings total as a floor whose distance from true spend filings cannot show.

Value drivers

Public filings AI data

This page covers public filings and AI data spending: a visibility map for 8 deal parts.Illustrative

Key takeaways

  • Of eight parts of an AI data deal, one is visible in aggregate by rule, four surface only when material or by choice, two are hidden by an exemption or redaction, and one sits outside filings entirely.
  • Backlog disclosures can leave out usage-based royalties on licensed intellectual property and contracts of one year or less, so a backlog decline can reflect pricing, not demand.
  • The SEC eliminated the contractual obligations table in November 2020, so a buyer's data commitments now surface only as narrative and only when material.
  • Alphabet's 2025 10-K explains an $11.2 billion rise in other cost of revenues partly by content acquisition costs, a licensed-content line for distribution, not training data.
  • Under illustrative inputs, true spend could be 2.5 to 16.7 times a filings total, and the coverage share driving most of that range appears in no filing.
  • ASU 2024-03 expense disaggregation starts with annual periods beginning after 15 December 2026 and lists no category for licensed data.

What can public filings reveal about AI data spending?

Only fragments, and nearly all of them from the seller's side. A public company that licenses data may report the revenue, may disclose fixed-price backlog and may name a deal; a public company that buys data reports the cost inside broad expense lines. A seller without registered securities files no periodic reports at all.

Three quantities get confused. Recognized licensing revenue is what a seller booked in the period. Backlog, which U.S. accounting calls remaining performance obligations, is contracted revenue not yet recognized. Buyer spend is what the licensee paid or committed, whoever it bought from. Filings measure the first two for a handful of sellers and almost never the third, which is the quantity a market size needs.

Adding up disclosed seller revenue therefore gives a floor, and it is a useful one, because it is dated and issued under securities law. The question on this page is what such a floor cannot see: which parts of a data deal reach a filing, which rule pulls each one in, which exemption keeps it out, and how far below true spend the floor could sit under stated assumptions.

Why a buyer's filing hides what it spends on data

Because financial statements group costs by function and nature, and licensed data is rarely large enough, or managed separately enough, to earn a line of its own. In a developer's filing it usually sits inside research and development or cost of revenues, next to compute, salaries and depreciation.

Alphabet's Form 10-K for 2025 shows how even a licensed-content cost disappears into a combined line. Its cost of revenues rose $16.2 billion from 2024, which the filing attributes to increases in other cost of revenues and traffic acquisition costs of $11.2 billion and $5.0 billion. In the passage we read, the rise in other cost of revenues is explained by content acquisition costs, depreciation and technical infrastructure, with YouTube named as the main source of the content increase, and no separate figure for content alone. Content acquisition costs pay for content that Alphabet distributes, mainly on YouTube. That is licensing for distribution to viewers, not for training, which makes it the classic lookalike line.

Newer segment rules help only at the margin. FASB's ASU 2023-07, effective for fiscal years beginning after 15 December 2023, requires public companies to disclose significant segment expenses that are regularly provided to the chief operating decision maker, plus an amount for other segment items with a description of what it contains. Licensed data would appear only if that executive reviews it as a significant expense in its own right.

The next change is broader. ASU 2024-03, issued on 4 November 2024, requires public business entities to break relevant expense captions into specified categories, namely purchases of inventory, employee compensation, depreciation and intangible asset amortization, starting with annual periods beginning after 15 December 2026. Licensed data is not one of the categories, so it stays in whatever remains of each caption; in the Board's proposal, that remainder was to be described in words rather than numbers. For a calendar-year company the first covered annual period is 2027, which makes those filings the first real test of whether any developer names data spend at all.

Which seller-side numbers survive the disclosure exemptions?

Two kinds survive: recognized revenue, when a seller reports licensing as its own line, and backlog for fixed-price contracts longer than a year. Everything else depends on how the deal is priced and how long it runs.

The backlog disclosure has practical expedients. FASB's ASU 2016-20 added optional exemptions from the remaining performance obligations disclosure, including for variable consideration that is a sales-based or usage-based royalty promised in exchange for a license of intellectual property, and the standard also lets companies leave out contracts with an original expected duration of one year or less. A company that uses an exemption has to say so and describe what it left out, but it does not have to put a number on it.

For data licensing the consequences are concrete. A license priced per record delivered, per API call or as a share of the buyer's revenue can be absent from backlog while producing revenue every quarter. Pilots and one-year renewals can be absent too. So a fall in reported backlog can mean that deals moved to usage pricing or shorter terms, not that demand fell, and a backlog figure is a lower bound on contracted demand even for the sellers that report it.

Named but not priced: redacted contracts and a retired table

A deal can be visible as a fact while its price stays private. Securities rules ask for material contracts to be filed as exhibits, but they also let companies withhold terms, and one rule change removed a table that used to show future commitments.

Item 601(b)(10)(iv) of Regulation S-K lets a company redact specific provisions or terms of a filed contract when it customarily and actually treats that information as private or confidential and the omitted information is not material, provided the exhibit is marked as redacted.

News Corp shows the named-but-unpriced pattern in its own words. Its 2025 preliminary proxy statement refers to implementing its historic, multi-year content licensing agreement with OpenAI; the passage we read gives no value for it, and a press estimate is not filing evidence.

On the buyer side, the SEC's November 2020 amendments to Regulation S-K eliminated Item 303(a)(5), the tabular disclosure of contractual obligations, and added Item 303(b)(1) on material cash requirements. Commissioners Lee and Crenshaw objected in a joint statement that the change removed a tabular presentation that gave investors insight into supply chains and risk management. A developer's multi-year data commitments, which could once have sat inside a purchase-obligations row, now surface only in narrative and only when material.

The visibility map: eight parts of an AI data deal

Of eight parts of an AI data deal, one is visible in aggregate by rule, four surface only when material or by choice, two are hidden by an exemption or a redaction right, and one sits outside filings entirely. The grades describe what a filing can show, not what any particular filer chose to show.

Use the map before reading a filing: decide which part a figure belongs to, then check whether the exemption in the third column could have removed some of it. The map follows.

The visibility map: eight parts of an AI data deal
Deal partWhere it could surfaceWhat lets it stay outVisibility
That a licensing deal existsA filed material contract, MD&A, a results release or a proxy statementDeals below materiality; release and proxy mentions are voluntaryConditional
Price or total value of one dealThe contract exhibit, if it is filedItem 601(b)(10)(iv) redaction of immaterial terms kept confidentialHidden by redaction
A seller's recognized licensing revenueA revenue line the seller chooses to report separatelyBundling inside broader revenue captions; materialityConditional
Fixed-price commitments longer than a yearRemaining performance obligations in the revenue noteReported in aggregate, not by deal or data typeVisible in aggregate
Usage-priced licenses and contracts of a year or lessNowhere in backlogASC 606 practical expedients, including the ASU 2016-20 royalty exemptionHidden by exemption
A buyer's cost of licensed dataSignificant segment expenses (ASU 2023-07); expense disaggregation (ASU 2024-03)No required category for data; sits in R&D or cost of revenuesConditional
A buyer's future data commitmentsMaterial cash requirements narrative, Item 303(b)(1)Contractual obligations table eliminated in 2020; narrative only when materialConditional
Spend paid to private sellersNo filing by the sellerCompanies without registered securities file no periodic reportsOutside filings

An extraction worksheet for any single filing

Record eight fields for every figure you lift from a filing, so a second analyst can recheck it and nobody adds a seller's revenue to a buyer's cost of the same deal. The worksheet covers what a filing does show; the map covers what no extraction can reach.

Each field below exists because a specific mistake keeps recurring in market models built from filings.

An extraction worksheet for any single filing
FieldWhat to recordMistake it prevents
Document and periodFiler, form, period end and fiscal calendarAdding a fiscal year ending in April to calendar-year figures without saying so
Side of the dealSeller revenue, seller backlog, buyer cost or buyer commitmentCounting one payment once as revenue and again as cost
Caption contentsEverything else the line includes, quotedTreating a bundled line as pure data licensing
Recognized or contractedRevenue in the period versus remaining performance obligationsReading multi-year backlog as one year's spend
Exemptions usedRoyalty and one-year expedients stated in the revenue noteReading a backlog decline as a demand decline
Use of the licensed contentTraining, evaluation, distribution or displayCounting distribution licenses, such as content acquisition costs, as training demand
Counterparty and overlapNamed buyer or seller, and whether the other side also filesDouble counting a deal both parties disclose
Evidence locationPage or note number and a short quoted excerptFigures nobody can find again

How far below true spend can a filings total sit? An illustrative sensitivity

Far, and no filing can say how far. If p is the share of annual licensed-data spend paid to sellers that file and report licensing as its own line, and b is the share of those sellers' licensing revenue they quantify separately rather than inside a broader caption, then a filings total captures p times b of true spend, and true spend equals the filings total divided by p times b.

Both inputs below are illustrative assumptions, not estimates: p at 15, 30 and 50 percent, and b at 40, 60 and 80 percent. In the central case, 30 percent times 60 percent is 18 percent, so true spend would be 1 divided by 0.18, about 5.6 times the filings total. With both inputs low the multiple is 16.7; with both high it is 2.5.

The coverage share dominates. Moving p alone across its range moves the multiple from 11.1 to 3.3; moving b alone moves it from 8.3 to 4.2. Breakout, b, can at least be checked caption by caption with the worksheet. Coverage, p, is exactly what filings cannot report, because the sellers it describes do not file. That is why a filings-based floor should be published as a floor with its distance from true spend left open, never grossed up into a market size.

The table shows the seven scenarios' arithmetic; multiples are computed from exact fractions and rounded half up to one decimal.

How far below true spend can a filings total sit? An illustrative sensitivity
ScenarioCoverage pBreakout bVisible share p x bTrue spend as a multiple of the filings total
Both inputs low15%40%6.0%16.7x
p low, b central15%60%9.0%11.1x
p central, b low30%40%12.0%8.3x
Both inputs central30%60%18.0%5.6x
p central, b high30%80%24.0%4.2x
p high, b central50%60%30.0%3.3x
Both inputs high50%80%40.0%2.5x
Range chart of how many times true spend could exceed a filings total: 2.5x to 16.7x with both illustrative inputs varied, 3.3x to 11.1x with coverage varied alone and 4.2x to 8.3x with breakout varied alone, with the central case marked at 5.6x.
Illustrative inputs; multiple = 1 / (p x b); not a market-size estimate.

What could make this reading of filings wrong

Six developments could make this reading too pessimistic or too generous. Each points in a known direction, which is the reason to track them rather than to stop using filings.

  • Expense disaggregation could expose more. ASU 2024-03 applies to annual periods beginning after 15 December 2026; if developers describe licensed data in what remains of their expense captions, the buyer side becomes partly visible and this map's buyer rows move up a grade.
  • Materiality moves with deal size. As licenses grow, more of them cross into material contracts and significant segment expenses, so visibility rises with spend and an old floor understates a new year.
  • Pricing form can fake a trend. A shift from fixed-price terms to usage pricing lowers reported backlog with no change in demand, because usage-based royalties on licensed intellectual property can be left out of the backlog disclosure under ASU 2016-20.
  • Lookalike lines mislead. Content licensed for distribution, such as the content acquisition costs Alphabet describes for YouTube, is not training-data demand and should not enter a training-data model.
  • Rules change in both directions. The SEC removed the contractual obligations table in 2020; a later amendment could remove another data point or add one.
  • Coverage is unobservable. The share of spend reaching sellers that file and break out the line drives the illustrative multiple, and no filing reports it, so any single grossed-up total rests on an assumption.

Questions for any company that cites filings as demand, SourceX included

Ask for the side, the caption and the exemptions before you accept any figure presented as demand. A company that cannot say whether a number is seller revenue, seller backlog or buyer cost has not yet separated evidence from assumption.

Direct licensing from private operating companies falls in the part of the market filings do not cover, because a company without registered securities reports nothing about the deals it signs. SourceX acquires licenses to business data straight from the companies that hold it and acts as the licensor to AI developers itself, and it publishes no demand, pipeline, supplier or buyer figures. Its investor page says company and commercial metrics are provided directly to investors.

SourceX is raising under Rule 506(c), so only verified accredited investors can participate; materials are requested through its investor page, and the questions below are fair to put to it as to anyone.

  • Which side of the deal does each figure come from: seller revenue, seller backlog, buyer cost or buyer commitment?
  • Which filings, captions and periods, and what else sits inside each caption?
  • Did any figure come from backlog, and did the filer use the royalty or one-year exemptions?
  • How do you treat deals a filing names without a price?
  • What coverage share do you assume for sellers that do not file, and what evidence sets it?
  • Where do you remove content licensed for distribution or display rather than for training?

Methods note: rule documents, company examples and gaps

The visibility map, the extraction worksheet and the coverage sensitivity were built for this article on 2026-10-28. Rule descriptions come from FASB's ASU 2016-20 and ASU 2023-07, FASB's In Focus summary of ASU 2024-03, the SEC's page for its November 2020 amendments to Regulation S-K, the joint statement of Commissioners Lee and Crenshaw, and the eCFR text of Item 601. Company examples come from Alphabet's Form 10-K for 2025 and from News Corp's 2025 preliminary proxy statement. All were verified through search excerpts on the publishers' own domains on 2026-10-28; the SourceX investor page was read directly. A grade in the map states what a rule allows, read from those documents; it is not a count of how filers behave.

The sensitivity reads no data. Coverage p and breakout b are illustrative; visible share is p times b; the multiple is 1 divided by p times b, computed from exact fractions and rounded half up to one decimal. Not measured: any company's actual licensed-data spend or revenue, how often filers use the backlog exemptions, how many deals filings name without a price, the treatment under IFRS for non-U.S. filers, and press-reported deal values, which are excluded throughout. No SourceX data is used, and nothing here is a SourceX measurement, estimate or projection.

Frequently asked questions

Can you size the AI data market from 10-K filings?

Not on their own. Filings show recognized licensing revenue and backlog for the few public sellers that break them out, which sets a floor. Buyer spend, usage-priced deals and private sellers mostly stay out, so the distance between that floor and true spend has to be stated as an assumption.

Why do remaining performance obligations miss some data deals?

Accounting standards let companies leave out of the backlog disclosure contracts of one year or less and, since ASU 2016-20, usage-based royalties on licensed intellectual property. A data license priced per use can therefore produce revenue every quarter without ever appearing in backlog.

Will ASU 2024-03 make AI data costs visible in 2027 annual reports?

Partly at best. It requires public companies to break expense captions into inventory purchases, employee compensation, depreciation and intangible amortization from annual periods beginning after 15 December 2026. Licensed data is not a listed category, so it stays in the remainder unless a company chooses to describe it.

Why don't large AI developers report what they pay for training data?

Accounting groups costs inside research and development or cost of revenues, and segment rules require separate lines only for expenses the chief operating decision maker reviews as significant. Unless data spend is managed and reported that way, no rule forces a separate number.

Does a press report of a deal's value count as filing evidence?

No. Only figures in the filing, the company's own release or a regulator's document count. A press estimate can prompt a question, but it should not enter a model as if a company had disclosed it.

Sources

  • Alphabet's 2025 Form 10-K reports a $16.2 billion rise in cost of revenues, $11.2 billion of it in other cost of revenues, attributed partly to content acquisition costs, primarily for YouTube. Source
  • FASB's ASU 2023-07 requires disclosure of significant segment expenses regularly provided to the chief operating decision maker and of other segment items with a description. Source
  • FASB's summary of ASU 2024-03 lists the required expense categories and an effective date of annual periods beginning after 15 December 2026 for public business entities. Source
  • FASB's ASU 2016-20 adds optional exemptions from the remaining performance obligations disclosure, including usage-based royalties on licenses of intellectual property, and requires more information when an exemption is used. Source
  • Item 601(b)(10)(iv) of Regulation S-K allows redaction of contract terms that are not material and that the registrant customarily and actually treats as private or confidential. Source
  • News Corp's 2025 preliminary proxy statement refers to implementing its historic, multi-year agreement with OpenAI. Source
  • The SEC's November 2020 amendments eliminated Item 303(a)(5), the tabular disclosure of contractual obligations, and added a material cash requirements item. Source
  • Commissioners Lee and Crenshaw said the amendments removed a tabular presentation of contractual information that gave investors insight into supply chain and risk management. Source
  • SourceX's investor page says it licenses business data directly from companies and onward to AI developers in its own name, and 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