Skip to content

AI data market

Fresh vs historical dataset market sizing: year one is not a run rate

By SourceX Editorial · Updated

10 sections

Short answer

Size fresh and historical data as two revenue lines with different drivers: archive licensing draws down a finite pool as archives sell, while refresh purchases accrue with new records and renewals. In this article's illustrative model, refresh reaches 29.5 percent of a market that shrinks to 46.5 percent of its first-year size by year six, and a year-one run rate overstates six-year spend by 40.4 percent.

System export

Fresh

This page covers fresh vs historical dataset market sizing: year one is not a run rate.Illustrative

Key takeaways

  • Archive licensing and refresh purchases are a stock and a flow; adding them as one annual market inflates every year after the first.
  • Informa reported Taylor & Francis underlying growth of 14.5 percent in 2024 including non-recurring data contracts and about 3.5 percent without them, then -2.1 and 3.6 percent in 2025.
  • A supplier with seven years of history growing 5 percent a year creates new records equal to 17.3 percent of its archive each year.
  • In the illustrative model, refresh is 29.5 percent of the year-six market and the year-one run rate overstates six-year spend by 40.4 percent.
  • The archive draw-down rate moves the result most: from 10 to 35 percent a year, the six-year total runs from 53.6 to 109.0 units.
  • Ask any data company to split licensing revenue into archive sales and later records, with the buyers who came back.

What does fresh historical dataset market sizing have to separate?

It has to separate two products that are counted in different units: a one-time license to an archive of records that already exist, and a recurring purchase of records created after that archive's cutoff. The first is a stock. The second is a flow. Adding a year of archive sales to a year of refresh purchases and calling the sum an annual market mixes the two and inflates every year after the first.

The investor's decision is how much of any data licensing estimate, or any company's licensing revenue, will recur without new archives being found. Three kinds of evidence answer it: contract terms that say whether later records are included, filings that split recurring from non-recurring licensing revenue, and the rate at which suppliers actually create new records relative to what they already hold.

None of these is a market size, and this article states none. It builds a model in index units, where 100 is what buyers would pay to license every eligible archive once, so a reader can see the shape of the two markets and plug in their own scale.

What one publisher's results show after a non-recurring data year

They show the spike-and-fade pattern a one-time data deal creates: growth that looks exceptional in the year of the deal and weak in the year after. Informa, the parent of the academic publisher Taylor & Francis, reported that Taylor & Francis revenue rose 12.8 percent reported and 14.5 percent underlying in 2024, to £698.2 million, and that underlying growth was about 3.5 percent excluding non-recurring data access contracts; it put those AI partnership contracts at more than $75 million.

The following year reversed the sign. In its 2025 results, Taylor & Francis revenue was £670.8 million, down 3.9 percent reported and 2.1 percent underlying, while underlying growth excluding non-recurring data access contracts was 3.6 percent. Informa's 2025 annual report says the 2024 AI revenue was something "which we expected would not repeat to the same degree in 2025".

These figures measure one publisher's licensing of academic content, in two currencies, as the company classifies it; they do not measure business records, other sellers or the market as a whole, and they do not say whether the 2024 contracts were archive licenses in this article's sense. What they do show is the arithmetic a market model must reproduce: a year of one-time sales lifts the total, and the next year's comparison falls even when the recurring business keeps growing at the same pace.

What one publisher's results show after a non-recurring data year
Taylor & Francis, as reported by InformaIncluding non-recurring data contractsExcluding themWhat the gap shows
2024, reported revenue growth12.8% (to £698.2m)Not givenThe headline includes the one-time contracts
2024, underlying growth14.5%c.3.5%About 11 points came from the non-recurring year
2025, reported revenue growth-3.9% (to £670.8m)Not givenThe prior-year base still holds the one-time contracts
2025, underlying growth-2.1%3.6%The recurring business grew while the total fell

Two drivers: how fast archives sell and how fast new records accrue

Archive revenue depends on how fast a finite pool of archives gets licensed, and refresh revenue on how many new records suppliers create each year relative to what they hold. The two need different drivers in a model, and the ratio between them is the number most market estimates never state.

Exclusivity makes the stock smaller still. On SourceX's published terms, licenses are exclusive for AI training, the term runs from Closing and covers only the selected data, and future data and unselected records are not included. Under terms like these an archive leaves the stock when it is licensed, and next year's records are a new sale rather than something the first license already bought.

The flow ratio follows from two facts about a supplier: how many years of history it holds and how fast its record volume grows. If volume grows at rate g, next year's new records divided by the whole archive is g times (1 + g) to the power of the history length, divided by that power minus one; with no growth it is one over the history length. A supplier with seven years of history growing 5 percent a year creates new records equal to 17.3 percent of its archive each year. SourceX's general guide is that buyers look for two or more years of consistent history, which is where the ratio is highest. The ratio for other cases, as a share of the archive per year:

Two drivers: how fast archives sell and how fast new records accrue
Years of historyNo volume growth5% a year15% a year
250.0%53.8%61.5%
425.0%28.2%35.0%
714.3%17.3%24.0%
128.3%11.3%18.4%

The six-year model: inputs and their labels

The model runs six years with six inputs, each given a low, central and high value and labeled Illustrative, because no public source measures any of them for business records. The archive stock starts at 100 index units. Each year a fixed share of the archive stock that is still unlicensed gets licensed. A share of those buyers then buys each later year's new records; the first refresh is worth the archive's value times the flow ratio times the relative price of fresh records, and each later refresh grows with record volume and survives at the renewal rate.

The relative price input matters for the size of the refresh market but not for whether it exists. A recency premium has to be proven case by case, so the central case prices a fresh record the same as an archived one, with a discount and a premium as the low and high values. The inputs:

The six-year model: inputs and their labels
InputLowCentralHighLabel
Years of history in a typical archive4712Illustrative; buyers typically look for 2+ years
Annual growth in record creation0%5%15%Illustrative
Share of unlicensed archive stock licensed each year10%20%35%Illustrative
Share of archive buyers who also buy refreshes20%40%60%Illustrative
Annual renewal of refresh purchases60%75%90%Illustrative
Price of a fresh record relative to an archived one0.81.01.5Illustrative

Results: an archive spike, then a smaller recurring base

At central values the market is largest in year one, when it is all archive, and by year six it has shrunk to 46.5 percent of that size with refresh at 29.5 percent of it. Archive licensing falls from 20.0 index units in year one to 6.6 in year six as the stock runs down; refresh starts at zero, rises to 2.8 in year five and eases to 2.7 as non-renewals offset growth. Over six years, buyers spend 73.8 units on archives and 11.7 on refreshes, 85.5 in all.

An analyst who annualized the first year would carry 20 units a year forward and report 120 over six years, overstating the model's own total by 40.4 percent. That is the same error the Informa figures expose in miniature: a year of one-time sales read as a run rate.

The recurring base is real but smaller than the spike, and its size depends on refresh take-up and renewal, which no public source measures for business records. The model also leaves out companies that become eligible suppliers later and add new archives to the stock, which would slow the decline. Year by year, with each figure rounded separately:

Results: an archive spike, then a smaller recurring base
YearArchive licensingRefresh purchasesTotalRefresh share of total
120.00.020.00.0%
216.01.417.48.0%
312.82.215.014.6%
410.22.612.920.3%
58.22.811.025.2%
66.62.79.329.5%
Stacked columns for years one to six in index units, with a shrinking archive segment and a growing refresh segment, labeled with totals for years two to six of 17.4, 15.0, 12.9, 11.0 and 9.3; a dashed line marks the year-one run rate of 20.0, and side labels show refresh at 29.5 percent of year six and the run rate times six overstated by 40.4 percent.
Illustrative model in index units (100 = every eligible archive licensed once); central inputs, figures rounded separately.

Sensitivity: how fast archives sell decides how fast the spike fades

The draw-down rate moves the result more than any other input: selling archives faster makes a bigger early spike, a steeper fade and a larger recurring share later. Moved alone from 10 to 35 percent a year, it takes the six-year total from 53.6 to 109.0 units and the year-six refresh share from 23.0 to 44.7 percent, while year six shrinks from 76.7 to 21.0 percent of year one.

History length works the other way. A deeper archive means each year's new records are a smaller share of it, so with 12 years of history the year-six refresh share falls to 21.5 percent, against 40.6 percent with 4 years. Refresh take-up and relative price move the refresh line in proportion: tripling take-up from 20 to 60 percent raises the year-six share from 17.3 to 38.6 percent.

For diligence, the ranking says which question to ask first. How fast a company is selling through the archives it can reach decides how long its early revenue lasts; the share of buyers who come back for later records decides what is left. Each input moved alone, with the others at central values:

Sensitivity: how fast archives sell decides how fast the spike fades
Input moved aloneLow value: year-6 refresh share; 6-year totalHigh value: year-6 refresh share; 6-year totalYear 6 as share of year 1, low; high
Archive draw-down (10% to 35%)23.0%; 53.644.7%; 109.076.7%; 21.0%
Years of history (4 to 12)40.6%; 92.921.5%; 81.455.2%; 41.7%
Refresh take-up (20% to 60%)17.3%; 79.638.6%; 91.339.6%; 53.4%
Record growth (0% to 15%)24.0%; 83.041.3%; 91.943.1%; 55.8%
Renewal (60% to 90%)22.1%; 83.238.3%; 88.542.1%; 53.1%
Relative price (0.8 to 1.5)25.1%; 83.138.6%; 91.343.7%; 53.4%

Could AI spending growth stand in for refresh demand?

No. Spending on AI measures something else, and growth in it cannot tell an archive sale from a refresh. Sequoia's 2023 essay AI's $200B Question, by David Cahn, argued that each year of GPU capital spending at the time implied roughly $200 billion of lifetime revenue needed to pay it back, and estimated a gap of $125 billion or more to fill. Those figures describe what data-center investment must earn, not what anyone spends on licensed records, and nothing in them recurs on a data refresh cycle.

Accounting can blur the line further. IFRS 15 distinguishes a license that gives a right to access intellectual property as it exists throughout the license period, recognized over time, from one that gives a right to use it as it exists when granted, recognized at a point in time (paraphrase of paragraphs B56 to B61). A one-time archive license can therefore land in a single year's revenue, and unless a filing splits the line, as Informa's results do, a reader cannot tell a spike from a base.

The test of the alternative is whether licensing revenue moves with AI spending. In the one public split read for this article, it did not behave like a stream that grows with spending: Informa classed its AI data contracts as non-recurring and said it expected them not to repeat to the same degree in 2025.

What could make the stock-and-flow model wrong

Six developments would change the shape of the two markets, and each pushes in a known direction:

At a glance

What could make the stock-and-flow model wrong

  • New suppliers: companies that become eligible later add archives to the stock, slowing the archive decline; the model holds the stock fixed.
  • Non-exclusive licensing: where the same archive can be licensed to several buyers, the effective stock is larger and runs down more slowly than under exclusive terms.
  • Model release cadence: if buyers retrain on newer cutoffs more often, refresh take-up and renewal rise; if they rely on retrieval at query time instead, refresh purchases of training data could shrink.
  • Synthetic substitution: generated records could replace some refresh volume for tasks whose acceptance tests do not require observed recent outcomes.
  • Buyer concentration: with few buyers, refresh take-up moves in large steps, and one lapse can remove much of the recurring base.
  • Retention rules and deletion: shorter retention shrinks history length, which raises the flow ratio but lowers the archive stock.

Questions for any data company that calls its revenue recurring, SourceX included

Ask for revenue split the way the model splits it. A company that cannot say how much of its licensing came from archives, how much from later records, and how many archive buyers came back has not shown recurring revenue, whatever its contracts are called.

SourceX's published terms sit mostly on the archive side. Its how-it-works page says it usually pays a one-time fee for a dataset, that some agreements set recurring deliveries, each paid as the agreement states, and that a license covers only the selected data, with future data not included. Its questions page says historical depth is usually the starting point and refreshes follow if a first deal works. SourceX does not publish deal counts, refresh rates or revenue; the investor page keeps those numbers for conversations with investors. Its offering is made under Rule 506(c) to verified accredited investors only, and a reader evaluating it can send the questions below with a materials request from that page.

  • What share of licensing revenue in each period came from archives, and what share from records created after the first license?
  • How many archive buyers bought later records within 12 months, and how many renewed after that?
  • How much eligible archive stock remains unlicensed, and how fast has it been drawn down?
  • Are licenses exclusive, and do any include future records automatically?
  • Which revenue is recognized at a point in time, and which over time?

Methods note: sources, model rules and what was left out

The stock-and-flow model, the flow-ratio table and the sensitivity table were built for this article on 2026-10-28. Public evidence: Informa's 2024 and 2025 full-year results and its 2025 annual report, IFRS 15 as published by the IFRS Foundation, Sequoia's 2023 essay, and SourceX's how-it-works, questions and investor pages. Every model input is Illustrative; no figure from the Informa results enters the model.

Rules, as the script applies them: archive licensing in year t is the draw-down rate times the stock still unlicensed at the start of year t, from a starting stock of 100. An archive licensed in year t produces refresh value from year t + 1, equal to its archive value times the take-up share, the relative price and the flow ratio, multiplied by (1 + growth) and by the renewal rate once for each year after the first refresh. The flow ratio is g(1 + g)^H / ((1 + g)^H - 1), or 1/H with no growth. Figures are computed in floating point and rounded half up to one decimal at the end; the naive run rate is year-one total times six.

Not measured: any real draw-down, take-up or renewal rate; prices of archive or fresh records; new supplier entry; and the share of real licensing revenue that is recurring at any company other than the one public split cited. The script and its output are kept as bin/xs0368_stockflow.py and bin/xs0368_output.txt in the writer's scratch folder.

Frequently asked questions

Is fresh data licensing a bigger market than historical archives?

Not at first, in this article's illustrative model: year one is all archive sales, and refresh reaches 29.5 percent of the market by year six. The recurring flow becomes a larger share as archives run down, but its size depends on refresh take-up and renewal rates that no public source measures for business records.

Why does a one-time data licensing deal distort growth figures?

Because it lifts revenue in one year and then sits in the comparison base for the next. Informa reported Taylor & Francis underlying growth of 14.5 percent in 2024 including non-recurring data contracts and about 3.5 percent without them, then -2.1 percent in 2025 including them and 3.6 percent without.

How do you estimate how much new data a supplier creates each year?

Divide next year's expected records by the archive it already holds. With steady volume that is one over the years of history; with growth it is higher. Seven years of history growing 5 percent a year gives about 17.3 percent of the archive each year.

Does a refresh contract make data licensing revenue recurring?

Only if buyers actually take later deliveries and renew. A contract that allows recurring deliveries is an option, not revenue. Ask how many archive buyers bought later records within a year and how many renewed, and check whether revenue is recognized at a point in time or over time.

Can AI spending growth be used to size refresh demand?

No. Figures such as Sequoia's 2023 estimate of the revenue needed to pay back a year of GPU spending describe data-center economics, not purchases of licensed records, and they cannot tell an archive sale from a refresh.

Sources

  • Informa 2024 full-year results: Taylor & Francis revenue £698.2m, up 12.8% reported and 14.5% underlying; underlying growth c.3.5% excluding non-recurring data access contracts; $75m+ of non-recurring data access revenue. Source
  • Informa 2025 full-year results: Taylor & Francis revenue £670.8m against £698.2m, down 3.9% reported and 2.1% underlying; 3.6% underlying growth excluding non-recurring data access contracts. Source
  • Informa 2025 annual report: the 2024 AI partnership revenue was expected not to repeat to the same degree in 2025. Source
  • IFRS 15 (B56-B61): a licence giving a right to access IP as it exists throughout the licence period is recognized over time; a right to use IP as it exists when granted is recognized at a point in time (paraphrase). Source
  • Sequoia, AI's $200B Question (September 2023): each year of GPU capex implied roughly $200B of lifetime revenue needed to pay it back, with a $125B+ gap. Source
  • SourceX how-it-works: usually a one-time fee with some recurring deliveries; licenses exclusive for AI training; the term covers only the selected data and future data are not included. Source
  • SourceX questions page: historical depth is usually the starting point; refreshes follow if a first deal works. Source
  • SourceX investor page: company and commercial metrics are shared with investors directly; investor materials are requested on the page. 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