AI data market
Historical business archive data TAM: five filters, a stock not a flow
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Historical business archive data TAM counts rights holders with surviving, usable, licensable archives, not bytes. Start from an official count, such as Eurostat's 306,000 EU enterprises with 50 or more people employed (2024, preliminary), then apply five shares: group structure, preserved depth, context, rights and buyer fit. Illustrative inputs leave 661 to 37,014 archive owners, about 7,746 at central values, and exclusive licenses take licensed history off the market.
System export
Historical business archive
Key takeaways
- Size historical archive supply in rights holders and archive-years; a dollar figure needs license prices nobody publishes for private business records.
- The only anchored input is the frame: 55,000 large plus 251,000 medium-sized EU enterprises in 2024, 306,000 with 50 or more people employed (Eurostat, preliminary).
- The law requires little depth: U.S. wage-hour rules require 3 years for payroll, sales and purchase records and 2 for order, shipping and billing records, so older history survives by choice.
- Illustrative shares shrink the frame to 661 to 37,014 archive owners, about 7,746 at central values; buyer fit and context completeness move the count most and have no public measure.
- Exclusive AI training licenses make archives a stock: at central inputs, new history accrues at about one-eighth of the stock each year.
What does a historical business archive data TAM count?
It counts rights holders that own surviving, usable and licensable archives, and the years of records they hold, not bytes stored or dollars spent. For historical business records, the total addressable market is a supply ceiling: how many companies could sign a license for years of operating history a buyer would accept.
At a glance
What does a historical business archive data TAM count
The unit matters because the dollar version cannot be built honestly. A dollar figure needs a price per license and a purchase frequency, and neither is published for private business records; SourceX publishes neither. Buyer budgets are the other side of the same market and need their own evidence. This page sizes the supply side, in archive owners and archive-years, and shows every multiplication.
Top-down numbers are the strongest competing explanation, and they overstate addressable supply. AI spending measures compute and services, global data volume measures storage, and venture funding measures investor appetite; none says how many companies kept usable records, may license them and hold something a buyer has asked for. An archive that exists but cannot be licensed, or that no request matches, adds nothing. Four definitions keep the count honest:
- Rights holder: the legal entity that owns the records and can sign a license through an authorized officer.
- Archive-year: one year of operating records from one primary system, preserved with enough context to be used.
- Eligible archive: one that clears the depth, context and rights filters below.
- Selected archive: an eligible archive whose record type matches a buyer request in the period.
Which official counts can anchor the frame?
Only the starting frame can be anchored to a statistical table today. Eurostat's release of 9 December 2025, on preliminary 2024 data, counts 55,000 large enterprises (more than 249 people employed) and 251,000 medium-sized enterprises (50 to 249 people employed) in the EU business economy: 306,000 enterprises with 50 or more people employed.
The 50-person cut is chosen to sit close to SourceX's own screen, which focuses on U.S., Canadian and EU companies with 50 or more full-time employees at peak. The fit is imperfect in three ways. Eurostat counts people employed, not full-time equivalents; it counts size today, not at peak; and it counts enterprises, which, as Eurostat notes, can belong to larger enterprise groups, so one rights holder may sit behind several entries.
The U.S. and Canadian frames stay as reader inputs. The U.S. Census Bureau publishes firm counts by enterprise size in its Statistics of U.S. Businesses tables, but this article did not read those tables, so it uses no U.S. count. Each anchor below carries its year, its definition and what it leaves out:
| Figure | Source and year | Definition | What it does not measure |
|---|---|---|---|
| 55,000 large enterprises | Eurostat, 2024 preliminary data, released 9 December 2025 | More than 249 people employed, EU business economy | Full-time equivalents, enterprise groups, size at peak, closed firms |
| 251,000 medium-sized enterprises | Eurostat, same release | 50 to 249 people employed, EU business economy | Same gaps as the large-enterprise count |
| 46% five-year survival | Eurostat, Key figures on European business, 2023 edition | EU enterprises born in 2015 still active in 2020 | Survival of firms that had already reached 50 people; whether their records survived |
| 34.7% ten-year survival | U.S. Bureau of Labor Statistics, 2024 | Private-sector establishments born in March 2013 still operating in March 2023 | Firms rather than establishments, large firms, records |
| 3 and 2 years of retention | 29 CFR 516.5 and 516.6, U.S. wage-hour rules | Minimum periods for payroll, sales and purchase records (3) and for order, shipping and billing records (2) | Longer periods in other laws; what companies keep by choice |
How many archives survive long enough to matter?
Fewer than the frame suggests, because the law requires only a few years of history to be kept and companies do not last. Under U.S. federal wage-hour rules, employers must keep payroll records, certain agreements and certificates, and sales and purchase records for at least 3 years (29 CFR 516.5), and supplementary records such as time cards, wage rate tables and order, shipping and billing records for 2 (29 CFR 516.6). Anything older survives because someone chose to keep it through migrations and staff changes.
European privacy law pushes in the same direction for records that name people. The GDPR's storage limitation principle requires personal data to be "kept in a form which permits identification of data subjects for no longer than is necessary for the purposes for which the personal data are processed" (Article 5(1)(e)), and Recital 39 says time limits should be set for erasure or periodic review. Operating archives full of customer and employee details may therefore run shallower in the EU than purely transactional ledgers. This is general information, not legal advice.
Firm survival trims the frame further back in time. The Bureau of Labor Statistics reported that in March 2023, 34.7 percent of U.S. private-sector establishments born in March 2013 were still operating. Eurostat put five-year survival at 46 percent for EU enterprises born in 2015. Both count new businesses of every size, so they overstate attrition among companies that had already reached 50 people; what they show is that a ten-year archive needs a rights holder that existed and kept records for ten years, and that many old archives belong to companies that no longer exist, whose records sit wherever an acquirer or liquidator left them.
The supply workbook: one anchored frame, five filters
The workbook multiplies one anchored count by five shares, each with a low, central and high value. Every share is an illustrative assumption written for this article, not a measurement, and each is there because it removes a distinct reason an archive cannot be licensed.
Group structure (g) removes enterprises that do not decide on their own records. Preserved depth (d) keeps those holding at least five years in one primary system, beyond the two- and three-year federal wage-hour minimums. Context completeness (c) keeps archives whose records link work to its outcome, such as a ticket to its resolution or an order to its delivery. Rights (r) keeps those that pass a rights review for at least one system, and buyer fit (b) keeps those whose record type matches a buyer request in the period. Usable archive-years per owner (Y) converts owners into a stock of history.
Replace any value with your own evidence; the next section shows how the cells combine. The inputs:
| Input | Low | Central | High | Basis |
|---|---|---|---|---|
| F. EU enterprises with 50 or more people employed | 306,000 | 306,000 | 306,000 | Anchored: Eurostat, 2024 preliminary |
| g. Decide on their own records, not through a group | 60% | 75% | 90% | Illustrative; Eurostat notes enterprises can belong to groups |
| d. Keep 5 or more years in one primary system | 30% | 45% | 60% | Illustrative; U.S. legal minimums are 2 to 3 years for many records |
| c. Records link work to its outcome | 30% | 50% | 70% | Illustrative; no public measure found |
| r. Pass rights review for at least one system | 40% | 60% | 80% | Illustrative; no public measure found |
| b. Record type matches a buyer request in the period | 10% | 25% | 40% | Illustrative; buyer requests name scope, volume, language and format |
| Y. Usable archive-years per owner | 5 | 8 | 12 | Illustrative |
Arithmetic: from 306,000 enterprises to an illustrative range
At central values the workbook leaves about 7,746 archive owners in the EU frame, inside an illustrative range of 661 to 37,014. The ends of that range are corners where every share sits at its low or its high value at once, so the true figure is unlikely to sit at either end.
The arithmetic, with nothing hidden. Central: 306,000 x 75% x 45% x 50% x 60% x 25% = 306,000 x 0.0253125 = 7,745.6, rounded to 7,746. Low: 306,000 x 60% x 30% x 30% x 40% x 10% = 306,000 x 0.00216 = 661.0, rounded to 661. High: 306,000 x 90% x 60% x 70% x 80% x 40% = 306,000 x 0.12096 = 37,013.8, rounded to 37,014.
Multiplying the unrounded owner counts by usable archive-years gives the stock of history: 661.0 x 5 = 3,305 archive-years at the low corner, 7,745.6 x 8 = 61,965 at the center and 37,013.8 x 12 = 444,165 at the high corner, each rounded once at the end. In owners the high corner is about 56 times the low one; in archive-years, about 134 times. That width is the honest size of what nobody has measured, and none of these figures is a SourceX estimate, a dollar market size or a measurement.
The range widens at every step, which the chart traces on a log scale. After group structure the running count spans 183,600 to 275,400; after preserved depth, 55,080 to 165,240; after context, 16,524 to 115,668; after rights, 6,610 to 92,534; and after buyer fit, 661 to 37,014, each rounded half up from the exact product.
Which input moves the count most?
Buyer fit. Moved alone from its low to its high value, with every other input central, it swings the count from 0.4 to 1.6 times the central case. Context completeness swings it from 0.6 to 1.4 times, rights and preserved depth from 0.67 to 1.33 times each, and group structure from 0.8 to 1.2 times.
In a product, the input with the widest range relative to its central value dominates, and here the two widest, buyer fit and context completeness, are also the two with no public evidence at all. That tells an investor where diligence time pays: ask a company that claims archive supply for measured selection and completeness rates from a disclosed sample, not for a bigger frame count.
Each row varies one input and holds the rest at central values:
| Input varied alone | Low to high value | Archive owners | Multiple of central count |
|---|---|---|---|
| b. Buyer fit | 10% to 40% | 3,098 to 12,393 | 0.4x to 1.6x |
| c. Context completeness | 30% to 70% | 4,647 to 10,844 | 0.6x to 1.4x |
| d. Preserved depth | 30% to 60% | 5,164 to 10,328 | 0.67x to 1.33x |
| r. Rights clear | 40% to 80% | 5,164 to 10,328 | 0.67x to 1.33x |
| g. Group structure | 60% to 90% | 6,197 to 9,295 | 0.8x to 1.2x |
Why archive history is a stock, not a flow
Because an exclusive AI training license takes the licensed history off the market while the exclusivity runs, and new history accrues only a year at a time. A TAM that treats the historical stock as annual supply counts the same archive in every year.
SourceX's terms show the mechanism. Its how-it-works page says "Licenses are exclusive for AI training: you agree not to license the selected data to anyone else for that purpose," with the term and any alternatives set in each Dataset Deal Confirmation. The term covers only the selected data, so future data and unselected records stay outside it, and the supplier confirms the data has not been licensed for AI training before, so an archive another buyer licensed first drops out of the pool.
Illustrative arithmetic makes the point. At central inputs, 7,746 owners hold about 61,965 archive-years; if each adds one year of records a year, new supply is 7,746 archive-years a year, about one-eighth of the stock. Once the first licenses are signed, repeatable supply is that accrual, plus unselected systems and newly eligible owners, not the stock. Non-exclusive licensing models behave differently, which is one more reason to ask which kind a company's supply figure assumes.
Where the archive supply count could break
Six weaknesses could push the true count well away from any total these inputs produce, in either direction. Each names the direction it pushes and the evidence that would settle it, so a reader can see which new data would move the estimate.
- Frame mismatch: Eurostat counts people employed today, not full-time staff at peak, and enterprises rather than groups; firms that were once larger fall out, and groups count more than once.
- Preservation is unmeasured in both directions: more companies keeping longer histories than assumed would raise d and Y, while retention schedules, system migrations and the GDPR's storage limitation principle push them down; a measured depth distribution would settle it.
- Rights outcomes are unmeasured: customer contracts, vendor terms and earlier licenses can each remove a system, and every exclusive license signed by any buyer shrinks the pool for the next.
- Buyer demand can move: if developers shift requests toward other record types or toward synthetic data, buyer fit falls without any change in supply.
- Selection bias in any measurement: companies that apply to a licensing program are not a random draw from the frame, so their pass rates overstate the frame's.
- Closed and acquired companies: the frame counts active enterprises only; archives of firms that closed may add supply where a successor holds them, or may be gone.
Measuring the five shares, and what to ask any company
Measure the five shares on a disclosed random sample drawn from an official business register, not on applicants. Stratify by size class and sector; for each sampled company record years preserved per primary system, the share of records that link work to an outcome, the rights-review result by system, whether the record type matches a dated set of buyer requests, and preparation hours. Report the sample size, response rate and nonresponse by stratum, keep companies with unfinished reviews in the denominator as open cases, and publish intervals for each share. SourceX has not published such a study; this is the design, and no results are implied.
SourceX is one application of the question rather than evidence for an answer. Its screen and guidance line up with three of the inputs: companies with 50 or more full-time employees at peak in the U.S., Canada and the EU (the frame), and, as a general guide, years of consistent history, typically two or more, with tens of thousands of records or more in one system (depth and context), assessed deal by deal. It publishes no supply counts, pipeline or selection rates, and shares company and commercial metrics with investors directly. For readers weighing SourceX itself: it is raising now under Rule 506(c), from accredited investors only, and its investor page takes requests for materials along with diligence questions like these.
The same questions apply to any company that cites a supply figure for business archives, SourceX included:
Checklist · 6 items
- Which official frame, year and size definition does your supply count start from?
- Which of the five shares are measured, on what sample, with what response rate?
- How do you count archives of acquired or closed companies?
- What share of your archive-years is already under an exclusive license?
- How many archive-years does a typical supplier add each year?
- What share of companies passed rights review, with the denominator stated?
0 of 6 ticked. Tick each item as you go.
Methods note: inputs, arithmetic and what was not measured
The workbook, its arithmetic and the sensitivity table were built for this article on 2026-10-28. The frame comes from Eurostat's 9 December 2025 news release on preliminary 2024 structural business statistics; survival figures from Eurostat's Key figures on European business (2023 edition) and the Bureau of Labor Statistics (2024); retention minimums from 29 CFR Part 516 on eCFR; the storage limitation text from the GDPR on EUR-Lex; and SourceX's terms from its live how-it-works page. Each was checked on its publisher's domain on that date. Products are computed with exact fractions and rounded half up to whole numbers once; sensitivity multiples vary one input at a time from the central case and are rounded half up to two decimals.
Not measured: every share in the workbook and the archive-years per owner; U.S. and Canadian frames; the age distribution of firms with 50 or more people; archives of closed firms; correlations between shares, such as companies with deep archives also keeping better context, which would make the corners more likely than independent inputs suggest; and any price or dollar value. No SourceX data is used, and nothing here is a SourceX measurement, estimate or projection.
Frequently asked questions
How many companies could license historical business archives?
No measured count exists. Starting from Eurostat's 306,000 EU enterprises with 50 or more people employed (2024, preliminary) and applying five illustrative shares, this article's workbook leaves 661 to 37,014 archive owners, about 7,746 at central values. Every share is an assumption until measured on a disclosed sample.
Why not size the historical archive market in dollars?
Because price per license and purchase frequency for private business records are not published, so a dollar total would rest on invented numbers. A supply count in owners and archive-years can be anchored and checked; the buyer-budget side needs its own evidence before the two are combined.
Do acquired or closed companies count in the supply frame?
Acquired companies usually sit inside the frame through the group or acquirer that now holds their records, although the group-structure filter may remove some entries. Closed companies are outside it: their archives count only where a successor holds the records and has authority to license them, which this workbook does not estimate.
Why does exclusivity change the TAM?
An exclusive AI training license removes the licensed history from the market for as long as the exclusivity runs, so historical archives behave like a stock. At this article's central inputs, new history accrues at about one-eighth of the stock each year, which is the repeatable part of supply.
What evidence would narrow the range most?
Measured buyer-fit and context-completeness rates, because those two inputs move the count most and have no public measure. A random sample drawn from an official business register, with response rates and open cases reported, would shrink the range far more than a larger frame count.
Sources
- Eurostat (9 December 2025, preliminary 2024 data): 55,000 large enterprises (more than 249 people employed) and 251,000 medium-sized enterprises (50-249 people employed) in the EU business economy; size classes use people employed only and enterprises can belong to larger groups. Source
- Eurostat's Key figures on European business, 2023 edition, reports a five-year survival rate of 46% for EU enterprises born in 2015 and still active in 2020. Source
- The Bureau of Labor Statistics reported that in March 2023, 34.7 percent of U.S. private-sector business establishments born in March 2013 were still in operation. Source
- 29 CFR 516.5 requires employers to preserve payroll records, certain agreements and certificates, and sales and purchase records for at least 3 years; 516.6 sets 2 years for supplementary records including time and earning cards, wage rate tables, and order, shipping and billing records. Source
- GDPR Article 5(1)(e) requires personal data to be kept in identifiable form no longer than necessary for the purposes of processing; Recital 39 says time limits should be established for erasure or periodic review. Source
- SourceX's how-it-works page says licenses are exclusive for AI training, with the term set in each Dataset Deal Confirmation and covering only the selected data, and that suppliers confirm the data has not been licensed for AI training before. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.