How Much Is Company Data Worth to AI Companies? Valuation Methods and Public Price Points

Company data is worth what an AI buyer will pay for the right to train on it, usually under a license rather than a sale, and no public rate card exists. The documented figures each have a named source: Mercor states listing ranges of USD 10,000 to 1,500,000 per company (checked 2026-11-09), a bankruptcy court records Google LLC's successful bid of USD 10,000,000 for Spirit Airlines' de-identified data (2026-08-14, not approved as of November 9, 2026), and Reddit's S-1 of 2024-02-22 discloses USD 203.0 million in aggregate contract value for data licenses with unnamed counterparties.

Valuing company data is one part of AI training data licensing: three methods (cost, market and income) apply, six drivers move the result, and a buyer program's payout range is a claim, not a deal value. The defensible number is a written offer on comparable terms from more than one buyer, and not selling remains a valid result.

DataPacts Editorial Team checked every figure here against its primary source on November 9, 2026; the page on how we source, date and check every figure sets out the rules.

What Data Valuation Means When the Buyer Is an AI Company

Data valuation is the estimate of what a data asset is worth to a specific buyer for a specific use, and for AI training that use is a license to train models. The buyer pays for training rights, not for the files: micro1's Data License Master Terms (updated 2026-11-08) take a worldwide, royalty-free, nonexclusive, irrevocable and perpetual grant, and the seller keeps its data. The seller can license the data again where the terms allow, but cannot take back what a model has learned: Labelbox states on its Prism FAQ that "Models already trained on the data cannot be unlearned" (checked 2026-11-09).

In machine learning research, Data Shapley (Ghorbani and Zou, ICML 2019, PMLR 97:2242-2251) scores each training example's contribution to a model and sets no price.

Three numbers therefore differ for the same archive: its value to one buyer, its value to the seller, and its book value, which for internally built data is usually absent from the balance sheet.

Three Methods to Value Company Data for AI Licensing

Data is valued with the cost, market or income approach, and each answers a different question. The table compares the three methods by their public evidence for AI training data, as of 2026-11-09, with buyer estimators as a fourth row.

Cost Approach: What the Data Cost to Create or Replace

The cost approach values data at what it cost to create or would cost to replace, which sets a floor for negotiation but not a price. The floor is rarely visible in the accounts, because IAS 38.63 bars recognition of internally generated customer lists and items similar in substance as intangible assets. The three cost groups in that floor are listed below.

• Collection and creation: the staff time and systems that produced the records, tickets, documents or code.
• Storage and extraction: keeping the archive and exporting it in a format a buyer can load.
• Cleaning and de-identification: removing personal data, duplicates and third-party material.

Mercor states that it covers all costs for extraction, anonymization and transfer (checked 2026-11-09), which lowers the seller's cost floor without raising the price.

Market Approach: Comparable Public Deals and Price Points

The market approach compares data with disclosed deals and published prices, and for AI training data few deal values are disclosed by a party. Tempus AI's 10-K for fiscal 2025 (filed 2026-02-24) states USD 200 million in data license fees from Pathos over three years, including USD 50 million upfront, "for the sole purpose of" training a foundation model. Its 10-Q for the third quarter of 2026 adds a Neuro Order Form of 2026-10-22 worth up to USD 55.0 million, at USD 100 per delivered record plus USD 5.0 million for updates. The Authors Guild reported on 2024-11-19 that HarperCollins offered USD 5,000 per title, split 50-50 with the author, for a license to an unnamed AI company.

Comparables transfer badly, because each deal prices one data type, one rights package and one term: clinical records and book titles set no rate for an archive of tickets or email. Press-only deal values are left out here.

Income Approach: Licensing Income Over the Term

The income approach values data by the license income it can earn over the term, discounted for risk, and listed sellers' filings show how uneven that income is. Reddit's S-1 (2024-02-22) discloses January 2024 data licenses worth USD 203.0 million in aggregate over two to three years, with at least USD 66.4 million expected as 2024 revenue. Its 10-Q for the third quarter of 2026 states that "substantially all" of that contract value comes from two counterparties whose agreements are "approaching the end of their respective current terms". A recurring fee is the other pattern: the Kopinor agreement with the National Library of Norway (signed 2025-12-19) sets NOK 45,000,000 for 2026, indexed from 2027 (s.17).

Wiley's 10-Q gives AI license revenue of USD 13.7 million for the quarter to 2026-07-31, against USD 28.9 million a year earlier; its fiscal 2026 earnings release reports USD 49 million of total AI revenue, a broader base. Informa reported AI-related revenues for Taylor & Francis of USD 75m+ in 2024 and called the data access revenue "non-recurring". Each company's record sits under AI data licensing revenue in company filings.

Buyer Estimators and Why They Are Not a Valuation

A buyer's data value calculator shows that buyer's illustrative range for its own program, not a valuation of the seller's data. Handshake AI's data partnerships page carries the hero line "Earn $100K to $4M+" and, on the same page, a calculator headed "Your data could be worth up to $100,000" (both checked 2026-11-09). micro1's estimator shows an example output of USD 50,284 to 92,776, below the USD 100K floor of its own lowest tier (checked 2026-11-09). Two more estimators are listed below.

• Nyne's estimator returned four ranges for the same default inputs (100 employees, 10 years, USA) on four page loads on 2026-11-09: USD 51,996 to 121,745, 56,037 to 131,207, 41,420 to 96,982 and 38,632 to 90,454.
• Pewter, a broker, shows USD 385,000 to 895,000 for 100 staff and 10 years, labeled "A scenario, not a valuation" (checked 2026-11-09).

Public Price Points and Why They Do Not Transfer to Company Data

Public per-unit prices for AI training data exist for five units (title, hour, image, token and record), each stated by a named party on a known date, and none is a market rate. The six reference points below, as of 2026-11-09, include one court settlement, which is not a license price.

• Per title: USD 5,000 per HarperCollins title, split 50-50 with the author (Authors Guild, 2024-11-19).
• Per hour: SpeechData.ai states USD 60 to 95 per hour of audio; Claru states USD 10 to 120 per hour of capture (both checked 2026-11-09).
• Per image: Bee Maps states USD 0.005 per street-level image and USD 6 per AI event video (checked 2026-11-09).
• Per token: Pangea states indicative rates of USD 100 per million tokens of source code and USD 40 for git history, changeable on 30 days' notice (checked 2026-11-09).
• Per record: USD 100 per delivered record under the Tempus and Pathos Neuro Order Form of 2026-10-22.
• Settlement, not a price: Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO) ended in a USD 1.5 billion non-reversionary settlement, final approval 2026-07-20 (Dkt. 680); it sets no license rate.

Unit prices come from sellers with standard formats and clean rights, such as speech catalogs. A company archive mixes formats, customer contracts and personal data and is priced as one package, so size times a unit rate misstates it. The dated table of AI training data prices per work, hour, image, token and record tracks each rate with its source.

What Drives the Value of Company Data Up or Down

Six drivers move the value of company data, and rights, consent and personal data can take it to zero. The table shows, as of 2026-11-09, when each driver raises or lowers value and the public evidence for it.

Rights and Chain of Title

Data is worth only what the seller can license, so unclear ownership lowers value more than any other driver. Under s.5.1 of micro1's master terms, the licensor warrants that it has not granted a third party rights to the same data on the same or substantially similar terms. Material a seller cannot warrant has no value to the buyer and, because a trained model cannot unlearn data, is excluded before pricing.

Customer Consent and Confidentiality

Customer contracts and confidentiality clauses can bar a license of data the company holds, which removes that data from the sale rather than lowering its price; no party has published a figure for that effect as of 2026-11-09. Buyers write the condition into their requirements: Mercor's listing for AI product companies asks for "User terms that allow licensing de-identified usage data, or a way to get consent", and its call center listing excludes "Recordings made without consent" (both checked 2026-11-09).

Personal Data, GDPR and De-identification

Personal data lowers value through the cost and the loss of de-identification, and the Spirit Airlines sale record shows the standards one court-supervised sale used. For EU personal data, the rules on GDPR and AI training data decide whether a dataset can be offered at all, starting with its legal basis.

The Spirit Airlines Bill of Sale in Bankr. S.D.N.Y. 25-11897 (Dkt. 1463, 2026-08-14) requires de-identification under Cal. Civ. Code 1798.140(m), plus 45 CFR 164.514 for health data, and names Tonic.ai as the Deidentification Agent, so Google LLC's successful bid prices de-identified data only (sale not approved as of 2026-11-09).

Exclusivity and Other Licensing Agreement Terms

Exclusivity raises what one buyer will pay and shuts out every other buyer, and published licensing agreement terms take three forms, all checked 2026-11-09.

• Exclusive by default: Labelbox states that Prism licenses are exclusive by default; the seller keeps ownership.
• Time-limited exclusive window: Turing states it is sole seller for the first 90 days of Project Lazarus; lab licenses are non-exclusive.
• Non-exclusive with a no-prior-grant warranty: micro1 pairs a nonexclusive grant with the s.5.1 warranty; Endless, an aggregator placing company data with buyers, states its licenses are non-exclusive.

No party has published an exclusivity premium for AI training data as of 2026-11-09, so the price effect is set in negotiation and written into the exclusivity clause of an AI data licensing agreement.

Uniqueness and Scarcity

Data a buyer cannot get elsewhere is worth more, and duplicated or third-party material is removed before pricing. SimpleClosure states in a blog post of 2026-09-03 that about 37 % of the raw code offered through its AssetHub is removed in cleaning, as third-party or generated code and duplication. micro1 reserves its highest payout tier for highly unique proprietary data (checked 2026-11-09).

Volume, History and Freshness

Volume sets deal size only after rights and quality are settled; history and freshness decide whether a buyer pays once or keeps paying. The Tempus and Pathos Neuro Order Form of 2026-10-22 prices updates separately at USD 5.0 million of its USD 55.0 million, and Reddit's January 2024 licenses run on continuous API access. Mercor's listing for consulting firms asks for 5+ years of engagements (checked 2026-11-09).

Buyer Program Payout Claims vs Public Deal Values

A buyer program's payout range is that buyer's claim about its own offers, while a disclosed deal value is a figure a party filed, signed or a court recorded. The table sets both side by side as of 2026-11-09: group A holds program claims by buyers and intermediaries, group B holds values stated by a party or a court, alphabetical within each group.

What Buyer Programs Claim, and Where Their Pages Disagree

Buyer programs publish ranges from tens of thousands to several million US dollars, each the buyer's own statement, checked November 9, 2026. The page of AI data partnership payouts and program terms compared keeps all of them with check dates. micro1 also states USD 1 to 2M a year for agent trajectories, and Handshake states "$250M, Paid out to partners" without saying whether that covers data licenses (both checked 2026-11-09).

Four programs contradict themselves on their own pages, all checked 2026-11-09. Mercor's feed shows USD 17,000,000 paid to companies in one block and USD 50,000,000 in another, and Handshake's hero range and calculator disagree. Labelbox promises a written quote within 48 hours on its home page and a preliminary valuation within one business day in its FAQ. micro1's master terms give 30 business days to reject with deemed acceptance (s.1.3), while its program page says "30 days". No seller has publicly confirmed an amount received under any of these programs as of 2026-11-09.

What Disclosed AI Data Licensing Deals Show

Disclosed AI data licensing deals are few, and each value carries the name of the party or court that stated it. The four in the table come from SEC filings by the seller (Reddit 2024, Tempus 2026), a signed agreement (Kopinor, 2025-12-19) and a bankruptcy docket whose sale hearing is set for 2026-11-17 (Dkt. 1779).

A deal value works as a comparable only within one data type and rights package, with its stating document attached. The AI data licensing deals tracker lists each deal with parties, dates, any party-stated value and its source.

Not Selling: When Keeping the Data Is Worth More

Not selling is a valid result when the license price is below the data's value to the company or below the risk the sale creates. A license is hard to reverse: micro1's master terms make the grant irrevocable and perpetual and, under s.4.4, require deletion within 120 days of delivery with no duty to retrain models (terms updated 2026-11-08).

Data income can also fall away. Shutterstock reported USD 203.3 million of Data, Distribution, and Services revenue (a revenue category including AI data licensing) for 2025 (10-K), then USD 15.8 million for the third quarter of 2026 against USD 65.7 million a year earlier, down 76 %, with its data offering down 91 % (10-Q). The four sides of the decision are listed below.

Benefits of licensing company data:

• Cash for an asset that usually carries no book value.
• Continued ownership and use under a non-exclusive license.

Risks of licensing company data:

• A perpetual, irrevocable grant that cannot be recalled after delivery.
• Breach of customer contracts or privacy law where consent or de-identification falls short.

Downsides of licensing company data:

• Staff time for the inventory, rights check and de-identification before any offer.
• Exclusivity that shuts out later buyers, and income that can drop after the first term.

Alternatives, including not selling at all:

• Keep the data and use it internally.
• License a de-identified subset, non-exclusively or for a fixed term.

Each option is set out under alternatives to selling data to AI labs.

How to Reach a Defensible Number for Your Data

A defensible number comes from written offers on comparable terms from more than one buyer, built on an inventory and a rights check, not from a calculator. The five steps run in this order.

1. Inventory the data: systems, formats, volume and years of history.
2. Check rights and consent, and exclude what contracts or third-party rights bar.
3. Set the de-identification scope and the standard it follows.
4. Describe the data in a manifest and send it to more than one buyer.
5. Compare written offers on the same terms: price, payment trigger, acceptance period and license grant.

Equal prices hide unequal terms: Mercor states payment typically 2 to 4 weeks after signing, Mode about three months, and micro1's terms allow 30 business days to reject a delivery (checked 2026-11-09). The method for evaluating an offer for your data puts those terms on one scale.

Where Valuation Sits in the Process to Sell or License Company Data

Valuation follows the inventory, rights and de-identification steps and comes before choosing a buyer and negotiating the contract. The seven steps around valuation are listed below in order.

1. Build a data inventory for an AI data sale.
2. Run a rights audit.
3. Review customer consent and confidentiality clauses.
4. Scope PII removal and de-identification.
5. Value the data, then describe it in the data manifest.
6. Move to choosing a buyer for your data.
7. Finish by negotiating an AI data licensing deal.

The full sequence to sell or license company data to AI companies runs from readiness to payment; whether to start is covered under should you sell or license your data.

Accounting and Tax Treatment of Data Value

Under IAS 38.63 internally generated data archives are not recognized as intangible assets (the standard names customer lists and items similar in substance), so a license fee usually has no book value to compare against. A gain on an outright sale "shall not be classified as revenue" under IAS 38.113, while license fees follow IFRS 15 B56 or ASC 606-10-55-58A and 55-58B (over time or at a point in time, per ASU 2016-10). See accounting for data licensing revenue and tax on data licensing income; this page is information, not accounting, tax or legal advice.

Valuing Data From a Defunct Company or in Insolvency

Data from a closed or insolvent company is valued the same way, but the seller is often an administrator or trustee and an auction or offers set the price. SimpleClosure states that AssetHub deals run from tens to hundreds of thousands of dollars, and micro1 states company compensation of USD 100K to 3M+ in its defunct-company program (both checked 2026-11-09). The same methods apply to selling data from a defunct company. The Spirit Airlines auction shows how a court sale of data assets in insolvency sets a price.

Buyer Profiles and Program Pages Cited on This Page

The programs and intermediaries cited above are listed alphabetically; each figure comes from their own pages, checked November 9, 2026.

• Endless (aggregator)
• Handshake AI: Handshake AI data buyer profile
• Labelbox Prism
• Mercor: Mercor data buyer profile
• micro1: micro1 data buyer profile
• Mode
• Nyne
• Pewter (broker)
• Turing Project Lazarus

The overview of company data partnership programs compares them, and who buys AI training data covers every buyer.

Frequently Asked Questions

How much do companies sell data for?

Public figures for company data run from five-figure buyer listings to nine-figure filed contracts, and most are buyer claims, not paid amounts. Mercor states listings of USD 10,000 to 1,500,000 (checked 2026-11-09); Reddit's 2024 S-1 discloses USD 203.0 million in aggregate. Advertised and confirmed amounts are compared under reported payouts to data sellers.

What is data valuation?

Data valuation estimates the monetary value of a data asset with the cost, market or income approach. For AI training, the asset is usually licensed, so the value is the license income a buyer pays over the term.

Is there a price per record for company data?

No general price per record exists; public per-record prices come from single contracts such as Tempus and Pathos at USD 100 per delivered record. That order form of 2026-10-22 covers 500,000 clinical records over five years; tickets, email and CRM entries have no party-stated rate as of 2026-11-09.

Is a buyer's data value calculator a valuation?

No; a calculator shows one buyer's illustrative range for its own program. Nyne's estimator returned four ranges for the same inputs on 2026-11-09, and Pewter labels its output "A scenario, not a valuation". The published calculators are run on one company profile under buyer estimators side by side.