Public figures for AI training data span a wide range. Buyers advertise roughly $100K to several million dollars per company for operational data, while reported publisher deals run from $1-5M to about $203M. Price depends on uniqueness, volume, linkage, exclusivity, refresh and rights clarity. These are reported or advertised numbers, not guarantees.
What these numbers are, and what they are not
There is no public price list for AI training data. Most licenses are private contracts, and the parties rarely disclose terms. What exists is a small set of figures that were reported by the press, compiled by analysts, or published by buyers on their own websites.
Every figure in this guide is a reported or advertised number. None is a guarantee, a quote, or a prediction of what a specific dataset will fetch. Reported deal values often bundle several things together, such as content access, API access and product partnerships. Advertised ranges describe what a buyer says it may pay, across all the companies it hopes to attract.
Read them as reference points that show the shape of the market. A real price comes only from a written offer for a specific, scoped dataset.
Public figures at a glance
The table groups the figures by source. Buyer-advertised ranges come from each company's own site and were checked on 2026-11-16. The rest come from the publications named in the last column.
| Deal or buyer | Data | Reported value | Source |
|---|---|---|---|
| Polyshares (advertised) | Company operational data | $100K to $2M and above | Polyshares site, checked 2026-11-16 |
| micro1 (advertised) | Company data partnerships | $100k-3M+ | micro1 site, checked 2026-11-16 |
| Mode (advertised) | Company operational data | $100K-$5M | Mode site, checked 2026-11-16 |
| Scale AI (illustrative) | Company data partnerships | $10K to $1M+ per partnership | Scale AI site, checked 2026-11-16 |
| Closed startups | Code base | About $25,000 | MIT Technology Review, 2026-10-13 |
| Closed startups | Workspace data | About $15,000 | MIT Technology Review, 2026-10-13 |
| Google-Reddit | Platform content | About $203M total contract value | Neudata, 2025-06-10 |
| Shutterstock | Stock media content | $25-50M per deal | Neudata, 2025-06-10 |
| Taylor & Francis-Microsoft | Academic publishing content | $10M upfront | Neudata, 2025-06-10 |
| Le Monde / Prisa-OpenAI | News content | $1-5M | Neudata, 2025-06-10 |
| Wiley | Published content | $23M one-time | Monda, 2024-10-28 |
| X firehose access | Post stream | About $42K a month | Monda, 2024-10-28 |
| Bright Data price list | Public web datasets | Up to $0.0025 per record, $250 minimum | Bright Data site, checked 2026-11-16 |
What the table shows
The figures fall into rough tiers. At the top are large content owners. Neudata's market analysis of 2025-06-10 puts the Google-Reddit agreement at about $203M in total contract value and Shutterstock's agreements at $25-50M per deal. The same analysis lists Taylor & Francis and Microsoft at $10M upfront and the Le Monde and Prisa agreements with OpenAI at $1-5M. Monda's deal list of 2024-10-28 records Wiley at $23M as a one-time fee and X firehose access at about $42K a month.
The middle tier is company operational data, which is the category most relevant to an ordinary business. Here the evidence is advertised, not reported. On their own sites, Polyshares advertises $100K to $2M and above, micro1 advertises $100k-3M+, Mode advertises $100K-$5M, and Scale AI gives an illustrative $10K to $1M+ per partnership.
At the lower end, MIT Technology Review reported on 2026-10-13 that closed startups were selling a code base for about $25,000 and workspace data for about $15,000. Public web data sits lower still. Bright Data's public dataset price list shows up to $0.0025 per record with a $250 minimum. That contrast matters. Data anyone can collect is priced per record. Data only one company holds is priced per deal.
What drives price
Across these tiers, the same factors explain most of the spread.
- Uniqueness. Data that is not public and cannot be scraped or recreated is worth more than data a buyer can get elsewhere.
- Volume and history. Larger archives that span more years show more cases, more edge conditions and more change over time.
- Cross-system linkage. Records that connect across systems, such as a support ticket tied to the order and the call that followed, show complete workflows. Linked data is harder to find than any single export.
- Exclusivity. A buyer may pay more to be the only licensee for a period. The supplier gives up other sales in exchange, so exclusivity should be a priced choice, not a default.
- Refresh. A dataset that can be updated on a schedule has continuing value. A one-time archive is priced once.
- Rights clarity. Buyers discount data when ownership, consent or contract terms are uncertain. Clean, documented rights remove a reason to pay less or walk away.
- Vertical. Specialized fields with expert-produced records, such as healthcare, law, finance and engineering, tend to draw more interest than generic office content.
How payment is structured
Price is only half of the commercial terms. The structure decides when money arrives and who carries the risk.
Many agreements combine these forms, for example an upfront payment plus a recurring fee. Neudata describes the Taylor & Francis payment from Microsoft as $10M upfront, a reminder that a headline figure may be only one part of the payment terms.
- One-time fee. A single payment for a defined copy of the data. Monda's list records the Wiley deal in this form. It is simple and certain, with no upside if the data proves more useful than expected.
- Annual license. A recurring fee for continued access, often with refreshed data. It suits datasets that keep growing.
- Usage-based. Payment tied to volume delivered or accessed, such as per record or per month. The X firehose figure and Bright Data's per-record pricing are examples of metered access.
- Revenue share. The supplier receives a portion of what the buyer or an intermediary earns from the data. It can pay more over time, but it depends on reporting the supplier cannot easily verify.
Why advertised ranges are so wide
A range that runs from $100K to several million dollars is not a price. It is a statement that outcomes differ sharply from one company to the next. Buyers publish a wide range because they do not know what a supplier holds until they see a sample and a description of rights.
The top of a range is typically reserved for large, well-linked, multi-year archives in fields buyers care about, with clear rights. A small export from a single system, with unclear consent terms, will sit near the bottom or receive no offer. Advertised ranges are also marketing. They are written to attract suppliers, and the buyer has no obligation to pay within them.
Getting from a range to a real number
The practical route is to scope a specific dataset. List the systems, the record types, the years covered and the approximate volume. Check which contracts and privacy commitments apply to each. Decide what has to be removed before anyone outside the company sees it.
Outrich Data licenses proprietary data from companies with 50+ employees and 2+ years of history, de-identifies it, and supplies it to AI labs. Suppliers pay nothing, and an offer arrives in writing with a stated validity period. That written offer, for a defined package and a defined scope of use, is the only number worth planning around.
Until then, treat every figure on this page as context. The public numbers show that proprietary data has a market and that the spread is large. They do not show what any one company's data is worth.
Common questions
Is there a standard price per record for AI training data?
Are the buyer-advertised ranges reliable?
Why are publisher deals so much larger than company data deals?
Does a non-exclusive license lower the price?
Sources
- Polyshares, advertised range for company data2026-11-16
- micro1, data partnerships page2026-11-16
- Mode, data licensing page2026-11-16
- Scale AI, data partnership page2026-11-16
- MIT Technology Review, AI labs buying data from dead startups2026-10-13
- Neudata, AI data licensing market analysis2025-06-10
- Monda, list of data licensing deals for AI2024-10-28
- Bright Data, public dataset price list2026-11-16