How much is business data worth to AI labs?

Nobody can tell you from the outside. Reported sales run from about $10,000 to hundreds of thousands of dollars, and the one average built from a count works out to about $10,000 a deal. Providers lead with $100,000 or more but do not say how many companies were paid. The number that matters for your company is a buyer's offer for your actual data, because no source we found publishes a rate per record, per gigabyte, or per employee.

Leaving aside the Spirit Airlines auction, a large airline's archive sold in bankruptcy, the honest summary for a single small or mid-size seller is five to six figures with no formula behind it. The marketplace average is the most useful number here because it is the only one built from a count: about $10,000 a deal suggests most sales sit near the bottom of that $10,000 to $100,000 range. That is our inference, not something the sources say.

• Different sellers. The counted average comes from shut-down startups selling archives of messages and tickets. The $100,000 deals came from running businesses selling operating records. There is no reason those would price alike.
• Different things counted. One figure is an average across nearly 100 sales, one is a single archive, one is a marketing line, and one is two deals.
• Selection. A provider's page leads with its best outcomes, and a marketplace average includes the small ones. Neither is wrong. They answer different questions.
• Different scale. A bankrupt airline's archive went to auction at eight figures. That says nothing about what a 50-person distributor's records would fetch.

That reading of the gaps is ours. The sources do not explain their own differences.

• Scarcity more than size. SimpleClosure's CEO told Upstarts that scarcity matters more than volume, and gave the example of a codebase a quarter of the size being worth double.
• What the data is. The same article lists what labs reportedly pay more for: rare programming languages, context from developer tools such as Jira, Linear and Figma, and the documents around them, such as emails, customer records and support logs.
• How far back it goes. Nyne asks for three years in business, and the large archive sale in the table covered 13 years. Neither source says how much history adds to the price.
• Whether the license is exclusive. An exclusive license can be worth more up front but means you cannot license the same data again. What is actually in a data licensing agreement covers that trade.
• Where the records live. Nyne names accounting systems, CRM and support records, and scheduling or project tools. Each is a system of record rather than a spreadsheet or someone's memory. That reading is ours, not something the sources state.

None of the sources we read says how much any one of these factors is worth. They are reasons a number moves, not a way to calculate it.

Both were arranged independently by a partner, not by Nova Solutions, and Nova Solutions has not closed a data-licensing deal itself. They show the category exists and can move fast. We are not claiming to know why either landed at $100,000, and two deals do not make a price list. They also sit well above the roughly $10,000 average implied by SimpleClosure's numbers, which is a reminder that these sources are not measuring the same sellers.

• A price list. Every figure above is a range, a headline, or an average from a different kind of seller.
• How many companies were actually paid. A page saying deals are commonly worth $100,000 or more does not say how many were done.
• What gets turned down. None of the sources we read said which data labs refused.
• A typical time to close, apart from the two deals above. We found no published average from any provider.

That is why we would treat any number quoted before someone has seen what you hold as a place to start asking questions rather than a price.

The only number that means anything is an offer for your actual data. Getting one takes a short call about your business, your systems and roughly how much data you have. Nova Solutions then prepares a de-identified sample, takes it to AI labs and model developers, and brings back the offer. You approve the buyer and the price before anything moves, you keep every original, and we are paid a percentage only if you accept an offer and the deal closes. The data licensing page explains the steps.

Before that call, it helps to know where your records live, how far back they go, and whose information is in them. How to sell business data to AI companies has the checklist and each provider's stated requirements, and Can a business legally license its own customer data? covers the legal side.

If your company is not a running business, or what you hold is research, scientific or personal data, Data licensing companies by seller type covers which routes exist for other sellers.

• AI Weekly, Mercor chases shuttered startups' Slack and Jira archives, August 15, 2026. Summarizes reporting by Forbes, The Information and Fast Company.
• Upstarts Media, What Fetches Top Dollar On The Dead Startup Data Marketplace?, September 3, 2026.
• Nyne, Company data partnerships, read October 4 and October 9, 2026.
• micro1, Data partnerships, read October 9, 2026.
• Chapter 11 Cases, The Spirit Airlines deidentified data sale: auction results, objections, and the September 30 hearing, September 13, 2026.
• Fortune, report on micro1's bid for Spirit Airlines' data, September 14, 2026.
• FlightGlobal, US lawmakers urge Google to exclude worker details from Spirit Airlines data sale, October 8, 2026.
• aitrainer.work, enterprise provider comparison and how companies get paid for operational data, read October 4, 2026. A referral site that discloses fees from some buyers.
• Nova Solutions, data licensing, for the two deals and the terms of our own arrangement.