The $10 million ghost of Spirit Airlines
$10,000,000
Google agreed to buy the internal data of a dead airline. Spirit had already stopped flying, yet its records — every email, message, ticket and spreadsheet — went to the highest bidder. Here is what was sold, what it cost, and why the price tells a much bigger story about data and privacy in the age of AI.
A yellow airline that ran out of runway
Spirit Airlines was the largest ultra-low-cost carrier in the United States. After years of losses from 2020 onward, it could not survive a debt load of roughly $8.1 billion and the collapse of merger talks. Operations ceased on 2 May 2026, putting about 17,000 people out of work.
What bankruptcy does
Bankruptcy converts a dissolved company into a bundle of assets. Planes, landing slots, trademarks, and now — data. The people described by that data are not creditors in the usual sense. They do not get a vote, a cut, or a notification that their email is being weighed by the pound.
The auction frame
By the time the data lot came up, Spirit no longer existed as a going concern. That matters: the privacy policy its passengers and employees agreed to was Spirit's, and Spirit was gone. The promise did not outlive the company that made it.
A price, a counter-offer, and a price on privacy itself
The bidding is one of the most telling parts of the sale. Google opened low. Mercor, a marketplace for AI training data, pushed back. Then Mercor offered materially more for the version that had not been scrubbed. Google closed it at $10 million, with a backup bid of $7.5 million on the table.[1][2]
The $1.8 million privacy premium. Mercor was willing to pay $7 million for raw, unscrubbed data — and only $5.2 million for the clean version. The difference is the market price the bidders put on the layer of de-identification.
The proposed sale was filed on 14 August 2026 in the U.S. Bankruptcy Court for the Southern District of New York before Judge Sean H. Lane. Objections were due by 17 August; a hearing was set for 19 August 2026. After a union objection, that hearing was postponed to 9 September 2026.[2]
Numbers that should not fit in a paragraph
The package contains roughly 647 million discrete items. More than half are Microsoft Teams messages and collaboration records. The rest is a portrait of a company in motion: emails, files, tickets, flights, crew pairings, fuel slips, payroll and tax records, source code, and more.
About 1.5 cents per item
Divide $10 million by roughly 647 million items and the result is the single most useful number for a beginner: the market put an average value of about 1.5 cents on each item in this institutional memory.
Sort and filter: what Google gets, what is contested, and what is excluded
The filing lists the archive item by item. Some are clear. Some are reported differently by different outlets. And some are explicitly carved out — though even the carve-outs come with a catch.[1][2][3]
| Item | Count | Notes |
|---|
Anonymous, but threaded together
The records will be stripped of names before Google receives them. Yet the sale requires referential integrity: the linkages between records must survive the scrubbing. That is not the same as true anonymity. The diagram below shows why.
Each circle is one kind of record. The lines show that the same scrubbed identifier — not a name, but the same anonymous label — appears across email, tickets, code commits, payroll, and flight crew pairings.
- Names become IDs: "J. Doe" becomes "A-1847".
- The threads remain: email, ticket, payroll, and crew pairing still point to the same ID.
- With enough links, the ID becomes a puzzle that can often be solved.
"Anonymous but still connected across decades" is a weaker guarantee than "anonymous." A long, linked behavioral trail is easier to attach back to a real person than a scattered pile of records ever would be. The more links, the less scrubbing hides.
Who scrubs, who certifies, and who is satisfied
Every record must be de-identified before Google receives it. But the details of the privacy layer are themselves a bargain.
The scrubbing arrangement
- Google chooses the third-party de-identification vendor.
- Google pays for the vendor, with no reduction in the purchase price.
- The vendor is not named in the filing.
- No outside auditor is specified.
- The vendor must certify its work "to Google's reasonable satisfaction."
The standard, and the promise
- The standard is California Consumer Privacy Act (CCPA) de-identification, voluntarily extended to all U.S. consumer records.
- Federal health-privacy standards apply to health-related material.
- Google commits to keeping the archive de-identified and not trying to re-associate it with any person or household.
- Google stated: "We will not receive any personal information from this dataset. Any data we receive will be rigorously scrubbed of any personally identifiable information by a third party before receipt."[1]
The promise is about the receiving end: what Google gets. The requirement is that the linkages survive. That creates a tension the page can state plainly: the records must be cleaned enough to satisfy a privacy standard, but linked enough to be useful as a training corpus.
The union asked: whose privacy is this protecting?
"We will fight this every way possible." — Association of Flight Attendants-CWA, union objection
Employees, not consumers
The union argued the deal's privacy provisions were written to protect passengers, not employees. Yet the sale includes disciplinary records, payroll history, decades of internal communications, crew pairings, and workplace records.[1][2]
Trivially re-linkable
With a workforce of known size, roster, and crew base, "anonymised" crew pairings combined with payroll and email are not hard to map back to named individuals. The union's objection was sharp enough to push the approval hearing from 19 August to 9 September 2026.
Some coverage frames the wider issue as hundreds of millions of worker-touching records moving under a 1978-era bankruptcy statute — a law drafted for selling factories and machinery, not for liquidating decades of human behavior.
Why a company's exhaust is now worth more than public text
AI labs have already mined the open web. What they still lack is process: the causal chain from decision to action to consequence inside a real, operating business.
Real work, real records
A Mercor spokesperson put it directly: "Companies are sitting on decades of records that show how real work gets done."[1] Spirit's archive is exactly that — an end-to-end record of how an airline scheduled crews, priced seats, bought fuel, fixed planes, and talked to its own staff.
Bankruptcy as a mechanism
Bankruptcy makes this sale possible. It converts a dissolved company's records into an ordinary asset to be liquidated for creditors. The people whose lives are recorded in those files do not vote on the sale, receive a share, or get a notice.
The promise did not outlive the company
The privacy policy passengers and employees agreed to was Spirit's. With the airline gone, the only rights that remain are those a bankruptcy court recognizes — and the people described in the records are not at that table.
A pattern, not an accident
Every company holds a comparable archive. Every company can fail. The Spirit auction is a template for what happens to the data of the next bankrupt firm, and the next.
Where the facts come from
- 1 The Register: "Google buys crashed airline Spirit's data at auction, because AI" by Simon Sharwood, APAC Editor, published 18 August 2026.
- 2 U.S. Bankruptcy Court, S.D.N.Y.: Sale motion filed 14 August 2026 before Judge Sean H. Lane; objection deadline 17 August 2026; hearing set for 19 August 2026 and postponed to 9 September 2026 following the Association of Flight Attendants-CWA objection.
- 3 Other reporting: Some outlets describe more than 30 million recorded customer-service calls, more than 15 million chat records, 13.7 million active email addresses from Oracle Responsys, and 11 million in-flight Wi-Fi sales as part of the trove; other coverage states that call recordings, chat sessions, and disability-service requests were excluded. Both readings are presented here without choosing one.