Google's $10 Million Purchase of Spirit Airlines' Internal Business Data

Google's winning bid for Spirit Airlines' internal business data landed at roughly $10 million ^[1]- a number that only makes sense once you see what it bought. The auction opened with Google offering $5 million; AI recruitment company Mercor countered at $7.5 million before Google's final offer closed it out, with Mercor left as the designated backup buyer ^[2]. What either company was chasing wasn't a single asset but an entire company's digital exhaust: roughly 100 million emails across 80,000 accounts, about 500 million Microsoft Teams messages, over 17 million OneDrive files, more than 20 million SharePoint files, and hundreds of thousands of IT tickets, with employee records reaching back to 1986 ^[3]. Layered on top of that is 30 million lines of source code across 516 repositories and more than 7.5 billion recorded transactions dating back to 2008 ^[4].

Divide the price by that volume and the deal looks less like a data purchase than a rounding error against what training corpora usually cost - which helps explain why an AI-focused bidder like Mercor, whose own business runs on paying people for labeled work data, was willing to fight for it. Bankruptcy auctions let a buyer acquire an entire workforce's communication trail in bulk, at a fraction of what licensing or crowdsourced collection would cost. The deal is narrower than the headline number suggests, though: it explicitly excludes Spirit's roughly 97.5 million passenger records, 52.4 million loyalty members, and 740,000 co-branded cardholders ^[5]. Google bought the airline's internal operations, not its customers - which is part of why the price could stay this low while the volume stayed this high. Short-form video commentary picked up on that irony directly, treating the purchase as a dark joke: the airline itself isn't coming back, but its data still found a buyer.

The sale agreement requires the data to be scrubbed of personal information before Google ever receives it, but it also requires the scrubbing to preserve what the contract calls 'referential integrity' across the dataset ^[4]- meaning a pseudonymized employee's email thread, support tickets, code commits, and payroll history all stay linked to the same underlying identity even after direct names are removed. That is a deliberate design choice, not an oversight: a system trained on the archive needs to see how one person's messages, decisions, and output connect over time, not a shuffled bag of disconnected text.

That detail sits awkwardly next to Google's own public reassurance. The company's spokesperson said Google 'will not receive any personal information from this dataset' because a third party will 'rigorously scrub' it first ^[6]. But Google is the one who designates, pays for, and approves that third-party scrubber, and Spirit's contractual obligation is only to give 'good faith consideration' to Google's comments on the process - not to agree with them ^[4]. Nobody outside Google has to sign off on whether the standard was actually met. Compare that to the 23andMe bankruptcy, where a court-appointed privacy ombudsman reviewed the sale of genetic data before a judge approved it ^[7]. That skepticism isn't confined to privacy researchers, either - Reddit threads independently raised the same doubt, with commenters arguing that anonymizing over 100 million emails and hundreds of millions of chat messages at this scale is functionally impossible and that aggregated records can still be re-identified once direct names are stripped. A dataset built to keep every record traceable back to one person, cleared by a process Google controls end to end, is a harder claim to take on faith than 'we removed the names.'

Spirit ceased operations and entered liquidation after failing to secure a government rescue package ^[3]. What remains, in a sense, is the paper trail: decades of emails, chat logs, HR files, and code commits that Spirit's employees generated while doing their jobs, now sold as a single lot to help train the systems built to do versions of those jobs going forward. The Association of Flight Attendants-CWA has filed a formal court objection to the sale, arguing the data 'has no business being sold' and warning that a workforce as small and identifiable as Spirit's flight attendants is harder to meaningfully anonymize than a mass consumer dataset ^[8].

The tension isn't really about whether the sale is legal. Standard employment agreements generally make workplace communications company property, and Google has publicly committed to receiving a dataset with no personal information ^[6]- a reading Reddit commenters largely converged on independently, treating the sale as legally uncontroversial even where it feels ethically uneasy. Where the debate splits is over consent that was never actually asked for: some online commentary framed the deal as a positive sign that AI labs are finally paying for data instead of scraping it, while others pushed back that the employees who wrote these messages were never consulted or compensated the way a company like Mercor - whose usual business model pays individual workers directly for old-job documents - would pay someone for comparable material. One widely-shared reaction did the per-unit math and made a similar point by contrast, framing this as bulk, wholesale acquisition of a workforce's communications rather than the retail, consent-based data deals AI companies increasingly pursue elsewhere.

This isn't the first time a bankruptcy court has had to decide what happens to a company's data when the company itself dies. In 2015, RadioShack tried to sell its customers' personal data as part of its Chapter 11 liquidation; the FTC and 38 state attorneys general objected on the grounds that the sale would violate RadioShack's own privacy policy, and the company ultimately agreed to destroy most of the data and sell only a limited slice under court-approved restrictions ^[9]. A decade later, 23andMe's 2025 bankruptcy put the genetic data of more than 15 million customers up for auction; that sale went through, for $305 million, but only after a court-appointed privacy ombudsman reviewed the transaction ahead of the bankruptcy hearing ^[10]. Both cases established something close to a norm: sensitive personal data changing hands in bankruptcy gets independent scrutiny before a judge signs off.

The Spirit sale routes around a version of that norm largely by exempting itself from the customer-data question entirely - passenger and loyalty records are explicitly excluded, which is likely why this deal hasn't drawn the same FTC attention RadioShack did. What it doesn't route around is the employee side, where there's no ombudsman in the mix and no independent audit of the de-identification standard ^[7]. That gap matters beyond this one deal: as public internet text runs low, AI labs have increasingly turned to licensing proprietary information at scale - News Corp's multi-year deal with OpenAI and Reddit's reported nine-figure annual arrangements with Google and OpenAI are part of the same shift ^[11]. A bankruptcy auction that treats a defunct company's entire internal archive as a training asset, with lighter oversight than a customer mailing list once received, is a template other distressed companies now have a clear incentive to copy ^[12].