Google wins auction for Spirit Airlines' data trove to train AI

Google agreed to pay $10 million for a trove of the defunct Spirit Airlines’ internal data, according to the bankruptcy court’s notice of the auction results, which named Google the successful bidder over a competing $7.5 million offer from Mercor, a company that supplies AI labs with human-annotated training data. The virtual auction was held on 14 August 2026. Spirit ended its flights in May 2026 after filing for bankruptcy, and the same notice scheduled a hearing to approve the sale for 19 August — but the court later postponed that hearing to 9 September after Spirit’s flight-attendant union objected.

Reporting on the sale said the trove includes roughly 100 million internal emails and 500 million Microsoft Teams messages, alongside corporate records covering operations, revenue, aircraft maintenance and pricing, and — per Tom’s Hardware — billions of flight-pricing records and anonymised passenger records. Coverage of the underlying sale filing said personal data was excluded, including Spirit’s roughly 97.5 million passenger profiles, its Free Spirit loyalty-programme records and customers’ credit-card information; Google said it would not receive personal information from the dataset and would use what remained to train its AI systems. Mercor, the underbidder, is named in the court’s notice as the alternate bidder should Google’s purchase not close.

The postponement followed a limited objection from the Association of Flight Attendants-CWA, Spirit’s flight-attendant union, which argued, Fortune reported, that “the privacy architecture of this transaction is consumer-facing” while “its payload is disproportionately employee-facing” — leaving payroll records, disciplinary files and internal communications less protected than passenger data — and that “referential integrity” preserved across the data set could still let de-identified employee records be traced back to individuals. The union asked the court to exclude all flight-attendant data from the sale as a condition of approval.

The transaction illustrated a route into AI training data that bypasses the licensing negotiations more familiar from disputes with publishers or platform operators: a company’s internal correspondence and operational records, generated by employees who had no say in the matter, sold as a liquidation asset once the business itself had stopped trading. It followed a pattern, seen elsewhere in the industry that year, of AI developers acquiring large bodies of proprietary corporate data through corporate distress rather than direct licensing deals.