Google paid $10 million for Spirit Airlines' internal data and custom software in a Chapter 11 auction, a purchase the company says will feed its next generation of AI models. The budget carrier collapsed earlier this year, leaving behind decades of operational records that Google's spokesperson described as helpful for improving products and AI models. Customer and credit card information were excluded from the sale.

The auction played out in steps

Google opened at $5 million. AI training startup Mercor countered at $5.2 million and signaled willingness to pay $7 million if it could receive raw data first and anonymize it itself. Google returned with $10 million and closed the deal. Court filings show the process was competitive, though the final price represents a fraction of what large language model labs routinely spend on compute.

Corporate data is the new scarce resource

Tech companies have largely exhausted the open internet for training material and are now racing to acquire proprietary corporate datasets. Mercor's spokesperson framed the Spirit bid as part of a broader pattern: companies sit on decades of records showing how real work gets done, and that material is now among the most valuable for training and evaluating AI. Micro1 earlier this year began paying 50 midsize companies between $100,000 and $2 million each for access to anonymized operational data.

The Meta precedent

Meta attempted a different shortcut this year, tracking employees' keystrokes and mouse movements to teach its AI how to use computers. The program was suspended after an internal data exposure and strong employee backlash. Google's approach, buying a bankrupt airline's sanitized archives, avoids the privacy and consent issues of monitoring live workers, but it also bets that historical airline operations generalize to broader AI capabilities.

What to watch

The Spirit purchase sets a market price for distressed corporate data: $10 million for a national carrier's internal systems. As more companies enter bankruptcy or restructure, their data estates will become a recurring asset class for AI labs. The next test is whether models trained on airline logistics, crew scheduling, and customer complaint histories actually improve general reasoning, or just produce better chatbots for lost baggage claims.