The most valuable thing Spirit Airlines owned at the end wasn't its planes. It was its inbox.

When a company dies, its assets get auctioned — gates, aircraft, brand, the usual wreckage of a bankruptcy. Spirit Airlines had something less obvious on the block, and Google wanted it badly enough to pay $10 million and outbid a rival to get it. Not the fleet. The data: roughly 100 million internal employee emails, 500 million Microsoft Teams chats, billions of flight-pricing and transaction records, and more than a hundred thousand employee records stretching back decades, Axios reported.

The purchase is small in dollars and large in signal. AI training data — specifically the proprietary operational record of how a real company actually ran — has become scarce and valuable enough that a trillion-dollar tech giant will comb through a bankruptcy estate to find it. The open web has been strip-mined. What Google bought in Spirit's corpse is the kind of data the web never contained: the internal texture of real work.

What Google actually bought, and what it left behind

The scale of the trove is striking. According to reporting on the deal, the package includes about 100 million employee emails, 500 million Teams chats, some 7.2 billion records covering competitors' flights, 7.5 billion passenger transaction records dating to 2008, and more than 175,000 employee records going back to 1986, as Business Insider detailed. Google won the auction at $10 million, beating a $7.5 million bid from Mercor, an AI-focused firm named as the backup buyer. A federal bankruptcy judge in the Southern District of New York was scheduled to sign off on the transaction this week.

What Google left behind matters as much as what it took. The deal explicitly excludes personal and customer data — the roughly 97.5 million passenger profiles Spirit kept and the 50.2 million records from its Free Spirit loyalty program are not part of the sale. The data is described as de-identified, and a third party is set to scrub it before transfer. Google says it will use the material for product development and to improve the training of its AI models.

Slow down on why operational data like this is different from a web scrape. A model trained on the public internet learns from articles, forums, and marketing copy — the polished, outward-facing exhaust of human activity. It rarely sees the inside: how a real organization coordinated a schedule change, escalated a problem, argued through a decision, or handled a thousand mundane operational tangles over email and chat. That interior record — messy, contextual, decision-rich — is exactly what a company trying to build workplace AI agents most wants, because it shows work as it is actually done rather than as it is publicly described.

Why AI training data is the new frontier

For years, the assumption behind large models was that data was effectively unlimited — the web was vast, and more scraping always meant more to learn. That era is ending. The highest-quality public text has largely been consumed, licensing fights have made new sources expensive, and the marginal web page adds little. The competitive question has shifted from how much data can you scrape to what data can you get that your rivals cannot.

Proprietary enterprise data answers that question. It is scarce by definition — locked inside private companies — and it captures the operational reality that generic web data misses. Google's willingness to chase Spirit's records through a bankruptcy court is a concrete demonstration that the frontier of AI advantage is moving from model architecture toward exclusive access to information no one else can train on. Bankruptcy auctions are an unusually clean way to acquire that: a distressed estate is motivated to sell, the data is real and complete, and the transaction can be structured to exclude the personal information that would otherwise make it radioactive.

This reframes what a company's data is worth. Every large organization is sitting on an internal record of how it operates — communications, decisions, workflows. In a world where that record is training fuel for workplace AI, it is an asset with a market, not just an IT liability to be retained and forgotten.

Stakeholder analysis: who this reshapes

For Google and its rivals, the deal is a template. If proprietary operational data is the scarce input for building capable workplace agents, expect more tech companies to pursue it wherever it can be legally acquired — bankruptcy estates, licensing deals, partnerships. The advantage goes to buyers with the capital and legal machinery to move quickly and structure acquisitions that avoid personal data. A $10 million check is a rounding error for Google against the potential value of teaching an AI how real enterprises actually work.

For companies with large data troves, the message is that internal records have latent value most balance sheets ignore. That cuts both ways: it is a potential asset to license, and a governance responsibility to manage. Data retained for years "just in case" is now data that could one day be sold, subpoenaed, or scraped into a model — which raises the stakes on retention policies, contractual terms with employees, and what "de-identified" really guarantees.

For employees and privacy advocates, the deal is unsettling even with personal data excluded. The material being transferred includes 100 million internal emails and half a billion chats written by people who never imagined their workplace communications would train an AI. A flight attendants' union publicly criticized the sale as outrageous. Even scrubbed and de-identified, the transfer raises hard questions about consent, about what can be inferred from "anonymized" communications at scale, and about whether the people who generated the data have any say in its afterlife.

For the AI industry broadly, Spirit's auction is a marker of how the data economy is maturing. The scramble is no longer for more public text but for exclusive, high-context, real-world data — and that changes the competitive dynamics. Advantage flows to those who can find, buy, and legally process proprietary information, not just those who can train the largest model on the same public corpus everyone else has.

The takeaway for operators

Two durable lessons come out of a $10 million line item in an airline's bankruptcy. The first is that in AI, the scarce resource is quietly shifting from compute and model design toward proprietary data. When the open web is exhausted, the edge belongs to whoever holds — or can acquire — information their competitors can't. Google reading the room and buying Spirit's operational history is a preview of a broader hunt.

The second is that every company should now understand its own internal data as a strategic asset with a real market and a real risk profile. The emails, the chats, the operational logs that accumulate as byproducts of running a business are, in the age of workplace AI, training fuel. That means value to be captured and exposure to be managed, and most organizations have thought carefully about neither.

Spirit Airlines could not make the economics of cheap flights work, and it went under. But the record of how it tried — millions of messages, billions of transactions, the full operational memory of a company — turned out to be worth more to an AI lab than to the airline that created it. The planes were the business. The data, it turns out, was the asset that outlived it.

Frequently Asked Questions

What did Google buy from Spirit Airlines?

Google agreed to pay $10 million for a large trove of Spirit Airlines' de-identified internal data through a bankruptcy auction, including roughly 100 million employee emails, 500 million Microsoft Teams chats, billions of flight-pricing and passenger-transaction records, and more than 175,000 employee records. It plans to use the data for product development and AI model training.

Does the data include customer personal information?

No. The deal explicitly excludes personal and customer data, including about 97.5 million passenger profiles and 50.2 million Free Spirit loyalty program records. The transferred data is described as de-identified and is set to be scrubbed by a third party before transfer.

Why is proprietary enterprise data valuable for AI?

Public web data has largely been exhausted and rarely captures how organizations actually operate internally. Proprietary operational records — real emails, chats, and decisions — show work as it is genuinely done, which is exactly what companies building workplace AI agents want and cannot get from scraping the open internet.

Who else bid on the Spirit Airlines data?

Google outbid Mercor, an AI-focused firm, which offered $7.5 million and was named as the backup buyer. A federal bankruptcy judge in the Southern District of New York was set to review and approve the transaction this week.

Editor's note — sources: Axios; Business Insider (via Yahoo Finance); Tom's Hardware; Forbes.

Share this article: