Where does the data left behind on a bankrupt company's computers and servers go after liquidation? It's easy to assume it's simply processed quietly as part of bankruptcy proceedings. But when Spirit Airlines halted operations and entered liquidation in May 2026, its internal data instead became the target of a bidding war among AI companies. Google put up $10 million (roughly ¥1.59 billion at ¥159 to the dollar) to win the entire data package, edging out AI training-data company Mercor, the runner-up, by a margin of just $2.5 million. The approval hearing is scheduled for today, August 19, 2026, before Judge Sean H. Lane of the U.S. Bankruptcy Court for the Southern District of New York. If approved, it will mark a turning point where a company's internal communications data is formally priced as an AI training asset.
Inside the $10 Million Winning Bid: From Emails to Call Recordings
The list of assets confirmed in the bidding includes 100 million emails, 500 million Microsoft Teams exchanges, 17 million files stored on OneDrive, and 30 million call recordings. Flight operations data is also part of the package — 763,000 flight records and 787,000 parts-purchase records are being transferred as-is. This is an archive that blends unstructured and structured data, effectively capturing the company's decision-making processes themselves.
The company that pushed Google hardest in the auction was Mercor, which supplies training data to AI companies. Mercor's bid came in at $7.5 million, a gap of only $2.5 million from Google's winning bid. The approval hearing is set to begin at 11:00 a.m. Eastern Time; for now, Google's win remains pending court approval, and the $10 million payment will only be made once approval is granted.
Simply adding up the counts of emails, Teams messages, OneDrive files, and call recordings comes to 647 million items. Dividing the $10 million price by that count works out to roughly $0.015 per item (a little over ¥2). The per-item price looks tiny, but the fact that it adds up to $10 million in aggregate shows that monetizing internal corporate data is a business built on sheer volume.
Spirit had been pursuing a merger with JetBlue in 2024, but the deal was blocked, severely weakening the airline's financial position. It filed for Chapter 11 bankruptcy protection in November of that year, completed a restructuring in March 2025, but failed to turn the business around and filed for Chapter 11 a second time on August 29 of that year. It ceased operations on May 2, 2026, entering liquidation. The current bidding is part of this second and final liquidation proceeding, through which Spirit as a corporate entity will effectively cease to exist.
Why Does Google Want Even the Call Recordings? The Demand Structure Behind Internal Data
A Google spokesperson told Forbes that "this data will help improve our products and AI models," according to reports. However, it hasn't been disclosed which products or training models the data will be used for, nor what technical methods will be used for de-identification.
Datasets made up of public web pages and news articles are said to have already been largely exhausted among major AI companies. Corporate internal emails, chats, and call records, on the other hand, contain information not found in public data — the actual process of business decision-making, such as the terms under which parts were ordered or how flight cancellations were coordinated internally, preserved in the natural, conversational style of real exchanges. Chains of request-and-response — replies to order emails, internal coordination during flight cancellations — are difficult to reproduce from standalone pieces of text. The general view within the training-data industry is that this kind of raw business dialogue is rarer than public data when it comes to teaching AI models practical contextual understanding.
The fact that Mercor countered with a bid as high as $7.5 million is itself evidence of how strong demand is for this type of data. Mercor's core business is producing human-curated training data for AI companies, and the gap between its bid and Google's was a mere $2.5 million. A specialized AI training-data company going head-to-head with a platform giant over a bankrupt company's data represents a new dynamic, distinct from the public-dataset scrambles of the past.
An airline's internal data includes highly specialized exchanges — records of flight cancellation decisions, parts procurement negotiations — that differ from ordinary business chat. These industry-specific decision-making processes are hard to replicate with generic web text, and are seen as especially valuable for companies developing industry-specific AI models. The inclusion of flight records and parts-purchase records among Spirit's assets suggests Google's interest may extend to leveraging industry-specific data as well.
Excluded: 97.5 Million Passengers' Data — Yet the Boundary Remains Blurry
Excluded from the winning bid were the personal data of 97.5 million passenger profiles, roughly 50 million loyalty program members, and 740,000 co-branded credit card holders. This is the data an airline accumulates for customer management purposes, and it represents the most sensitive information cluster, linking names, contact details, and flight histories. A Google spokesperson told Forbes that the company "will not be receiving personal information," according to reports.
The winning bid does, however, include 13.7 million email addresses accumulated in the Oracle Responsys marketing distribution system, described as primarily a marketing list used for promotions to members. The contract terms stipulate that the practical work of de-identification will be entrusted to a third-party vendor, but that vendor must be one "the buyer can approve of, or the buyer designates," with Google bearing the full cost. While structured to involve a third party, the arrangement leaves the selection of the vendor, the cost-bearing, and the approval all in Google's hands as the buyer — this is not a setup that amounts to independent auditing.
The de-identification standard set out in the contract reportedly requires compliance with the level of California's Consumer Privacy Act (CCPA) and federal medical privacy regulations, but the specific procedures for how which identifiers will be processed have not been disclosed. If only direct identifiers such as names and customer IDs are mechanically stripped out, there remains room for individuals to be inferred from surrounding details — flight numbers, dates, amounts — still present in the text of emails and call recordings. The fact that the contract requires de-identification to preserve referential integrity across different records — the linking relationships between records — also fails to fully close off this possibility.
"Outrageous": The Union's Objection to Non-Consensual Secondary Use
Ahead of the approval hearing, the flight attendants' union AFA-CWA raised an objection over this deal. President Sara Nelson told Forbes: "This is outrageous! We are filing a court objection to Google's attempt to buy data that has no business being sold." The sheer force of this statement reflects the union's alarm over the nature of the data in question.
What the union objects to includes not only passengers' personal information but also employees' work communications. The vast majority of the 100 million emails, 500 million Teams records, and 30 million call recordings included in the winning bid consists of business communications exchanged between employees, and between employees and business partners. The core of the union's objection is whether, simply because the company went bankrupt, the statements and judgments left behind by its employees can be repurposed as AI training material without their consent.
Whether, and how far, the court weighs the court objection Sara Nelson has signaled will be the focal point of today's approval hearing. In Chapter 11 asset sale proceedings, labor unions have standing as parties to present their views, but the final decision rests with Judge Sean H. Lane, and it remains unclear at this point whether the union's objection will affect the outcome.
The question of non-consensual secondary use of data could well resurface at other bankrupt companies beyond Spirit. Liquidation proceedings have traditionally focused their scrutiny mainly on consumer protection and distribution to creditors, and it's hard to say that current bankruptcy procedures incorporate a clear framework for obtaining consent when the business data of former, already-departed employees is included among assets being sold.
Two Precedents, Two Different Outcomes: RadioShack in 2015, 23andMe in 2025
This isn't the first time friction has arisen over a bankrupt company selling customer data. In 2015, when electronics retailer RadioShack tried to sell customer data as part of its bankruptcy proceedings, the FTC (Federal Trade Commission) and 38 states intervened, and the scope of the sale was significantly narrowed. In the end, the data sale went through only in a limited form, restricted to uses consistent with the privacy policy RadioShack had promised customers at the time the data was collected.
In 2025, a similar battle played out over the customer data of genetic testing service 23andMe. In that case, TTAM Research Institute, a nonprofit led by the company's co-founder, outbid pharmaceutical giant Regeneron Pharmaceuticals in the final round of the auction. TTAM's offer of $305 million exceeded the $256 million Regeneron had previously agreed to. In redirecting the data's destination from an external for-profit company to a nonprofit led by the founder, this outcome differed from RadioShack's regulator-driven scope reduction. In both cases, the nature of the data and the intent of the buyer ultimately determined whether the bankruptcy court would approve the asset sale.
What sets Spirit's case apart from both is that the primary focus of the data in question is not passengers' personal information but internal communications between employees. The interventions in the RadioShack and 23andMe cases centered on protecting consumer data, but here, since personal information itself has already been excluded, the grounds for regulatory intervention may be comparatively weaker. The two prior precedents, centered on consumer protection, leave a gap in the framework that doesn't directly apply to the issue at stake here — employees' business data.
100 Times the Going Rate: What the Spirit Deal Reveals About Pricing a New Asset Class
At Asset Hub, a specialized broker that handles liquidation data from bankrupt companies, comparable internal data deals typically go for $10,000 to $100,000 per company. Dividing Spirit's $10 million deal by the upper end of that range — $100,000 — yields a multiple of 100x. In other words, a corporate internal data asset in liquidation was, for once, priced at 100 times the norm.
The sheer scale of the data — 100 million emails, 500 million Teams records — is one factor that made the Spirit deal stand out. But internal data assets of comparable scale could well emerge at other large bankrupt companies going forward. The 100x multiple should be read as a number signaling that the AI training-data market has begun seriously pricing a new asset class: corporate internal communications. Cases in which the internal data of failed companies becomes the target of a bidding war among AI companies during liquidation proceedings may well spread beyond Spirit going forward.
Specialized brokers that bridge liquidated companies' data toward AI training use are said, as an industry-context observation, to exist in some number beyond Asset Hub alone. However, most deals to date have stayed in the range of tens of thousands of dollars per company, making the publicly reported $10 million figure in this case exceptionally large. If today's winning bid is approved, it could serve as a price benchmark for other bankrupt companies and buyers entering similar negotiations.
For Mercor, the outcome means falling short even after putting up a bid as high as $7.5 million. Whether the company continues participating in similar auctions going forward, or instead focuses on smaller-scale liquidation deals, will be a telling indicator of how far price competition within the AI training-data industry will go from here.
For Japanese readers, $10 million comes to roughly ¥1.59 billion at ¥159 to the dollar. In Japan too, how accumulated data is handled when a company goes bankrupt and enters liquidation could become an issue under the Act on the Protection of Personal Information, but there are few precedents of internal communications data being bought and sold for AI training purposes, meaning Spirit's case could well be referenced as a leading example. If a similar case were to occur in Japan, even more careful procedures than seen here might be required regarding the handling of sensitive personal information and the consent requirements for third-party data transfers.
If Judge Sean H. Lane approves Google's winning bid at today's hearing, the practice of pricing a company's internal data as an AI training asset will gain the public endorsement of a bankruptcy court's approval process. Conversely, if conditions are attached in response to the union's objection, it will be recorded as a case where RadioShack-style after-the-fact intervention worked again. Whichever way it ends, this will stand as the first case testing just how much value the AI industry is willing to place on corporate internal data — data that only ever hits the market during the one-time event of a company's collapse.