Who owns a data estate when the company dies?

01

What was sold, and for how much

The lot is large. According to the court notice as reported by Bloomberg Law, it includes roughly 100 million emails, 500 million Microsoft Teams messages, 7.2 billion records of competitors' flight prices, 7.5 billion passenger transaction records going back to 2008, more than 175,000 employee records dating to 1986, and about 30 million lines of code ¹. The Register's reading of the same court document adds 17 million OneDrive files, 20.5 million SharePoint items, operational data covering 763,000 flights, 5 million crew pairings and 1.2 million fuel slips ². The notice also listed some 30 million recorded customer service calls and 15 million chat records ², though The Next Web reports that the agreement with Google excludes call recordings and customer chat sessions ⁸. Passenger profiles and loyalty records are excluded ¹.

Google's $10 million beat a $7.5 million bid from Mercor, an AI recruiting and training-data company and the designated backup bidder ¹. On 4 September, after the auction had closed, a second training-data start-up, micro1, filed a $12.5 million offer with no financing condition and a promise to bear all de-identification costs ⁷⁸. The court will consider that late bid at the 9 September hearing ⁷.

Spirit listed about $8.1 billion of debt at filing ¹. Nearly two decades of transaction history, the full operational record of a carrier and the personnel history of a workforce of around 17,000 ⁹ went for $10 million.

The sale is conditioned on de-identification. The data is to be "rigorously scrubbed of any personally identifiable information by a third party before receipt" ¹. Google says it "will not receive any personal information from this dataset" and that the purchase "can be helpful in improving our products and AI models" ⁴.

02

Why AI labs want operational exhaust

The bidders are not buying an airline. Two of the three are training-data companies. TIME frames it this way: the public web was enough to teach models to code, because code has a binary signal (it runs or it does not). Office work has no such signal. To train agents that carry out white-collar tasks, labs need records of how people made decisions inside a real organisation, including the failures ³. Nick Heiner of Surge AI told TIME that synthetic data helps, but that real operational records are worth more because the data "encodes many useful things about how humans carry out their jobs" ³.

That is what 500 million Teams messages are. They are a trace of who escalated what to whom, how a fuel cost got approved, how a crew pairing was rebuilt when a flight was cancelled. The 7.2 billion competitor price records are a revenue management team watching a market and reacting. The 30 million lines of code are the systems those decisions ran through. For a lab building agents, this is the signal the open web cannot supply.

This market will recur every time an organisation with a large operational record fails.

03

The consent and provenance gap

None of the people inside the data were parties to the sale. The Association of Flight Attendants-CWA, which represents more than 5,500 Spirit flight attendants, filed a limited objection ⁵. It puts the problem in one sentence: "The privacy architecture of this transaction is consumer-facing; its payload is disproportionately employee-facing" ⁴. Customer profiles were carved out. Employee records dating to 1986 were not ¹⁴.

The sale agreement requires that de-identification preserve "referential integrity across the data set" ⁴. Referential integrity is what makes a database usable: the link between a payroll row, an HR file, an email thread and a Teams channel survives after the names are stripped. It is also what makes re-identification feasible. The union argued that because links across datasets must be preserved, "information about individuals or small groups could potentially be reconstructed" ⁶. For a population of 5,500 cabin crew with known bases, seniority tiers and training dates, the set of candidates for any linked record is small.

This is the "de-identified but linked" problem. The buyer wants the links, because they carry the workflow signal that justifies the purchase. The data subjects want the links cut, because the links are the identity. A third-party scrubbing process cannot satisfy both. The union asked the court to strip flight attendant training records, time cards and payroll, to exclude Microsoft 365 content containing flight attendant information, and to require notice when third parties access the data ⁵. The late bid from micro1 offers an ombudsman chosen by Spirit's advisers, US-only storage, and exclusion of disciplinary and collective bargaining material ⁸. The bidders are now negotiating privacy terms that the original process never put to the data subjects.

Nobody is litigating the provenance gap. Once the scrubbed corpus leaves the estate, there is no public record of which files went, in what version, at what time. The receipt is private.

04

The sovereignty angle

Everything in this case was domestic: a Florida carrier, a New York court, a California buyer. For almost every other country the same event would be cross-border.

Transpose it. A national carrier in a mid-sized economy fails. Its data estate holds the price history of every route into and out of the country, the crew and maintenance records of a national workforce, the fuel and slot logistics of the state's main gateway, and decades of correspondence with regulators. The estate is administered by a court whose duty is to creditors. The highest bidder is a foreign hyperscaler that wants the corpus to train agents. The price is a rounding error against the debt. The data subjects have no standing. Neither does the state, unless it happens to be a creditor.

A bankruptcy court is built to maximise recovery for creditors, not to ask where a country's operational memory should live. That is its mandate, and no criticism of Judge Lane. Whether a nation's institutions should be able to keep their data estates onshore, or at least keep a record of where they went, is not a question the insolvency process is built to ask.

Governments already treat airlines, ports, grids and payment systems as critical infrastructure. Their data exhaust is now training material for the systems meant to run the next generation of that infrastructure. A state, a sovereign fund or a national statistics office has reason to ask where that material goes and on what terms, before the auction.

05

What this means for verified records

Owner-signed provenance would not have stopped this sale. Nothing about anchoring a hash of a record to a public chain gives a data subject standing in a bankruptcy court.

What it does is narrower. If an organisation anchors what it publishes, signed from its own wallet, there is a permanent public record of what existed, when, and under whose authority, independent of the organisation's survival. A buyer, a court, a union or a regulator can check a scrubbed corpus against that record. The estate can die; the receipt does not. That is the modest thing LocalTruth is built to do for the official records of Mongolia's institutions, on Injective. It fixes provenance before anyone has to argue about it.