Google won a $10 million bankruptcy auction for Spirit Airlines’ internal business data, a package that court filings put at roughly 100 million emails and 500 million Microsoft Teams chats plus documents, code and operational records. A federal judge in New York is set to hear the sale today.
The deal, still pending approval, hands the AI giant de-identified workplace exhaust from a carrier that shut down in May after its second Chapter 11. Google says the material can improve products and models. The runner-up was AI data firm Mercor at $7.5 million.
What the $10 million package actually contains
According to the August 14 auction results notice and related filings, Google is buying Spirit’s de-identified internal archive. The core counts that have circulated from the papers are about 100 million emails and 500 million Microsoft Teams chats.
Beyond the messages sit calendars, spreadsheets, marketing materials, HR files, project documents, financial databases, audits, presentations, IT tickets, revenue-management records, pricing models, booking curves, refund histories, aircraft operations data, and roughly 30 million lines of production code across hundreds of repositories. Some records reach back to 1986; operational transaction rows go to at least 2008. Filings also reference more than 175,000 employee records.
Taken together, the message volume and the supporting systems form a single connected trail. A pricing model in a spreadsheet can be read against the email thread that approved it and the Teams chat that handled the exception when the curve broke. That linkage is what the filings describe, and it is what the bidders paid for.
- Included: internal communications, custom software, finance and ops databases, maintenance and crew data, code repositories.
- Excluded: passenger profiles, Free Spirit loyalty records, customer personal information, credit-card data, and any material that qualifies as personal data under privacy laws.
- Process: a third-party de-identification agent, chosen or approved by the buyer and paid by Google, scrubs the set before delivery. Google agrees not to re-identify.
A Google spokesperson told Axios and Business Insider the company is acquiring “part of an enterprise dataset from Spirit Airlines, which can be helpful in improving our products and AI models” and “will not receive any personal information from this dataset.” Any data received “will be rigorously scrubbed of any personally identifiable information by a third party before receipt.”
How the bidding climbed from $5 million to $10 million
The virtual auction ran on August 14, 2026, under U.S. Bankruptcy Judge Sean H. Lane in the Southern District of New York. Spirit’s advisers at PJT Partners and Davis Polk moderated. Three qualified bidders took part.
The Dylan Friesner declaration on the bidding, filed by the PJT vice president who advised Spirit, lays out the rounds. Google opened at $5 million cash with a third-party de-identification process it would fund. Mercor.io Corporation overbid at $5.2 million on the same form, or $7 million if it could de-identify in-house with its own tools. Spirit’s team preferred the third-party route for privacy certainty and treated the $5.2 million bid as the higher one that round.
| Bidder | Key offer | De-id approach | Outcome |
|---|---|---|---|
| Google LLC | $5M opening, then $10M final | Third-party agent, buyer pays | Successful bidder |
| Mercor.io | $5.2M / $7M option, then $7.5M alternate ($10M if in-house) | Preferred own tools for higher price | Alternate bidder at $7.5M |
- Open: Google bids $5 million cash and funds a third-party scrub.
- Overbid: Mercor offers $5.2 million on the same scrub terms, or $7 million with in-house tools.
- Preference: Spirit’s advisers score the third-party path higher for privacy certainty and carry $5.2 million as the lead that round.
- Close: After about two and a half hours of rounds, Google reaches $10 million; Mercor’s $7.5 million bid (or $10 million in-house, which Spirit declined) becomes the alternate.
After multiple rounds lasting about two and a half hours, Google landed at $10 million. Mercor’s $7.5 million offer (or the $10 million in-house alternative that Spirit declined) became the backup. Friesner wrote that the jump from $5 million to $10 million and the $2.5 million gap to the alternate showed the value had been tested. Non-economic factors, especially a documented scrub process that kept consumer data out, shaped the outcome as much as cash.
Why a dead airline’s email threads suddenly have a price
Public internet text has been scraped nearly dry for large-language-model pretraining. What remains scarce is the connected record of how real organizations actually work: the email arguments that produce a pricing change, the Teams threads that resolve a crew shortage, the ticket-to-code trail of a software fix, the finance spreadsheets that close a quarter.
That material never lived on the open web. It sits inside companies. Agentic systems that are supposed to use software, follow multi-step processes and handle exceptions need exactly those traces. Synthetic data can approximate patterns; it does not reproduce the messy, time-stamped decisions that led to real outcomes.
Mercor’s own framing, quoted in coverage of the auction, captures the buyer view: “Frontier AI labs can’t build enterprise-grade models from public data alone. The next leap in AI capability requires operational data: how teams communicate, close deals, ship software, and run finance. That data lives inside your company.” A Mercor spokesperson added that companies “are sitting on decades of records that show how real work gets done, and that data is now some of the most valuable material for training and evaluating AI.”
Spirit’s archive is one of the first large, public price signals for that category. At $10 million for roughly 600 million messages plus code and ops data, the per-record cost looks tiny. The usable-example cost after cleaning and structuring is far higher. The auction still confirmed demand strong enough to double the opening bid.
The same scarcity logic explains why a failed low-cost carrier drew two serious AI-side bidders rather than a single stalking-horse offer. Once public text is exhausted, the next training edge is process history: how a refund was denied, how a crew swap was approved, how a pricing model was patched in code. Spirit’s estate happened to put that history on the block in one lot.
The scrub that kept passenger data off the table
Spirit’s estate ran a parallel marketing track for its customer list; that asset is not part of this sale. The de-identified package expressly excludes information “that relates to, describes, or is reasonably capable of being associated with a consumer.” Privileged attorney-client material is also carved out.
Google designates or accepts the de-identification agent and bears the full cost. Delivery happens only after scrubbing. The bill of sale language makes clear the buyer receives Deidentified Data and commits not to reverse the process.
We are buying the company’s internal data and custom software, but we are not buying their customer or credit card information.
That is the Google spokesperson’s line to Business Insider. The same message appears in the Axios statement. Privacy advocates still note that workplace communications can carry residual signals even after names and IDs are stripped, and that the buyer’s control of the scrub agent is a point of leverage. An earlier Google personal data fine case shows how courts in other jurisdictions treat the company’s handling of sensitive records. Here the bankruptcy court and the third-party process are the current safeguards.
The split is deliberate. Consumer files stay on a separate track. Workplace exhaust moves only after a third party strips identifiers. The buyer’s public statements and the sale papers both turn on that line.
What the sale does to the market for corporate knowledge
The second-order effect is larger than one airline. Once a $10 million bid clears for a defunct carrier’s email and code, every restructuring adviser and every distressed investor has a new line item to model. Unstructured operational data, the digital exhaust that never appeared on a balance sheet, now has observable buyers among the largest AI labs and the specialized data firms that feed them.
Live companies face a choice. They can wait until failure and let creditors auction the archive, or they can license curated, rights-cleared slices while the firm still operates and capture the value themselves. Continuous licensing produces cleaner, current data with upstream consent. Bankruptcy sales are sporadic, raw and backward-looking by definition. Both channels now exist.
- Bankruptcy path: creditors monetize a raw, historical archive after failure; price is set in open bidding.
- Live licensing path: the operating firm sells or licenses curated slices with consent already upstream; data is current rather than frozen at shutdown.
Similar pressure is already visible in other asset classes. Banks have explored banks converting old data sites into housing, and large tech grant changes have produced sudden data losses for nonprofits, as in the Microsoft grant cut that erased nonprofit data. The Spirit deal simply attaches a hard number to the AI side of the ledger.
On X, the reaction mixed dark humor with recognition. Posts described “cyber immortality” for a failed firm’s internal life inside frontier models. Others treated the package as pure inventory: emails and chats now sit on the same shelf as GPUs and power contracts. The sharpest observation is structural. Labs that once scraped books and websites are now bidding in bankruptcy court for the ordinary digital residue of running a business.
Why the Third Party Route Carried the Room
Cash alone did not decide the auction. Mercor’s papers offered a path to match Google’s final figure if Spirit would accept in-house de-identification. Spirit’s advisers declined that branch and kept the third-party requirement in place.
The preference was about certainty. A scrub agent chosen or approved for the sale, paid by the buyer, and finished before delivery gave the estate a cleaner record on consumer data staying out. Google’s opening bid already funded that structure. Mercor’s higher in-house alternative asked the estate to trust the buyer’s own tools instead.
Friesner’s declaration treated the documented scrub as a non-economic factor on par with price. The $2.5 million gap between Google’s winning $10 million and Mercor’s $7.5 million alternate is the visible cost of that choice. Privacy process, not only headline dollars, sorted the winner from the backup.
How the Price Signal Travels Beyond One Carrier
At $10 million against roughly 600 million messages plus code and ops systems, the headline unit cost looks small. After de-identification, cleaning, and structuring, each usable training or evaluation example costs far more. The auction still proved that two qualified bidders would fight into the eight figures for one airline’s internal trail.
That proof changes how other estates and advisers model recoveries. Email, chat, ticket, and repository archives stop reading as storage liabilities and start reading as assets with named buyers. Google and Mercor both showed up for the same category of material: how teams communicate, ship software, and run finance when the work is real.
The hearing will test only whether this sale closes. The demand signal does not wait on the order. Workplace process data now has a public clearing price, and the first large airline package set it at ten million dollars.
Spirit’s planes, people and the rest of the wind-down
While Google and Mercor fought over the data, the physical fleet headed for desert storage in Arizona, where dry heat slows corrosion. Plane-leasing executives have described the routine: maintain the aircraft, wait for the next operator or part-out. Employees were pointed toward networking platforms such as Kitty Hawk to find new roles in a seniority-heavy industry. Older workers face the steepest restart.
The data sale is one slice of a larger disposition. Creditors get $10 million if the court signs off. Google gets three decades of how a low-cost carrier actually scheduled crews, priced seats, handled refunds and wrote software. Mercor, if the primary deal collapses, stands ready at $7.5 million with its own plans to feed the same kind of operational history into enterprise models.
The hearing is set for 11 a.m. Eastern on August 19. Until the order enters, the sale is not final. What is already final is the price signal. Real workplace process data now has a public auction market, and the first large airline archive just cleared at ten million dollars.
