Google–Spirit: The First Public Price on a Company's Memory

Ten million dollars, an auction that doubled, and a sale nobody has approved

On 14 August 2026 the debtors in Spirit Airlines' chapter 11 filed the results of an auction almost nobody was watching. The asset was not aircraft or gates. It was the airline's deidentified data, and the successful bidder, at $10,000,000, was Google LLC, with Mercor.io alternate at $7.5 million ^[1]. Three days later the estate's banker, Dylan Friesner of PJT Partners, swore a declaration describing how that number was reached: three qualified bids, a $5 million opening, an auction of roughly two and a half hours ^[2]. On the evening of 18 August the flight attendants' union filed a limited objection ^[3]; three hours earlier the debtors had moved the sale hearing from 19 August to 11:00 a.m. on 9 September 2026 ^[4]. Nothing is approved, and by the union's account no data has been transferred ^[5]. What has happened already is the part that matters: a company's operational memory went to auction and cleared at a price.

• $10M Successful bid, Google LLC, for Spirit's deidentified data; Mercor.io alternate at $7.5m ^[1] Auction result filed 14 Aug 2026 · sale contested and not approved; hearing 9 Sep 2026 MEASURED court filing
• $5M → $10M Opening bid to successful bid across an auction of about two and a half hours, three qualified bidders ^[2] Banker's declaration filed 17 Aug 2026 · sworn, in support of approval MEASURED
• 175,658 Employee records, back to August 1986, marked Included — alongside ~100m emails and ~500m Teams items ^[1] Assets Schedule, filed 14 Aug 2026 · same figures restated by the objecting union MEASURED

What is on the block is narrower than the coverage suggests, and stranger. Marked Included: roughly 100 million emails, 500 million Teams items, 7.25 billion competitor flight observations, 7.51 billion revenue-system transactions back to May 2008, 175,658 employee records, 3.4 million payroll records, 516 source-code repositories and redlined draft-to-final contract histories. Marked Not Included: 97.5 million customer profiles and 50.2 million loyalty records ^[1]. Several outlets called this a sale of Spirit's customer data; it is close to the opposite: the estate is not selling what Spirit knew about its customers, but what Spirit knew about itself.

Five days from auction result to adjourned hearing

1. 14 Aug 2026

   Notice of auction results filed: Google LLC successful at $10,000,000, Mercor.io alternate at $7,500,000, with the sale agreement and Assets Schedule attached ^[1].

   Filed · sale not approvedPriceMEASURED

2. 17 Aug 2026

   The objection deadline passes at 4:00 p.m. ^[4]; the banker's declaration describing the auction is filed the same day ^[2].

   Filed · deadline expiredProcessMEASURED

3. 18 Aug 2026 · 17:04

   Debtors' counsel files a notice adjourning the data-sale hearing to 11:00 a.m. on 9 September 2026, stating no reason ^[4].

   Notice by the debtors — not a court orderDelayMEASURED

4. 18 Aug 2026 · 20:25

   The Association of Flight Attendants-CWA files a limited objection, three hours and twenty-one minutes after the adjournment notice ^[3].

   Filed · late, and narrow by its own termsObjectionMEASURED

5. 9 Sep 2026

   Hearing on the sale, 11:00 a.m., before Hon. Sean H. Lane, U.S. Bankruptcy Court, Southern District of New York ^[4].

   Scheduled · outcome unknownNext markerFORECAST

The order of those two 18 August filings is worth holding onto, because the coverage inverts it. Wire copy framed the delay as coming amid the objection, and downstream write-ups hardened that into cause ^[6] INFERENCE. The adjournment was docketed first, without explanation; the objection arrived afterwards. The union objected and the hearing moved. That the objection moved the hearing is supported by nothing on the docket.

The clause that makes the corpus valuable is the clause the union is objecting to

The sale agreement requires the data to be deidentified to consumer-privacy standards — the California statute's test whether or not it applies — and then adds six words that do all the work: the deidentification must be performed “while preserving referential integrity across the data set”. The buyer names or vetoes the deidentification agent, certification runs to the buyer's satisfaction, and the buyer pays that bill separately, with costs that “shall not reduce the amount of the Purchase Price” ^[1]. So $10 million is a floor, not a gross.

A corpus is worth buying because its records still know each other. Deidentification is safe when they forget. This agreement asks for both and pays for one.

That is the union's objection, and it is narrower and better argued than the headlines. Deidentification, it says, addresses whether a record can be traced to a named individual; “it does not address whether the contents of the record are confidential”. The privacy architecture borrowed here was built for customers, while the payload conveyed is “disproportionately employee-facing” — and because § 3(c) preserves the linkages by design, the agreement “not only omits an employee-confidentiality screen, but it functionally forecloses one”. The union makes no technical re-identification claim and says it need not: the crew base population in the schedule is 4,600, described across linked systems spanning more than a decade, and the flight attendants are expressly denied third-party beneficiary status in the contract that moves their record ^[3].

Note what the union does not ask for. It does not seek to unwind the auction or stop the estate monetising data, it concedes the privacy-ombudsman trigger is not met here, and it filed a day late ^[3]. Its procedural hand is weaker than the coverage implies. Sara Nelson's public line was that the plan is “outrageous” ^[7]; the filing asks only for carve-outs and a review protocol, noting that the buyer already bears the deidentification cost with no reduction in price ^[3]. That is why it matters more than an angrier filing would: this is not an attempt to close the market but the first absent party walking into it and putting a price on the friction. When 23andMe's data was sold in bankruptcy in 2025, the people in that database were consumers and a privacy ombudsman examined the sale ^[8]; Spirit's employees get no equivalent office. My judgement is that every future whole-corpus sale inherits this cost layer — a confidentiality screen, a review protocol, carve-outs, someone to negotiate with — whatever Judge Lane rules on 9 September.

The auction had already priced part of that layer, twice. Mercor's first overbid was not one bid but a pair: $5.2 million to take Google's form bill of sale, under which a third party deidentifies anything containing personal information at the buyer's cost, or $7 million if the terms let Mercor complete the deidentification with its own tools. The estate took the lower cash — the debtors and their advisers “determined that $5.2 million was the highest bid for that round despite the lower price”. The split returned at the close: Mercor also offered $10 million, conditional on an in-house rather than third-party process, and the estate designated its $7.5 million bid as the alternate instead ^[2]. Control of the scrubbing method was priced twice, at $1.8 million and then $2.5 million, and a wind-down estate twice let the larger number go to keep that control in third-party hands. On this record it is not a term of the sale; it is a second thing being sold.

The appetite is the story; the deal is only the receipt

Why a frontier lab would want thirty-four years of an airline's internal record is legible from the data-supply market. In June 2025 Meta paid $14.3 billion for 49% of Scale AI and took its founder in-house ^[9]. Within days, Scale's largest customer — Google, reported to have planned some $200 million of spending with it that year — was reported to be ending the relationship, and OpenAI was phasing it out ^[10]. That is supplier neutrality collapsing, and it is the strongest argument I know for owning a corpus outright rather than renting a vendor's.

The market that answers is priced. Turing's Project Lazarus, live since December 2025, offers to buy private repositories, design documents, testing logs and internal wikis for “a single payout of up to $1M”, addressed to “operating and winding-down companies” COMPANY CLAIM ^[11]. Distress is the cheapest entry point, not a precondition. And the design the union objects to is sold as a feature elsewhere in the stack: the deidentification vendor Integral markets “entity-preserving” methods that keep “the longitudinal record intact” COMPANY CLAIM ^[12]. The objection and the product description describe the same thing from opposite ends.

The honest limit: every whole-corpus transaction I can verify is a wind-down, and that this appetite reaches going concerns rests on one vendor's marketing page and no observed deal. The priced market for corporate memory is, so far, a salvage market — and it is not finished with this carcass. The estate says it is separately marketing Spirit's customer list, with approval to be sought later ^[2]. Two priced data transactions out of one bankruptcy is the shape of a channel rather than an accident.

What would change my reading

Three dated tests, so this note can be wrong in public. None resolves on whether the sale is approved: if it depended on Judge Lane, this would be a court report, not a reading of a channel. One: if by 30 June 2027 no estate other than Spirit's has run a court-supervised or publicly reported sale of a whole-company operational corpus — communications, code and business-systems records as one lot — drawing at least two qualified bidders at $1m or more, the salvage channel was a one-off and I over-read it; Spirit's own customer-list sale does not count. Two: if by 31 March 2027 Turing's Lazarus page no longer publishes a payment figure and no comparable priced offer is live elsewhere, corporate memory has returned to unpriced, deal-by-deal negotiation and the appetite reading weakens; a higher ceiling or new entrants confirm rather than falsify. Three: if by 31 December 2026 any buyer, here or elsewhere, accepts a protocol that severs cross-system linkage without a reduction in consideration, linkage was not what was bought and the mechanism above is wrong. The standing counter-evidence is the auction itself: bidders repriced according to whose tools did the deidentification ^[2].

Method, evidence classes and disclosure. Evidence window: 12 June 2025 to 19 August 2026, every figure re-checked against its source on 19 August 2026. The four court documents behind this note were read as complete primary filings, free from the claims agent's public portal; no paywalled reporting is cited, and contract language is quoted from the filed agreement rather than press descriptions of it. Evidence classes: MEASURED a filed document's own count or text; COMPANY CLAIM a figure a party publishes about itself; FORECAST a scheduled future event; INFERENCE my judgement, inputs shown. Status: the $10 million is an auction result filed with the court — press-reported first, verified here in the filing — not an announced or closed transaction; the sale is contested and unapproved. Incentives: PJT Partners is the estate's banker and its declaration supports approval; the union is an adverse party; Turing and Integral sell the products cited. Disclosure: written in a personal capacity from public sources only. The author works within the Türkiye venture ecosystem; to avoid conflicts of interest, no Türkiye-based fund or startup is named or evaluated. The author holds no positions in, and no client relationship with, any company named.

This is a note, not a thesis

Notes are conjunctural: same sourcing, same badges, same conflict-of-interest discipline, a shorter spine, no scorecard. This one sits under no standing thesis. The mechanism it describes — corporate memory becoming a priced, transferable asset, and the absent parties inside that memory learning to show up — is one this series is watching, and a thesis position on it is forming rather than settled. Until then the argument lives here, graded by the tests above.

The Agents & Margins briefing is launching — until then, follow on LinkedIn

Sources

1. 1.ECF 1463 — Notice of Auction Results and Scheduled Hearing for the Deidentified Data, filed 14 August 2026 in In re Spirit Aviation Holdings, Inc., et al., No. 25-11897 (SHL), U.S. Bankruptcy Court, S.D.N.Y. Successful bid $10,000,000 (Google LLC), alternate $7,500,000 (Mercor.io Corporation); Sale Agreement at Exhibit A (§ 1(a)(ii)–(iii) deidentification agent and costs, § 3(c) referential integrity, § 12 third-party beneficiaries) and the Assets Schedule. Filings in this case are free of charge from the claims agent at dm.epiq11.com/case/spirit.
2. 2.ECF 1470 — Declaration of Dylan Friesner (PJT Partners) in support of the sale, filed 17 August 2026: three qualified bids, $5m opening bid, auction of roughly two and a half hours, ¶ 13 on Mercor's paired first overbid ($5.2m on Google's form bill of sale, $7m with its own deidentification tools) and the estate taking the lower price, ¶ 15 on Mercor's $10m offer conditional on an in-house process against the $7.5m designated as the Alternate Bid, and note 4 on the separate customer-list marketing process.
3. 3.ECF 1489 — Limited Objection of the Association of Flight Attendants-CWA, AFL-CIO to the Proposed Sale of the Deidentified Data, filed 18 August 2026, 20:25, counsel Charles M. Rubio. All quoted union language is from this filing.
4. 4.ECF 1488 — Notice of Adjournment of Hearing on the Sale of the Deidentified Data, filed 18 August 2026, 17:04, by Davis Polk & Wardwell LLP for the debtors: objection deadline was 4:00 p.m. on 17 August 2026; hearing adjourned to 11:00 a.m. on 9 September 2026. A notice by the debtors, not a signed order, and it states no reason.
5. 5.AFA-CWA — Spirit Bankruptcy: Objection to Sale of Your Data, 18 August 2026: the objecting party's own notice to members, stating that the sale has been announced but not approved and that no data has been transferred.
6. 6.Reuters — US court delays hearing on Google's purchase of Spirit Airlines data as union objects, 19 August 2026. The temporal framing here is the origin of the widely repeated causal claim; the docket does not support it.
7. 7.Forbes — Suzanne Rowan Kelleher, 18 August 2026. Cited only for AFA president Sara Nelson's public reaction and the union's stated representation of more than 5,500 Spirit flight attendants; its description of the data being sold conflicts with the Assets Schedule and is not used.
8. 8.NPR — 23andMe's sale of consumer genetic data approved in bankruptcy, 30 June 2025; the sale ran with a court-appointed consumer privacy ombudsman examining it. biopharmadive.com
9. 9.CNBC — Scale AI founder Alexandr Wang exits for Meta as part of a $14.3bn deal for 49% of the company, 12 June 2025.
10. 10.CNBC — Google, Scale AI's largest customer, plans split after Meta deal, 14 June 2025, reporting roughly $200m of planned 2025 spending; anonymously sourced. techcrunch.com, 18 June 2025
11. 11.Turing — Project Lazarus vendor page, read 19 August 2026: “A single payout of up to $1M”; “No retainer, no recurring commitment”; “For operating and winding-down companies”. Vendor marketing, cited for the terms of the offer. Launch, 23 December 2025
12. 12.Integral — data privacy programs for AI, 4 May 2026: “entity-preserving” methods that keep “the longitudinal record intact” while carrying a defensible re-identification posture. Vendor marketing, cited as the industry's own product definition.