In August 2026, Google agreed to pay $10 million at a bankruptcy auction for 40 years of Spirit Airlines’ internal emails, chat logs, source code, and employee payrolls. This is a beginner’s visual guide to how corporate death turns your daily work into artificial intelligence training fuel.
Before examining the digital files, consider how the bankruptcy market valued physical asphalt compared to human history.
When people think about corporate data, they imagine proprietary trade secrets worth millions per file. But when you divide $10 million across 647 million items, the illusion evaporates.
At a penny and a half each, four decades of human labor, aircraft maintenance panics, and executive negotiations become disposable training tokens.
Click an everyday object to see how many Spirit Airlines files, emails, or code commits cost the exact same amount.
100 million emails and 500 million Microsoft Teams messages cannot be understood as text on a page. They are shapes: stacked, nested, and proportional areas of corporate behavior.
Microsoft Teams Messages & Collaboration Logs
Emails across ~80,000 Corporate Accounts
SharePoint Items (20,577,677) + OneDrive Files (17,082,644)
Attendant & pilot flight schedules
>763K flight logs + >1.2M fuel slips
787,452 parts + 667,563 tickets
516 repos (30M lines) + 175,658 HR
If each of Spirit’s 647M digital messages & documents were printed on a single sheet of standard 20lb copy paper (0.1 mm thick):
This staggering vertical tower of human intent, crisis dispatch, and engineering telemetry was purchased for the cost of approximately two suburban houses in Palo Alto.
Here is what a single 1.5¢ item in the auction looks like when parsed as an AI training token:
{
"record_id": "SERVICENOW_INC_00667563",
"source_system": "ServiceNow Infrastructure IT",
"purchase_price_usd": 0.0154,
"timestamp_utc": "2024-03-12T04:19:22Z",
"category": "AIRCRAFT_GROUND_DISPATCH_ALERT",
"base_airport": "FLL",
"aircraft_tail_number": "N682NK",
"de_identification_standard": "CCPA_HASHED",
"reporter_hash": "UUID_USR_894f2b109e",
"summary": "APU generator feeder cable clamp chafing observed during A-check preflight.",
"tokens_estimate": 348,
"referential_integrity_pointers": {
"flight_telemetry_ref": "FLIGHT_NK_0402_20240312",
"parts_purchase_ref": "PO_PARTS_787452",
"teams_chat_thread_ref": "TEAMS_THREAD_f94a028",
"payroll_overtime_ref": "WORKDAY_PAYROLL_2024_06"
}
}The bidding sequence was not just a bidding war between Google and an AI startup. It revealed the exact dollar premium an AI company places on raw, unredacted human records.
When Mercor bid $5.2 million for scrubbed data, but escalated to $7.0 million for raw, unscrubbed access, it explicitly quantified the commercial penalty of privacy. Raw data carries unredacted syntax, authentic human conflict, real pilot names, and complete operational causality. Redaction damages data fidelity — and the market priced that damage at nearly $2 million.
Even the basic question of what changed hands is not settled in public reporting. Rather than smoothing over this ambiguity, look at the two conflicting readings side by side.
Decades of passenger complaint audio, pilot-dispatch voice comms, and phone bookings.
Direct web and app support conversations detailing cancellations, baggage, and disputes.
Active promotional subscriber lists, opening telemetry, and interaction metrics.
Device MAC identifiers, passenger device session logs, and onboard connectivity sales.
Wheelchair, medical oxygen, and passenger accommodation historical logs.
Headlines reassured travelers that passenger data was excluded from Google’s lot. But reading the bankruptcy filing reveals the reality: passenger records were not saved — they were merely held back for another auction.
When the bankruptcy filing states that 97.5 million passenger profiles, 50.2 million Free Spirit accounts, and 740,000 credit card holder records are excluded, a beginner assumes those people's privacy was protected by federal law.
It was not. The debtor estate expressly retained the legal right to sell the customer list separately, including complete passenger annual spending histories. Passenger data was unbundled so creditors could auction it off to credit card issuers, rival airlines, or hotel chains in a separate proceeding.
Why "anonymous" and "anonymous but connected across decades" are two completely opposite things.
If you completely scramble every name and delete all cross-file IDs, an AI model cannot learn how a company works. An isolated email saying "Check the turbine gasket" is meaningless noise to an LLM.
To train an AI model to operate an enterprise, the model must see the chain of consequence:
Therefore, the court order required Referential Integrity: replacing the real name with an anonymous hash (`USER_894f`), but keeping that exact same hash linked across all 647 million files.
Toggle de-identification and watch how threads across disparate databases make a person unmistakable.
Passengers had CCPA fine print. But 17,000 employees found that their performance write-ups, disciplinary histories, and union chats were sold as scrap metal.
When the sale motion was filed on August 14, 2026, the Association of Flight Attendants-CWA (AFA) realized something alarming: the privacy standards in the contract explicitly protected consumers (passengers), but provided zero protections for workers.
Inside the 647 million items are:
Because airline crews operate from known domicile bases (Orlando, Fort Lauderdale, Las Vegas, Detroit) on fixed FAA schedules, an "anonymized" crew pairing with date, time, and aircraft type is mathematically unique. Cross-referencing public flight records with payroll dates trivially re-identifies individual flight attendants by name.
Spirit’s records were liquidated under Section 363 of the U.S. Bankruptcy Code.
The code was drafted in 1978. In 1978, the assets of a bankrupt company were hydraulic presses, industrial shelving, office typewriters, and inventory in a warehouse.
When a press is sold, the metal doesn't have an opinion. But in 2026, an enterprise asset is 500 million live human chats and emails. The law still treats both as identical: property of the debtor estate to be converted into cash for senior debt holders.
The public internet has been scraped to exhaustion. What frontier AI models lack is not more words — it is the messy, painful process of how businesses actually work.
"Companies are sitting on decades of records that show how real work gets done."
Public text (Wikipedia, Reddit, news) teaches an AI model how to write fluent essays. But it does not teach a model how to resolve a mechanical delay when an aircraft is stuck on tarmac in Detroit with 180 angry passengers and a missing fuel slip.
Spirit’s corpus contains the full lifecycle of corporate decisions: from the initial panic on Microsoft Teams, to the ServiceNow dispatch ticket, to the budget spreadsheet adjustment, to the flight log sign-off. Models need cause-and-effect data.
Every employee and customer agreed to Spirit’s privacy policy. But Spirit Airlines ceased to exist on May 2, 2026. A privacy promise does not outlive the legal entity that made it. Once dead, commitments dissolve into auctionable creditor assets.
Spirit is not unique. Every hospital, bank, logistics provider, and software startup generates an identical archive of Slack messages, Jira tickets, and emails. In the AI era, bankruptcy turns your daily labor into a training set for your replacement.
Enter your company's headcount to see what Google or Mercor would pay for your entire internal memory at Spirit's 1.5¢ bankruptcy rate.
All figures, case filings, and quotations on this page are drawn directly from public bankruptcy filings and authoritative reporting.
Author: Simon Sharwood, APAC Editor • Published: 18 August 2026
U.S. Bankruptcy Court for the Southern District of New York (SDNY) • Judge Sean H. Lane presiding
Counsel for AFA-CWA representing ~17,000 active and displaced flight crews
Purchase price: $58,500,000 for 22 daily flight operations
A standalone visual explainer produced from first principles without external backend dependencies.
Dedicated to the 17,000 workers whose decades of labor were auctioned for 1.5 cents a record.