BusinessMarkets, money & strategy
Google wins $10 million auction for bankrupt Spirit Airlines' emails and chats to train AI
Google's $10 million winning bid for bankrupt Spirit Airlines' internal emails and chats is the largest known deal in a growing market where failed companies sell code and workplace data to AI labs, chiefly to build agent training environments.
From the MIT Technology Review · AI article of October 13, 2026 and cited sources: Google won the auction for Spirit Airlines' internal records with a $10 million bid. [2] [4] [5]
SOURCES[2] Axios: Google buys Spirit Airlines emails, chats, documents axios.com[4] Yahoo Finance: Google Paid $10M for Spirit Airlines' Dead… finance.yahoo.com[5] TIME: Google's $10 Million Bid for Spirit Airlines' Data time.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Google won the auction for Spirit Airlines' internal records with a $10 million bid. [2] [4] [5]
- Original source
- MIT Technology Review · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 13, 2026 article published
- Covered
- Oct 13 edition
- Evidence
- 9 sources across 9 sites
- 1MIT Technology Review: AI labs are buying data from dead…technologyreview.com
- 2Axios: Google buys Spirit Airlines emails, chats, documentsaxios.com
- 3Gizmodo: Failed companies are selling old Slack chats and…gizmodo.com
- 4Yahoo Finance: Google Paid $10M for Spirit Airlines' Dead…finance.yahoo.com
- 5TIME: Google's $10 Million Bid for Spirit Airlines' Datatime.com
- 6Technobezz: SpaceX Weighed Buying Failed Startups' Data…technobezz.com
QuietHow it’s measured
Hacker News · 1 like
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Google won the auction for Spirit Airlines' internal records with a $10 million bid. The records cover about 100 million emails and 500 million Teams messages, deidentified, with passenger data excluded. Mercor was the next-highest bidder at $7.5 million, and Micro1 submitted a $12.5 million offer after the auction closed. The sale still needs a judge's approval. [2] [4] [5]
Shutdown brokers such as SimpleClosure and Sunset now sell failed companies' Git repos, ticketing systems, Slack and email archives to frontier labs and to firms that build simulated workspaces. SimpleClosure reports more than 150 such deals, and says 47% of its deals now include a code or data sale. These figures are self-reported. Proceeds usually go first to creditors and investors. [1] [3]
Buyers mainly want the data to build reinforcement-learning environments where agents practise office tasks. They argue that real records avoid errors common in synthetic worlds. The difficulty is that realistic environments need the links between records kept intact, and those same links can make it possible to re-identify employees. [1] [5]
Spirit's privacy ombudsman recommended approving the sale on three conditions: Tonic.ai deidentifies the data, Google publicly commits not to re-identify anyone, and booking records are excluded. The flight attendant and pilot unions still object, and a hearing is scheduled for October 14. Legal experts warn that this kind of data can expose medical and family details, and some argue workers should have to give express consent. [7] [9] [1]
Documented: labs are paying for workplace records, and a bankruptcy ombudsman has proposed privacy conditions.
Read the full assessment
Implied: company chat and code archives are becoming sellable assets, so founders, employees and buyers face consent, deidentification and creditor questions. Whether real data trains better agents remains untested.
Executive brief
Google bid $10 million for the internal records of bankrupt Spirit Airlines: about 100 million emails and 500 million Teams messages. Its aim is to train AI, not to fly planes (Axios). The deal is the largest known example of a fast-growing market. Shutdown brokers such as SimpleClosure and Sunset now sell failed companies' code, Slack, email and documents to AI labs and builders of training environments. At SimpleClosure, 47% of deals now include a code or data sale (MIT Technology Review). Employee privacy and who gets paid are still unresolved. A court hearing on the Spirit sale is scheduled for October 14.
What changed and event timeline
SimpleClosure opens a data marketplace
Asset Hub lets companies that are shutting down license their code and workspace data, with personal information removed. SimpleClosure had done about 100 deals the year before, each paying $10k–$100k ().
Union objection and a late rival bid
The flight attendants' union argued that keeping links between data types intact could allow employees to be re-identified. Micro1 submitted a $12.5M bid after the auction closed ().
SpaceX's AI unit explores the same idea
Bloomberg reported that SpaceXAI teams had informally discussed buying data from struggling or defunct companies. The talks were early and no deal was confirmed (;).
Privacy ombudsman backs the Spirit deal
Lucy Thomson recommended approval on conditions: Tonic.ai deidentifies the data, Google publicly commits not to re-identify anyone, and booking records are excluded. The hearing is set for Oct 14 ().
MIT Technology Review maps the market
SimpleClosure reports more than 150 Asset Hub deals. Sunset is a growing rival. Sellers say proceeds mostly go to creditors and investors ().
Capabilities and access
- No model or version has been publicly tied to the Spirit data.
- Assets for sale: Git repos, ticketing systems and documentation, plus workspace data from Google, Notion, Figma and Slack (MIT Technology Review).
- Buyers: frontier labs, companies that build simulated workspaces, and mature tech firms training their own models.
Read the full section
- No model or version has been publicly tied to the Spirit data. Reports only say Google is expected to use it for LLM training (Axios).
- Assets for sale: Git repos, ticketing systems and documentation, plus workspace data from Google, Notion, Figma and Slack (MIT Technology Review).
- Buyers: frontier labs, companies that build simulated workspaces, and mature tech firms training their own models.
- Seller visibility: founders have little say over, or insight into, who ends up with the data.
Technical analysis for researchers and developers
- Main use: reinforcement-learning environments. Surge AI buys archives to build "RL gyms" where agents practise tasks such as IT debugging or drafting legal briefs.
- Why real data: it avoids errors common in synthetic worlds, such as dates that don't exist.
- Deidentification tension: realistic environments need the links between records kept intact, and the union argues those same links make re-identification possible (TIME).
Read the full section
- Main use: reinforcement-learning environments. Surge AI buys archives to build "RL gyms" where agents practise tasks such as IT debugging or drafting legal briefs. It values variety and depth, including disagreements and judgment calls (MIT Technology Review).
- Why real data: it avoids errors common in synthetic worlds, such as dates that don't exist. Office tasks have no pass/fail signal like code tests do, so grading them still needs human feedback (TIME).
- Deidentification tension: realistic environments need the links between records kept intact, and the union argues those same links make re-identification possible (TIME).
- Reproducibility: no one has published a comparison of training on real data versus synthetic data.
Claims and evidence
- Spirit price and data volumes: these come from court-auction reporting, so they are relatively well documented (Axios). MIT Technology Review cites 176,000 employee records.
- SimpleClosure figures are self-reported: more than 150 deals, an expected $10M+ returned to founders in 2026, and the 47% share (MIT Technology Review).
- "Real data trains better agents": this is a buyer's argument. No independent evaluation exists.
Read the full section
- Spirit price and data volumes: these come from court-auction reporting, so they are relatively well documented (Axios). Counts differ by outlet. MIT Technology Review cites 176,000 employee records. TIME adds 30M lines of code.
- SimpleClosure figures are self-reported: more than 150 deals, an expected $10M+ returned to founders in 2026, and the 47% share (MIT Technology Review). An April count put deals at about 100 (Gizmodo).
- "Real data trains better agents": this is a buyer's argument. No independent evaluation exists.
- Startup shutdowns rising 26%–56% a year: estimates from Carta and AngelList, as reported by MIT Technology Review.
Context and prior work
- Labs have hit the limits of public web data and synthetic data, and demand for real company data rose sharply in 2025 (MIT Technology Review).
- Anthropic has reportedly discussed spending up to $1B on RL training infrastructure (Gizmodo).
- Mercor and Micro1, both AI data vendors, bid for the Spirit data alongside Google (Forbes).
Read the full section
- Labs have hit the limits of public web data and synthetic data, and demand for real company data rose sharply in 2025 (MIT Technology Review).
- Anthropic has reportedly discussed spending up to $1B on RL training infrastructure (Gizmodo).
- Mercor and Micro1, both AI data vendors, bid for the Spirit data alongside Google (Forbes).
Limitations, safety and contested findings
- Privacy: attorney Greg Ewing calls this kind of data "legally radioactive." Emails and calendars can reveal medical appointments and family logistics (MIT Technology Review).
- Consent: Emory law professor Ifeoma Ajunwa argues that selling workers' data should require their express consent. No federal law bans these sales.
- Spirit dispute: the flight attendant and pilot unions still object. The ombudsman says the privacy risks have been "eliminated or mitigated" (The Next Web; Bloomberg Law).
Read the full section
- Privacy: attorney Greg Ewing calls this kind of data "legally radioactive." Emails and calendars can reveal medical appointments and family logistics (MIT Technology Review).
- Consent: Emory law professor Ifeoma Ajunwa argues that selling workers' data should require their express consent. No federal law bans these sales.
- Spirit dispute: the flight attendant and pilot unions still object. The ombudsman says the privacy risks have been "eliminated or mitigated" (The Next Web; Bloomberg Law).
- Pricing: buyers quote very different numbers because valuations rest on untested "research theses" about how many useful tasks a dataset yields.
Business and practitioner implications
- Founders: a wind-down can now recover cash. Proceeds usually go to senior creditors first (MIT Technology Review).
- Employees and HR: workplace chat should be treated as a likely future asset sale.
- Buyers: budget for deidentification certification (the Spirit deal uses Tonic.ai), non-re-identification commitments and court review in bankruptcy sales (The Next Web).
Read the full section
- Founders: a wind-down can now recover cash. One founder netted about $25k for code, with roughly $15k pending for workspace data. Proceeds usually go to senior creditors first (MIT Technology Review).
- Employees and HR: workplace chat should be treated as a likely future asset sale. Expect pressure for consent clauses.
- Buyers: budget for deidentification certification (the Spirit deal uses Tonic.ai), non-re-identification commitments and court review in bankruptcy sales (The Next Web).
- Watch: the Oct 14 Spirit ruling is likely to shape how future data sales are structured.