Google Paid $10 Million for a Dead Airline's Inbox

Spirit Airlines is being sold off in bankruptcy court piece by piece. Aircraft, gates, slots. This month the auction reached the part nobody expected to be worth anything: the company's inbox.

Google won it with a $10 million bid. The package includes roughly 100 million internal emails, 500 million Microsoft Teams messages, 17 million OneDrive files, 20 million SharePoint files, and 516 source code repositories. HR records, financial databases, audits, pricing history, calendars, presentations. Two decades of how a 17,000 person company actually ran, day to day.

Google says the data will be scrubbed of personal information by a third party before handoff, that no passenger or loyalty data is included, and that it will be used for "product development and AI model training." The runner-up bidder was Mercor, an AI hiring startup, at $7.5 million. A bankruptcy judge still has to approve the sale. The hearing was pushed to September 9 after the flight attendants' union objected.

Why a search company wants an airline's email

Ten million dollars doesn't register on Google's balance sheet. The interesting question is why it's paying anything at all for a bankrupt carrier's Teams history.

The answer is that the labs are running low on the thing that made large language models work in the first place: text written by humans. Epoch AI estimated the total stock of usable public human-written text at around 300 trillion tokens and projected that frontier models would consume all of it somewhere between 2026 and 2032. They've since nudged the window out to 2028, but the direction hasn't changed. The public web has been scraped. Books, Wikipedia, GitHub, Reddit, Stack Overflow are already in the training sets. Every new model needs more, and there isn't much more lying around for free.

So the labs have been buying. OpenAI signed a $250 million deal with News Corp. Reddit licenses its archive. Publishers, stock photo libraries, and forums have all been turned into line items. That data is public-facing though. It's articles and posts, written for an audience.

What Spirit's archive offers is different. It's private, internal, working communication. How a revenue team argues about pricing in a Teams thread. How an ops manager writes up an incident. How a finance lead explains a variance to an auditor. Nobody wrote it for publication, which is exactly why it's valuable. Models trained on the open web are good at sounding like a blog post. They're weaker at sounding like a real coworker inside a real company, because that text almost never leaves the building.

The sale agreement reportedly requires the links between records to stay intact, so a thread can be followed across email, chat, and files. That's what makes the dataset useful for training an assistant that understands how work actually flows. It's also what the union is worried about: with the links preserved, a scrubbed name is not much protection when the pattern of who talked to whom about what is still there.

The part that should get your attention

Set aside whether the judge approves this one. The precedent is the story.

Every company generates this archive. Your Slack, your email, your Notion, your Jira, your Git history. Right now that data is a liability you pay to store and a compliance headache when someone files a request. Google just established that, in the right circumstances, it's an asset with a market price. Roughly $10 million for a company of Spirit's size. Bankruptcy trustees have a legal duty to maximize the value of the estate, and they now know there's a buyer.

That means a few things for anyone running or advising a business.

First, what you put in your chat tools is now potentially permanent in a way it wasn't before. Not because someone will read it, but because a model might learn from it. Retention policies were a legal question. They're becoming a strategic one.

Second, the data scarcity problem isn't going away, so this kind of purchase will get more common, not less. Mercor bid on the same package. Micro1 also reportedly showed interest. Once a market exists, sellers appear. Expect to see data appear as an explicit asset class in more bankruptcies, acquisitions, and eventually as something healthy companies license on purpose.

Third, if you're building with AI rather than training it, this is a reminder of where the moat actually is. The general models are converging. What Google is paying for here is context, the specific texture of how a real organization operates. Your company's own operational data is that same asset for you. Most businesses never structure it well enough to use, which is why so many internal AI projects stall. The labs are paying millions for a stranger's version of something you already own.

Spirit Airlines spent 30 years flying people to Florida for cheap. Its last sale might be the most valuable thing it ever produced, and it was sitting in an inbox nobody thought to protect.