Why AI Labs Are Quietly Buying Bankrupt Companies' Data
The top bid came from Google: ten million dollars for the inbox of a dead airline.
Tech companies are quietly buying the emails, chats and code that bankrupt companies leave behind. This time a court put each bid on the record.
On August fourteenth, twenty twenty-six, Google won the bankruptcy auction for Spirit Airlines' inbox: one hundred million emails, plus five hundred million messages from Microsoft Teams.
Spirit stopped flying on May second. About seventeen thousand people lost their jobs.
The messages they wrote at work went up for sale with the rest of the airline.
If you've sent a work email or a Slack message this week, you've made the same kind of thing Google bid on.
A dead airline's mail is worth that much to an AI company for one reason. Since AI labs have trained on most of what's public, the next lesson has to come from somewhere private. A model that's meant to do office work needs to see office work: the question, the ticket that follows, the fix and the sign-off, all in order.
A company's inbox holds years of that. It works like a flight simulator for an AI, a copy of a company where it can practice the job before it does the job for you.
The losing bidder said it plainly.
A spokesperson for Mercor, a company that sells training data to AI labs, told reporters that companies are sitting on decades of records that show how work gets done.
Spirit's archive ran that long. The schedule filed in bankruptcy court lists eighty thousand email accounts, five hundred sixteen code repositories and employee files going back to nineteen eighty-six.
When a company goes bankrupt, a court sells what's left to pay back the people it owes. It works the way an estate sale clears out a house.
Spirit's planes, engines and airport gates went that way. This time the court also sold the airline's memory.
And the sale turned into a fight.
The bidding opened at Google's five million dollars. Mercor came back at five point two million.
Then Mercor tried a different offer: seven million dollars if it could take the raw records and strip out the names itself.
Google closed at ten million.
Mercor stands as the backup buyer at seven and a half million if Google's deal falls through.
Google's win came with conditions. Google picks the firm that scrubs the records and pays for the scrub on top of the price.
Scrubbing means removing names and personal details, the way a court blacks out names in a public filing.
The records also stay linked to each other after the scrub. So one unnamed employee can show up in an email, then in a support ticket, then in a change to the code.
That chain is what a buyer pays for.
It shows one person's work from start to finish.
Google says a third party will scrub each record of personal information before Google receives it. Customer profiles, loyalty records and call recordings were kept out of the sale.
Spirit is the headline because a court made its sale public. Smaller deals close in private.
The same buyers are shopping for something far smaller and far more common: the Slack of a startup that's closing.
Spirit's flight attendants' union objected to the sale on privacy grounds.
Work email carries private life.
An attorney who advises AI startups points to the details that hide in inboxes and calendars: doctor appointments, school pickups and date-night plans. He calls this kind of data legally radioactive.
Then comes the question of who gets paid. The money from these sales usually goes to creditors and investors. The people who wrote the messages get nothing from it.
Those are fair objections. A seller who ignores them is selling a lawsuit.
They also set the rules for doing this well.
Now shrink the seller.
Keelan Miskell ran a crypto startup for about two and a half years. This year he began closing it.
He figured its data was worth little. The users were anonymous and the company was young.
A wind-down firm called Simple Closure sold his code for about twenty-five thousand dollars. His team's Slack, email, Notion and Figma files were on track to bring in about fifteen thousand more.
He told MIT Technology Review, "At least all that stuff we were doing for two and a half years is worth something to someone somewhere."
Simple Closure started selling data for closing startups in April. It has handled more than a hundred and fifty of these deals.
Forty-seven percent of its closings now include a code or data sale.
Live companies are selling too. micro one, a company that supplies data to AI labs, publishes a starting tier of a hundred thousand dollars or more for a company's operating data.
A dead company sells its records once. A live one keeps making new records each year, so it can keep selling them.
When I covered this market for the Trends dot V C report on data-for-cash apps, one gap stood out between Spirit and the founder.
Spirit had two buyers who wanted the same thing. Google and Mercor pushed each other up, bid by bid, until the price doubled.
A small seller gets something else.
One buyer told MIT Technology Review that buyers quote wildly different numbers for the same kind of data. Each one is guessing how many useful lessons a pile of code can teach.
The same reporting found that founders can't see who's buying.
They have little assurance the offer is fair.
So picture a founder with a twelve-person team and four years of Slack. She's closing the company this month.
Her tools get cancelled when the bills stop.
Once a workspace closes, the records may be gone.
Somebody makes her an offer. She can't tell if it's high or low. The clock is running, so she takes it.
That's the opportunity: bring a contest like Spirit's to the companies too small to draw one. Take a cut of each sale.
Picture the same founder with a second option.
On Monday she fills out a short form: twelve people, four years, Slack, email and code. Nothing leaves her laptop yet.
Her direct messages, HR files, pay, legal advice and anything she holds for clients stay out. The names come off the sample before anyone sees it, because the union's objection is the first thing a seller will hear.
You show that clean sample to a handful of buyers you've approved: AI labs and data companies that have bought records before.
Spirit's bidders raised each other out loud. For a small sale, use sealed bids instead. Each buyer writes one number on a card and seals it in an envelope, the way some houses get sold. The envelopes open on Friday at noon.
Since a buyer can't see the other bids and gets one shot, it bids close to the most it'll pay.
On Friday she sees the offers side by side. She picks one or walks away. The sale closes before the workspace shuts.
Your cut comes out of the winning bid, so you earn more when she does.
The customer is a founder like her or the wind-down firm running her closing. Later it's the owner of a live company that wants to sell its records each year.
The pieces are on the market.
Korra lets an owner describe a company's records in about five minutes so approved buyers can make offers. The buyer pays Korra's fee, so the seller pays nothing. Simple Closure sells for closing startups, while micro one buys from live ones.
What's missing is the contest. Spirit's sale shows what a contest does to a price. A closing startup rarely gets one.
Spirit's one hundred million emails drew ten million dollars because two buyers wanted them at the same time. A twelve-person startup's Slack holds the same kind of work. Today it goes to whoever asks first.
Visit trends dot V C and join more than fifty thousand founders to get the next opportunity.