I’m Back. And Google Just Bought a Dead Airline’s Inbox for $10 Million.

Hi guys, sorry for the radio silence. This last month has been a lot: I packed up, moved to Atlanta, and started a master’s in machine learning and economics at Georgia Tech. Orientation, apartment hunting, and syllabus week ate every spare hour I had. But I’m back, and to the fifteen of you who actually read this thing regularly, thank you, genuinely. Let’s get into it.

I wanted my first post back to be about something I’ve been chewing on since before the move, and this week handed me the perfect hook. It involves an airline that no longer exists, an email server, and ten million dollars.

A Bankrupt Airline’s Inbox Just Became a $10 Million Asset

Spirit Airlines shut down for good this year after failing to claw its way out of a second Chapter 11 filing. No more planes, no more famously yellow livery, no more $39 fares with $200 in fees bolted on. What Spirit did have, sitting in a bankruptcy estate, was data, and this week a judge is being asked to approve Google’s winning bid to buy it for $10 million, beating out a competing $7.5 million offer from Mercor, an AI-training-data company.

Sit with that for a second. A company that could not generate enough revenue to keep 175 aircraft in the air produced, almost as a byproduct of a decade of ordinary operations, a dataset that a trillion-dollar company found worth fighting over in bankruptcy court. Nobody at Spirit was building that Teams archive to sell it someday. It was exhaust. And exhaust, it turns out, is now a line item.

Data Has Always Been Worth Something. Now It’s on the Balance Sheet.

The Spirit deal isn’t a one-off curiosity: it’s the cheapest, weirdest entry in a pattern that’s been building for two years. Every major AI lab has been quietly signing checks for access to data that used to just sit around being a company’s internal exhaust.

Here’s the fuller picture:

Look at the spread on that list. Ten million dollars for an entire defunct airline’s internal record versus fourteen billion for a stake in a labeling company. The number isn’t really about volume of data; Spirit’s trove is, by AI standards, small. It’s about what’s in it: real operational decisions, real customer complaints, real pricing logic, made by real people doing a real job for over a decade. That’s the kind of texture a model can’t get from scraping the open web, because it was never public in the first place.

Why Synthetic Data Is the Other Half of This Story

Here’s the tension underneath all of it: proprietary, real-world data like Spirit’s is exactly what labs want most, and it’s also the hardest kind to get more of. You can’t generate a decade of a real airline’s customer service transcripts. Researchers have been warning for a while now that the supply of easily scraped, high-quality public text is on track to be effectively exhausted sometime this decade. So the industry is pulling on two levers at once: buy what real proprietary data still exists (see: Spirit), and manufacture the rest.

This isn’t a fringe technique anymore. It’s becoming default infrastructure in a handful of fields where real data is either too scarce, too dangerous, or too regulated to use directly:

Estimates of the synthetic data market’s future size vary wildly depending on who’s selling the report, which is a good reason to treat any single number skeptically. But the direction every forecast agrees on is the same: this goes from being a research niche to being standard tooling, growing at somewhere north of 20% a year through the early 2030s. Real data buys you authenticity. Synthetic data buys you scale, coverage of rare cases, and a way around privacy law. Increasingly, teams need both, which is exactly why a $10 million bankruptcy auction for a dead airline’s Teams archive made national news the same month synthetic data tooling is turning into its own category of startup.

What This Actually Means If You Work in Tech

1. Your company’s exhaust is an asset now

   The Slack history, the support tickets, the internal wikis nobody reads: that’s not just clutter to be deleted at the next audit. Treat data retention and data governance as a strategy decision, not just an IT policy.

2. Synthetic data is a skill, not a hack

   Knowing how to generate and validate synthetic datasets, and knowing when they quietly diverge from reality, is turning into as core a skill as SQL was a decade ago.

3. De-identification is the new moat

   Google didn’t just buy Spirit’s data. It bought the ability to say, credibly, that it scrubbed it first. The tooling and process around anonymization is becoming a product category of its own.

4. Domain-specific beats general-purpose

   Nobody thinks Google needs Spirit’s data to make Gemini smarter at poetry. The bet is narrower: operational, industry-specific intelligence, exactly the kind of thing a boring internal dataset is full of.