Google just paid $10 million for a bankrupt airline's internal emails.

Google's winning bid was $10 million. The lot: roughly 100 million emails and 500 million Microsoft Teams messages, plus spreadsheets, calendars, marketing materials, HR files, project management documents, financial databases, audits and presentations.

Six hundred million artifacts. About 1.7 cents each.

That is the first public price on an asset class that every company on earth owns and almost none has ever valued. Your Slack history. Your Teams threads. Twenty years of arguments about Q3 forecasts. Until this week, that was a compliance liability — something you pay to store and hope is never subpoenaed. A bankruptcy court just discovered it clears at a price, with competitive bidding.

First, two things worth getting right

Google hasn't paid. It won the auction. A federal judge still has to approve the sale, at a hearing scheduled for August 19. Until that gavel falls, this is a winning bid, not a completed purchase.

And the data is narrower than the headlines suggest. Per the court filing, the lot explicitly does not include passenger profiles or frequent flyer information. No customer list. No card data. The information is being "deidentified" by a third party before Google receives it, and Google has agreed not to attempt re-identification. Google's own statement: "We will not receive any personal information from this dataset."

So this isn't Google buying what Spirit knew about its passengers. It's Google buying what Spirit's employees said to each other. That's a stranger and more important purchase.

The underbidder is the real story

Everyone is looking at the winner. Look at second place.

The next-highest bid — $7.5 million — came from Mercor, an AI training-data company. Mercor's annualised revenue crossed $1 billion in February 2026 and hit $2 billion by June. It has been in talks at a $20 billion valuation, with roughly 90% of revenue coming from foundation model labs.

Read that again. This was not Google versus an airline creditor or a records-management firm. It was two AI data buyers bidding against each other for a defunct carrier's paperwork.

A liquidation auction quietly became a training-data auction. That's the market signal — not the $10 million.

Why an airline's inbox is worth anything

The public web is largely scraped. What frontier labs are short of now isn't knowledge — it's process.

Not what is the answer, but: how did eleven people disagree about a fuel hedge over four days and arrive at a decision? What did the spreadsheet look like on revision three, and on revision fourteen? What does an escalation actually sound like when it's real and consequential and nobody is performing for an audience?

If you're building models that answer questions, you want text. If you're building agents that do work, you need examples of work being done — messy, sequential, with the dead ends left in. Six hundred million of them, from a single organisation, spanning years and covering finance, HR, marketing and operations.

That is a very hard corpus to obtain legitimately. There is essentially one situation in which it becomes available.

And that's the uncomfortable part

This data is purchasable precisely because the company failed.

Spirit ceased operations earlier this year after it couldn't emerge from its second Chapter 11. A functioning airline would never sell its internal communications at any price — the legal exposure alone would kill the idea. The archive became an asset only when there was no longer a company to protect.

Which means the emerging corpus of "how organisations actually operate" will be assembled disproportionately from organisations that didn't survive. Every future dataset of this kind arrives via the same door: liquidation.

I don't know what that does to a model trained on it. But "learn how companies work, from companies that stopped working" is a survivorship bias running in the wrong direction, and I haven't seen anyone raise it.

Where I'd push back

Deidentification is carrying enormous weight here. Stripping names from 100 million emails does not strip the fact that colleagues describe each other, discuss compensation, name projects, and complain about specific managers in identifiable ways. Re-identification from context is a well-documented problem. Google's commitment not to attempt it is a contractual promise, not a technical impossibility.

And nobody consented to this. Not one Spirit employee wrote a Teams message anticipating it would become model training data. They had no seat at the auction. Their words were transferred as an estate asset, the same way a gate lease or a spare engine would be.

In fairness: this was disclosed in open court, scrubbed by a third party, and subject to judicial approval — which is more process than most data acquisition in this industry receives. The problem isn't that Google did something underhanded. The problem is that the mechanism is now proven, and every bankrupt company is sitting on inventory.

Question for you: if your employer went under tomorrow, would you be comfortable with a decade of your work chat clearing at 1.7 cents a message? Because that price is now discovered, and nobody asked us.

Address

Boston, Massachusets

Contacts

news@techproduit.com

Get your weekly newspaper