Google pays $10 million for Spirit Airlines' internal emails to train AI

By Nate from Nate's Substack · October 8, 2026.

In August, Google won a bankruptcy auction — still pending court approval — with a $10 million bid for Spirit Airlines' internal business data. The proposed purchase includes emails, Microsoft Teams messages, spreadsheets and other records accumulated while employees did their jobs. Google has said it wants that material for AI training and product development.

Nate uses this news as a starting point to reflect on something he finds strange: seeing a price put on all those days of work, including the meeting someone stayed late for, the document that needed another revision, or the conversation that finally solved a problem. A buyer can acquire the records as a whole, but according to the author it still has to work out what each of them actually achieved.

From there he poses the article's central question: what is work for? He illustrates it with his own experience writing product requirements documents at Amazon, where version numbers could reach the thirties. He notes that saying version 37 was better than version 36 cannot be explained simply by comparing the changes in the file: usually what mattered was that an argument had become clearer, that the team had discovered it was solving the wrong problem, or that two people who had spent entire meetings agreeing finally understood they meant different things. The next document might contain only a few altered sentences, even though the conversation behind them had changed what the team was going to build.

Nate extends this idea to presentations: a beautiful presentation can accomplish very little, because getting people to admire the slides is a different achievement from helping them understand a decision well enough to make it. The real work, he says, lies in alignment, and it sometimes leaves very small traces: a meeting that ends early because a hard question was finally settled, an unnecessary project that never gets started, or a feature that is cut from the roadmap because someone better understood what the customer needed.

The author connects this to the unease AI generates: we use what we produce to understand our own contribution, and when a machine produces something similar in a minute it can feel as if someone had revised the value of all that effort without consulting us. He suggests taking that feeling seriously and asking what that effort was really helping to achieve, honestly acknowledging that sometimes we developed a valuable understanding of a problem and other times we simply got very good at meeting expectations that benefited no one.

Still, he insists he does not want to turn the article into a defense of the value of past work, because what matters most to him is the possibility that opens up ahead: if a team can research a problem, build a prototype and test an idea with less time and money, some projects that were previously out of reach become possible. He gives the example of a product team preparing a proposal: if AI helps it write the proposal faster, it could use that time to produce more proposals, or instead build a small working version of the idea and put it in front of the people it is supposed to help — which changes what the team can learn before asking the company to commit.

He closes the article by noting that Google's bid puts a price on the possibility of learning from work already done, while he is more interested in what can be learned by examining your own work while there is still time to change it: where the effort leads, who benefits if it succeeds, and what things had been ruled out on the assumption that they would take more time, money or people than were available.

The email also includes promotion for 'Nate's Library MCP', a read-only MCP server that gives any AI access to the complete archive of articles Nate has published, and for his Slack community, along with links to other related posts from his Substack ('Nobody knows what you're worth anymore', 'How to tell if your job is on thin ice', 'The default trap: why Anthropic's data policy change matters', among others).