SpaceX Weighs Buying Customer Data From Failed Startups to Feed Grok's Training Pipeline

SpaceX is exploring an unconventional shortcut to sharpen its AI models: scooping up customer records and operational data from startups that are struggling or have already shut down.

Teams inside SpaceXAI, the company's artificial intelligence division, have held preliminary conversations about acquiring such datasets as a lower-cost way to feed training material into Grok and other models, according to people familiar with the matter who asked not to be identified because the discussions are private. The talks remain informal, the people said, and may never produce an actual transaction.

The idea mirrors a move Alphabet made earlier this year. After Spirit Airlines ceased operations, Google parent Alphabet offered $10 million for the carrier's business data to support AI training. That bid drew objections from some former flight attendants and sparked broader questions about whether customer information should be repurposed for machine learning.

SpaceXAI is chasing the same kind of real-world information that makes enterprise AI systems more useful, said the people. Public data can teach a model how to write a memo or summarize a contract, but it cannot easily replicate how an actual company logs transactions, handles support tickets, or manages inventory. Buying that operational history from a defunct firm could give Grok a window into genuine business behavior at a fraction of the cost of generating or licensing equivalent datasets.

The push comes as SpaceXAI fights for enterprise customers against Anthropic and OpenAI. Both rivals have spent heavily on proprietary data pipelines, and SpaceXAI has so far leaned on two main sources: the firehose of content flowing through Musk's social platform X, and a team of internal specialists known internally as AI tutors.

Those tutors help refine Grok's performance in areas ranging from finance and science to humor. But the group has been through a turbulent stretch. Reports in June said SpaceXAI paused hiring for the tutor program, followed by a management reshuffle. Jack Garabedian, a longtime Starlink veteran, recently took over the team. Since then, he has worked to ease internal frustration, establish clearer workflows and meeting rhythms, and set more explicit goals for how training data should be produced, the people said.

An internal SpaceXAI communication reviewed by Bloomberg summarized July achievements that included progress on new AI models and a coding agent. The message credited the tutor team's behind-the-scenes work, saying that without "the rigorous, meticulous labeling, preference alignment, and validation work this team does quietly every day," none of those products would have matured into finished offerings.