SpaceX has held informal internal talks about buying customer and operational records from troubled or defunct startups to train its AI models, Bloomberg reported. The buyer would be SpaceXAI, the unit formerly known as xAI, which wants cheap operational data and customer information to feed Grok. The talks may not produce a deal, but the very direction matters: a company with launch and satellite operations is now looking at other firms' customer lists as a training resource.
What stands behind the talks
The template is Google's $10M offer for the records of the collapsed airline Spirit, which was later topped by a counterbid. That archive held 100 million emails and 500 million Teams messages. SpaceXAI is not bidding at one auction; it is describing a method, which means the company is testing whether distressed corporate data can become a repeatable supply channel rather than a one-off purchase. SpaceX did not respond to Bloomberg's request for comment on the discussions.
The mechanics differ from the usual way AI companies obtain training material. Instead of scraping public web pages or licensing content from publishers, a buyer would take over records that already sit inside a failing business: support tickets, account histories, internal correspondence. Such data is structured, tied to real transactions and often cheaper than licensed archives, because the seller is winding down and needs to monetise remaining assets. The buyer becomes a new controller of the data and must establish its own lawful basis for processing, which is a harder problem than price.
The shift is visible in how SpaceXAI has trained Grok so far. The unit used posts from X and in-house specialists it calls AI tutors. Hiring for that team was paused in June, and a Starlink staffer, Jack Garabedian, has since taken it over. Elon Musk has been open about the wider ambition: he told staff in August that Grok would be trained on the sum total of all SpaceX information, according to Bloomberg. That covers his own employees; buying a failed startup covers other companies' customers.
What this means for business
For companies that adopt AI, the practical consequence is that operational records are turning into a tradable asset class. A small firm that shuts down may find its customer database has a buyer, while a large enterprise evaluating vendors should expect questions about where training data came from. The difference between a small and a large company here is leverage: a startup in distress rarely has the legal team to negotiate processing terms, while a bigger seller can set conditions on how records are used and for how long.
Limitations and risks remain substantial. In Europe, Ireland's Data Protection Commission opened an inquiry in April 2025 into whether EU and EEA users' public posts were lawfully processed to train the Grok models, after using emergency High Court powers against X the previous August. SpaceXAI's European privacy notice relies on legitimate interests to improve its models and names a US entity as the controller for European users, with all processing in American data centres. Under the European Data Protection Board's 2024 opinion, that basis turns partly on what people would reasonably expect, and a customer of a company that no longer exists has no expectations about Grok. Consent does not travel with a database either. SpaceXAI's record on data it gathered itself is already mixed: it asked staff for tax returns in exchange for $420 and did not pay.
The marker to watch is whether SpaceXAI moves from describing a method to signing a first purchase agreement, and whether any such deal covers European records. A signed contract involving EU customer data would test the legitimate-interests argument in practice and show other AI developers whether distressed archives are a viable route or a regulatory trap. Until then, the talks remain informal and may not produce a deal.
