SpaceX discusses buying failed startups’ data to train Grok

SpaceX has held informal internal discussions about purchasing customer and operational records from troubled or defunct startups to feed its Grok artificial intelligence models, Bloomberg reported on Thursday. The talks, which may not lead to a deal, represent an unusual new frontier in the AI industry's hunt for proprietary training data.bloomberg

A New Data Strategy

The effort is being driven by SpaceXAI, the company's AI division formerly known as xAI, which is looking for affordable datasets beyond the public X posts and in-house "AI tutors" it has relied on so far. The hiring program for those tutors was paused in June, and a Starlink staffer, Jack Garabedian, has since taken over the team, according to Bloomberg.thenextweb

The approach mirrors Alphabet Inc.'s ↗ 1.35% reported $10 million offer to acquire business data from Spirit Airlines after the discount carrier wound down operations — a bid that drew privacy complaints from former flight attendants. But where Google targeted a single auction, SpaceXAI appears to be developing a broader acquisition method rather than bidding on one archive.investing+2

Elon Musk has signaled the ambition internally. "We are going to be training Grok on the sum total of all SpaceX information," he told staff in August, according to Bloomberg. He added that the model would also be trained on employees themselves — a scope that buying defunct companies' records would extend to other firms' customers.thenextweb

Privacy Questions in Europe

The potential purchases raise fresh regulatory concerns, particularly 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 Grok, having already invoked emergency High Court powers against X the previous August.thenextweb

Under the European Data Protection Board's 2024 guidance, the "legitimate interests" basis that SpaceXAI claims for processing turns partly on what data subjects would reasonably expect. Customers of a company that no longer exists would have had no expectation their information would end up training an AI model. Consent, moreover, does not transfer with a database — any buyer becomes a new data controller and must establish its own legal basis for processing.thenextweb

The Broader AI Data Race

The discussions underscore how the competition among AI developers is shifting from raw computing power to access to unique, real-world datasets. Failed startups can hold years of customer interactions, operational logs, and proprietary business information unavailable through conventional web scraping. SpaceX did not respond to Bloomberg's request for comment.investing