SpaceX Weighs Buying Failed-Startup Data to Train Grok, Bloomberg Reports
SpaceX has discussed buying customer and operational data from struggling or failed startups to train its AI models, Bloomberg News reported, as Elon Musk seeks cheaper real-world data for Grok.
The talks remain informal inside SpaceXAI and may not lead to a transaction, Bloomberg reported. SpaceX did not respond to a request for comment.
The data push fits Grok's existing training strategy. X, formerly Twitter, has provided real-time content, debates, memes, sarcasm and fragmented expression that shaped Grok's personality. A team of AI tutors, made up of domain experts and annotators, added professional capabilities in areas such as finance, science and code. Bloomberg reported in June that xAI had paused AI tutor hiring and later changed the team's leadership.
Diego Pasini, the former lead, entered the University of Pennsylvania in 2023 after high school and later took leave to join xAI, where he helped manage Grok's data annotation team, according to the report. He was replaced by Jack Garabedian, a SpaceX veteran who joined the company in 2021 and worked at Starlink for years. Starlink's global business spans satellite networks, terminal devices, subscription services and customer operations.
The leadership change signals that Grok's data training is moving from an early sprint to an engineered production line, the article said. Musk made the direction explicit at a recent SpaceX employee meeting, saying, "We will use all SpaceX information to train Grok, including you." That means Grok would learn not only from SpaceX code, documents and processes but also from traces employees leave in their work.
Customer information and operational data would fill a third gap: real business scenarios. The article said X can shape personality and AI tutors can add professional skills, but neither easily teaches how customers place orders, how tickets move, how services are delivered or how a company runs day to day. Operating companies may be unwilling to sell such data or may charge a high price, making distressed or failed startups a practical target.
Musk is not alone. After U.S. budget carrier Spirit Airlines went through bankruptcy restructuring, a batch of internal business data entered an auction. Google bid $10 million for de-identified business data, operational records and software code to train AI models and improve products, according to the article. The data included internal emails, Teams chats, collaboration records, revenue, flight operations, employee productivity, audit and fraud-related information. The deal was quickly shelved over possible privacy leaks.
The logic is similar for Google and SpaceXAI: if agents are to work inside enterprises, they must understand how real businesses run. Whoever obtains more real business data has a better chance of moving models from chatting to doing work, the article said.
The context is a competitive AI market. The Wall Street Journal reported in May that Grok downloads fell from more than 20 million in January to 8.3 million in April. Paid conversion was also weak. An engineer quoted in the report compared OpenAI to Coca-Cola, Anthropic to Pepsi and Grok to RC Cola, adding that he had never seen anyone drink it.