SpaceX’s artificial-intelligence division is considering an unusual source of training material: other people’s failed companies. SpaceXAI has held informal internal discussions about buying customer and operating data from startups that are struggling or have already shut down, according to people familiar with the matter, with the data intended to train its Grok models.
The idea is to acquire high-quality external datasets that would improve the models’ performance. Grok has so far been trained largely on data from X, the social platform owned by Elon Musk, along with input from an internal team of “AI tutors” who refine its responses. A move into purchased third-party data would broaden that base.
The strategy has a precedent. Google previously bid to acquire business data from Spirit Airlines for AI training, treating the operational records of a carrier in distress as feedstock for its models. SpaceXAI’s discussions echo that approach: when a company fails, its accumulated data, from customer interactions to operational logs, can outlive the business itself.
The talks are early and may not produce a deal. The discussions are informal, the people said, and SpaceXAI has not committed to any purchase. SpaceX declined to comment. Even so, the exploration signals a shift in how the division thinks about data, away from exclusive reliance on Musk’s properties and toward a more diverse set of sources.
The context is a company trying to catch up. SpaceXAI has been working to expand its enterprise customer base and close ground with rivals such as Microsoft and OpenAI, which hold large leads in the market for AI services. In a field where model quality is the product, the quality of training data is a competitive lever, and the easiest way to get more data is sometimes to buy it.
The division has also been through internal turbulence. Its AI tutor team recently underwent a leadership change and a hiring pause, the people said, and the new head, Jack Garabedian, has been working to stabilize the group and set clear data objectives. A defined target for the data the models are trained on is part of that effort.
The external-data discussions do not mean SpaceXAI is stepping away from internal information. Musk has told employees that Grok will undergo a comprehensive training effort to master the company’s internal knowledge, a reminder that the model is meant to be useful to SpaceX and X as well as to outside customers. Internal data remains part of the plan.
The economics of buying data from failed startups are unproven. A dataset’s value depends on whether it contains the kind of signal that improves a model, and a company that went out of business may have left behind data of uncertain quality and unclear ownership. Negotiating rights to it can be as difficult as the purchase itself.
Grok has been trained primarily on X’s data, a resource that gives the model a distinctive voice and access to the real-time conversation on the platform, but one that is finite in the categories it covers. A social feed is rich in some kinds of text and poor in others, and the models competing for enterprise customers need the kind of structured, domain-specific data that lives inside operating companies rather than public timelines.
The scramble for training data has become one of the defining constraints of the AI buildout. The high-quality text on the public web has largely been consumed, and the developers have turned to licensing deals with publishers, synthetic data, and, increasingly, the acquisition of datasets from businesses. Buying data from failed startups is an extension of that logic into a market that barely exists yet, with unclear prices and unclear rights.
The legal and reputational questions are as real as the technical ones. Data from a company that failed may carry privacy obligations, contractual restrictions, and questions about who still owns it, and a model trained on it could inherit those liabilities. SpaceXAI’s discussions are early, and the company has not committed to any purchase, but the very existence of the talks shows where the industry’s search for data has begun to lead.
The division’s enterprise push is the commercial context for the data search. Competing with Microsoft and OpenAI means offering models that are good enough to justify a switch, and model quality in turn depends on data the competition does not have. A unique dataset, even one bought from a failed company, is a way to differentiate in a market where the leading models otherwise converge on the same public sources.
Still, the search is telling. The largest AI developers have begun to look beyond the public web for training data as high-quality sources grow scarce, and the assets of dead companies are one of the few untapped pools left. SpaceXAI’s willingness to consider them places it in the same scramble as its larger rivals, even if its own talks remain at the earliest stage.
