According to Bloomberg, which cited people familiar with the matter, the AI department under SpaceX once had informal internal discussions about considering acquiring customer data and operational records left by a defunct startup company for use in training Grok. This idea is still at the discussion stage and may not be pursued, but it reflects that AI is turning its attention to data sources that are more difficult to obtain but also more commercially valuable.

The goal is to obtain real business data.

This type of data is different from content scraped from public web pages; it is more closely related to emails, internal collaboration records, customer information, and process documents that are generated in the daily operations of enterprises. For training models, these materials provide a more complete business context, and therefore their value is on the rise.

The report mentions that SpaceXAI was established in February of this year, resulting from the integration of SpaceX and xAI. The focus of their internal discussions is not on purchasing licensed data from companies that are still in operation, but rather on the residual data assets of enterprises that have gone bankrupt or ceased operations. Since the original companies no longer exist, such data is often sold along with other assets during bankruptcy liquidation.

There are precedents for similar practices.

Google has previously taken similar actions in bankruptcy auctions. Reports indicate that Google won the internal business records of Spirit Airlines for $10 million. This set of documents includes a large number of emails, Microsoft Teams messages, and employee profiles, and was incorporated into the AI training process.

However, such transactions have also sparked debates regarding the boundaries of privacy and data use. The union of employees related to Spirit Airlines once raised objections in the bankruptcy court, arguing that even if the data is “de-identified,” long-term internal communication records could still be reassembled to identify specific individuals. These legal disputes are still ongoing at present.

From the perspective of bankruptcy disposal processes, customer databases, like office equipment, are generally considered sellable assets that can be used to repay debts. The issue is that the customers and employees whose data is involved often do not directly participate in such decision-making regarding the sale.

Musk also demanded the excavation of internal records.

In addition to external data, Musk is also pushing for the use of SpaceX's own internal materials to train models. Reports indicate that during a company-wide meeting in August this year, he told employees that Grok would use nearly all of SpaceX's internal records for training, and claimed that this model would "inherit" the employees' ideas and ways of working.

The report also states that Musk described this approach as part of the model's development, hoping that employees would see themselves as "cultivating" the model. However, there is currently no clearer public explanation regarding which employee data will be included in the training process or how it will be used specifically.

Grok 4.5 was released in July, and this is the first model version after the integration of xAI. As the model iteration continues to advance, the sources of training data, the scope of its use, and how employee and customer information are handled may become areas of continued concern for the outside world.