SpaceX (SPCX.US)'s AI division targets “bankruptcy data”! Proposed acquisition of troubled startup customers and operational information to enhance Grok model training
The Zhitong Finance App learned that according to people familiar with the matter, Musk's SpaceX (SPCX.US) has discussed purchasing customer and operating data from troubled or bankrupt startups, hoping to obtain more high-quality data at a relatively low cost to improve the performance of its artificial intelligence models.
People familiar with the matter said that the relevant discussions are currently mainly taking place within SpaceX's AI division, SpaceX. They are still in an informal stage, and no deal will necessarily be reached in the end. SpaceX did not respond to requests for comment.
This idea is quite similar to Google's previous approach. Earlier this year, after the US low-cost airline Spirit Airlines ceased operations, Google proposed to buy its commercial data for AI training for $10 million. However, the deal also raised data privacy concerns, including some former Spirit Airlines flight attendants questioning it.
AI competition intensifies SpaceXAI seeks more external high-quality data
SpaceXAI was formerly XAI. As Musk's space, satellite communications, and artificial intelligence businesses are further integrated, the company is competing with AI companies such as Anthropic and OpenAI, and is seeking to attract more enterprise customers.
Like other AI companies, one of the key tasks SpaceXAI is currently facing is to find high-quality training data that can improve the model's performance on different missions. According to people familiar with the matter, the company is particularly concerned about business operation data and customer information, and plans to use this external data for training models such as Grok.
If finally implemented, this will mean some changes in SpaceXAI's data strategy.
In the past, the company mainly relied on data from Musk's social networking platform X and an “AI mentor” team composed of internal professionals to train and optimize artificial intelligence models. These AI mentors will help engineers improve the model's performance in professional fields such as finance, science, and even humor.
Today, SpaceXAI is considering introducing more external data sets, reflecting that as AI model capabilities continue to improve, it may be difficult to fully meet model training and optimization needs by simply relying on internal data sources. For leading AI companies, being able to legally obtain data with professional, authentic, and structured characteristics is becoming an important resource for improving model performance.
Reorganize the AI training team to speed up the improvement of the Grok training system
Meanwhile, SpaceXAI's internal data team responsible for model training has also recently experienced a series of personnel and strategic adjustments. In June of this year, SpaceXAI temporarily suspended the recruitment of AI mentors responsible for training Grok, then adjusted the team's leadership. Jack Garabedian, who worked for a long time at Starlink, a satellite internet business owned by SpaceX, recently took over the team to replace young engineer Diego Pasini.
People familiar with the matter said that Garabedian has been trying to improve the team's internal conditions since taking office, including establishing more clear work processes, meeting mechanisms, and AI training data goals.
A SpaceXAI internal communication seen by the media showed that the progress made by the data team in July included advancing the development of a new AI model and programming intelligence. The newsletter emphasized that without basic work such as labeling, preference alignment, and verification carried out by the team on a daily basis, it would be difficult for these models to eventually become mature products. This also highlights that in addition to computing power and model architecture, data processing and manual feedback are still an important part of AI model training.
External procurement and internal data go hand in hand, Musk plans to let Grok learn “all SpaceX information”
Although SpaceXAI is studying to obtain more external data sources, Musk's huge commercial footprint itself is still an important source of data for Grok, including even information generated by SpaceX employees.
Musk recently said at a SpaceX internal meeting that the company plans to use all SpaceX internal information to train Grok and directly told employees, “It will also use your data for training.”
This statement means that SpaceXAI's future data strategy may adopt an “internal+external” parallel model: on the one hand, it will continue to use data generated by Musk's companies and platforms such as SpaceX and X, and on the other hand, it will further expand the breadth and expertise of training data by purchasing external data sets.
It is worth noting that while purchasing data from troubled companies or companies that have gone out of business may provide a relatively low-cost data source, the use of customer information may also involve issues such as privacy, authorization, and scope of data use. The controversy caused by Google's previous bid for Spirit Airlines data has shown the potential compliance and privacy risks of such transactions.
As companies such as OpenAI, Anthropic, and SpaceXAI compete for model performance and enterprise customers, high-quality, specialized data is becoming another key resource for AI competition after chips, computing power, and electricity. SpaceXAI is considering purchasing operational and customer data from troubled startups, which also shows that the competition for training data among leading AI companies is further expanding into the field of private enterprise data.