SpaceX Eyes Buying Data from Failed Startups to Train Grok AI Models

SpaceX Pursues Unconventional Data Strategy for AI Training

SpaceX has begun internal discussions about buying customer and operational data from troubled or defunct startups to train its AI models1

1

2

2

. The talks, currently informal and not yet resulting in confirmed deals, represent a strategic shift as the company seeks cheaper alternatives to fuel Grok's development. The buyer would be SpaceXAI, the artificial intelligence division that emerged in February when SpaceX merged with xAI4

4

. This approach targets digital assets of defunct startups that can no longer object to how their data gets used, offering access to proprietary datasets at bargain prices compared to licensing agreements with operating companies.

The strategy addresses a critical challenge facing AI development: securing high-quality AI training data. SpaceX committed $15.8 billion to AI development in its previous quarter, even as xAI ended with a $1.3 billion net operating loss3

3

. Currently, Grok lags behind competing models from Anthropic, OpenAI, and Meta across multiple metrics. The company already uses X posts to train Grok models unless users opt out, but these models need additional real-world business information to improve their outputs and use cases1

1

.

Google's Spirit Airlines Deal Sets Precedent

The template for this approach comes from Google's recent $10 million bid for Spirit Airlines' business data at a bankruptcy auction1

1

. That archive contained 100 million company emails, 500 million Microsoft Teams messages, aircraft operations data, employee productivity information, 1 million time-card records, and hundreds of thousands of employee records including tax forms and employment contracts3

3

. The difference is that SpaceXAI is describing a method rather than bidding at one specific auction, signaling a systematic approach to data acquisition2

2

.

Chapter 11 bankruptcy proceedings treat customer databases like office furniture, auctioning them to the highest bidder without requiring customer consent4

4

. By the time the gavel drops, there's no company left for customers to complain to, creating a legal grey area that AI companies are increasingly exploiting during this AI gold rush.

Privacy Concerns and Regulatory Scrutiny

The strategy immediately triggers ethical and privacy concerns. Ireland's Data Protection Commission opened an inquiry in April 2025 into whether EU and EEA users' public posts were lawfully processed to train the Grok models2

2

. The regulator had already used emergency High Court powers against X the previous August. Under data protection laws, consent does not travel with a database—the buyer becomes a new controller and must find its own lawful basis, which presents a harder problem than price2

2

.

A customer of a company that no longer exists has no expectations about Grok. SpaceXAI's European privacy notice relies on legitimate interests to improve its models and names a US entity as the controller for European users, with all processing happening in American data centers2

2

. However, the European Data Protection Board's 2024 opinion states this basis turns partly on what people would reasonably expect.

The Association of Flight Attendants filed an objection with the bankruptcy court regarding Google's Spirit Airlines purchase, arguing that "de-identifying" data only covers consumers, leaving employees' private data vulnerable3

3

. "A pseudonymized dataset can still disclose which crew bases generated grievances, how a small subset of flight attendants performed on recurrent training, which employees were subject to investigation, what compensation adjustments followed which events, and what employees said to one another about management, staffing, or their union," the objection stated. The bankruptcy court has postponed approval of the purchase3

3

.

Ambitious Plans for Grok Development

Elon Musk has positioned this data acquisition as essential for Grok's advancement. xAI's current flagship model is Grok 4.6, with plans to release Grok 4.7 this month and Grok 5 launching before year's end1

1

. Musk claims Grok 5 will reach AGI-level capabilities3

3

. "With Grok 5, which should be out before the end of this year, we will be incorporating the entire corpus of SpaceX data," Musk said during SpaceX's first earnings call last month. "Basically, all the data that SpaceX has ever produced, which is a tremendous amount over the course of a quarter century, will be incorporated into Grok training. We think this will make Grok by far the best engineer"1

1

.

In August, Musk told SpaceX employees the company would train Grok on the entirety of SpaceX's internal records. "It will inherit your thoughts," Musk reportedly told staff, framing it as employees raising the model since it would absorb their outlook along with their work4

4

. This represents a change of direction from SpaceXAI's previous approach of training Grok on X posts and using in-house AI tutors, a hiring program that was paused in June2

2

.

Broader Industry Implications

This development reflects a wider industry struggle to secure training data ethically. Numerous AI companies face legal battles over web scraping efforts. The New York Times filed a lawsuit in 2023 against OpenAI for violating intellectual property rights, alleging "millions of articles" were used without permission5

5

. Anthropic triggered controversy when it emerged the company was purchasing old books en masse, scanning and destroying them to train its AI models5

5

.

Facilitating failed companies' sale of employee data to AI labs has become a niche market of its own as agentic AI makes its way into the workplace3

3

. The 23andMe collapse offers another example—Regeneron Pharmaceuticals paid $256 million for the company in May 2025 but committed to abiding by existing data protection protocols5

5

. Watch how bankruptcy courts handle data privacy objections and whether regulators establish clearer guidelines for AI companies acquiring data from troubled startups. The outcome will shape whether buying data from failed startups becomes standard practice or faces significant legal barriers.