xAI's Next Grok Is Training on SpaceX's Own Data

Elon Musk has given xAI's next model a launch window and an unusual ingredient. In a post at the start of September he said Grok 4.7 would arrive in about ten days, putting the target around September 12. The headline number is size. Grok 4.7 is reported at roughly 2.1 trillion parameters, about 40 percent larger than the 1.5 trillion of Grok 4.6.

The detail worth pausing on is what the model is reading. After finishing its main pre-training run, Grok 4.7 is said to be absorbing a body of data no rival can copy: SpaceX's internal engineering records. According to reporting on the release, that includes Starlink satellite telemetry, rocket development and failure logs, and internal documents produced by a workforce of roughly fourteen to fifteen thousand engineers.

A private moat, not a bigger scrape

That choice runs against the grain of the field. Most labs have spent the past year worrying that they are running out of fresh public text to train on. Musk's answer is to reach for material the open web never had. Proprietary industrial data, especially hard-won records of what works and what blows up on the launch pad, is exactly the kind of information a general web crawl will never contain. If it makes Grok measurably better at engineering and physical reasoning, it becomes an advantage competitors cannot simply download.

It also raises questions the company has not answered. Feeding one firm's confidential records into a commercial model that anyone can query invites obvious concerns about what might leak back out, and about whether SpaceX's own people signed up for their work to become training data. Musk's habit of routing resources between his companies has drawn scrutiny before, including a recent legal setback for xAI in Minnesota.

For now the claims sit on Musk's word. xAI has published no model card, no benchmark results, no context window, no final pricing, and no confirmation of how many of those 2.1 trillion parameters are active at once. Musk says the model beats Grok 4.6 across the board and uses tokens more efficiently, while conceding it may run a little slower given its size. The company has been busy on other fronts, having just added Grok 4.6 to Microsoft's Foundry platform alongside its enterprise push. Whether the SpaceX gamble pays off is a question the benchmarks, once they exist, will have to settle.