Elon Musk says SpaceX will pour its engineering data into Grok’s next training run. He excluded any material restricted under U.S. arms-export law from the set.
Musk shared the plan on X on July 21, calling it a direct boost to Grok’s engineering skills.
SpaceX Data Joins the 2 Trillion Parameter Run
Musk wrote that SpaceX’s engineering data will enter supplemental training for a new model built on 2 trillion parameters. That figure nearly doubles the 1.5 trillion parameters behind the V9 architecture that powers Grok 4.5’s public launch.
The billionaire excluded data covered by the International Traffic in Arms Regulations, the U.S. export control law that restricts sharing defense and space technology information abroad. That carve-out keeps SpaceX compliant. It still feeds rocket and satellite engineering knowledge into the model.
SpaceX built that data trove during its rapid rise as a public company. The firm completed a record initial public offering earlier this year, giving it fresh capital to expand engineering and AI work side by side.
Grok Build Draws a Split Verdict From Coders
Musk’s post responded to a developer who ran weeks of side-by-side tests between Grok Build and OpenAI’s Codex. The developer said Grok Build now beats Claude on complex coding tasks, though Codex remained solid too.
That praise follows a rougher stretch for the tool. Reviewers previously flagged Grok 4.5’s mixed benchmark results even as Musk framed the model as Opus-class, a reference to Anthropic’s top-tier Claude line.
SpaceX Leans on Its New Scale
The data pledge follows SpaceX’s $60 billion push into coding tools, completed through its acquisition of Cursor’s parent company weeks after the IPO. Musk has separately made a trillion dollar revenue bet on SpaceX by 2030.
xAI kept expanding Grok’s reach at the same time. The official Grok account announced that Grok for Excel is now live. The tool lets users build financial models, analyze market data and generate charts with Grok 4.5.
That launch shows xAI layering new products onto the same foundation model SpaceX’s engineering data will soon help train. Feeding proprietary rocket and satellite data into Grok ties that scale directly to the model’s roadmap.
It also hands Grok an engineering data set that rivals such as OpenAI and Anthropic cannot easily match. Whether the 2-trillion-parameter model closes the performance gap with Claude and Codex will depend on real benchmark results, not on Musk’s own claims alone.