TESLA PLANS TO REVIVE DOJO3 FOR SPACE-BASED AI COMPUTE

The comment marks a notable shift for a project that was once positioned as a cornerstone of Tesla's in-house Al ambitions on Earth.

- 

Dojo was originally conceived to reduce reliance on third-party chips for training neural networks used in autonomous driving. Dojo3, however, was paused as Tesla prioritized near-term compute needs using Nvidia hardware. The decision to revive the project for space use suggests a reassessment of where custom silicon could offer long-term advantages.

WHAT DOJO3 WAS ORIGINALLY DESIGNED TO DO

Tesla introduced Dojo as a custom Al training system intended to process massive volumes of video data collected from its global vehicle fleet. Earlier iterations focused on accelerating neural network training for Full Self-Driving development by optimizing performance per watt and cost per training run.

Dojo2 advanced those goals incrementally, but Dojo3 was widely understood to be more experimental. Development was halted as Tesla leaned more heavily on commercially available GPUs, which offered faster iteration cycles and broader software support. At the time, Musk acknowledged that building competitive training silicon was resource-intensive and difficult to justify against rapidly advancing off-the-shelf alternatives.

WHY SPACE CHANGES THE EQUATION

Space-based computing presents a different set of constraints than Earthbound data centers. Power availability, heat dissipation, latency, and autonomy all become more critical once systems are deployed in orbit or beyond. In that environment, reliance on terrestrial cloud infrastructure is limited, and specialized hardware can offer practical benefits.