Elon Musk has announced that Tesla plans to restart development of Dojo3, the company’s third-generation AI chip project that was previously shelved. However, this time around, the focus will shift dramatically: Dojo3 will not be used for training self-driving systems on Earth, but instead will power “space-based AI compute,” Musk said in a post over the weekend. The announcement marks a reversal of Tesla’s earlier strategy. Five months prior, the company effectively shut down its Dojo initiative after Peter Bannon, the project’s lead, departed. Around 20 team members followed him to DensityAI, a new AI infrastructure startup founded by Bannon, former Tesla engineer Bill Chang, and Ben Floering. At the time, Tesla signaled it would rely more heavily on external partners like Nvidia and AMD for computing power, and on Samsung for chip manufacturing, stepping back from its ambitions to build custom silicon. Musk’s latest statement suggests a pivot back toward in-house chip development, driven by progress on Tesla’s AI5 chip, which is being produced by TSMC and designed to support the company’s autonomous driving systems and Optimus humanoid robots. The AI5 chip was unveiled last summer, and Tesla later signed a $16.5 billion agreement with Samsung to produce its next-generation AI6 chips—meant to power vehicles, Optimus, and high-performance AI training in data centers. Now, Musk is positioning the next phase as a bold leap forward. “AI7/Dojo3 will be for space-based AI compute,” he wrote on X, framing the project as a long-term, high-risk endeavor. To bring it to life, Tesla is beginning to rebuild the Dojo team, with Musk directly recruiting engineers via a post urging candidates to email [email protected] with three bullet points on the toughest technical problems they’ve solved. The timing of the announcement is significant. At CES 2026, Nvidia introduced Alpamayo, an open-source AI model for autonomous driving that directly competes with Tesla’s Full Self-Driving (FSD) software. Musk acknowledged the challenge of handling rare, complex driving scenarios—what he called the “long tail”—and even expressed hope that Nvidia’s effort succeeds, calling it “super hard” to solve. Musk and other AI leaders have increasingly speculated that the future of data centers may lie beyond Earth. With terrestrial power grids under strain, off-planet computing offers a potential solution. Reports from Axios indicate that OpenAI CEO Sam Altman is also exploring the idea of orbital data centers. Musk has a distinct advantage: he controls SpaceX and its Starship launch vehicle, which could be used to deploy a constellation of satellites designed to operate in continuous sunlight, enabling 24/7 solar-powered AI computation. Still, major technical hurdles remain—particularly cooling high-power processors in the vacuum of space. Despite the challenges, Musk’s announcement follows a familiar pattern: proposing a seemingly far-fetched vision, then attempting to build it through sheer engineering force. Whether Dojo3 becomes a reality in orbit or remains a conceptual moonshot, it underscores Tesla’s ongoing ambition to control every layer of its AI stack—now extending beyond the road and into the stars.