Tesla hitting “restart” on its Dojo supercomputer project isn’t just a random engineering update — it’s another sign that Elon Musk still wants Tesla to own as much of its AI stack as possible, even as much of the industry keeps leaning harder into Nvidia and cloud-based compute.
What makes the move notable is that Dojo was essentially shelved not that long ago. Tesla disbanded the Dojo team last year after its internal chip roadmap started converging on newer architectures like AI5 and AI6, and key talent moved on — including Dojo lead Peter Bannon. Several former Dojo employees even left to join DensityAI, a new AI infrastructure startup. At the time, reports suggested Tesla would rely more heavily on outside partners like Nvidia and AMD for compute and Samsung for manufacturing. Musk’s latest comments suggest Tesla is shifting course again.
According to Musk, Tesla is restarting Dojo3 now that the AI5 chip design is “in good shape.” In other words, this isn’t a reboot driven by hype or temporary GPU shortages — Tesla is bringing it back because it believes the technical foundation is finally solid enough to justify putting resources into it again. That matters because building custom silicon and training infrastructure is expensive, slow, and risky, and most companies decide it’s easier to buy what they need rather than build it themselves.
The bigger twist is what Musk says Dojo3 is actually for. Historically, Dojo was designed to train Tesla’s machine learning models using massive volumes of video and driving data collected from the Tesla fleet — all aimed at improving Full Self-Driving. Now Musk is framing “AI7/Dojo3” as something much bigger: “space-based AI compute.” It’s a classic Musk leap from practical to sci-fi, and while he hasn’t laid out the exact plan, the idea fits a broader argument he and other AI leaders have floated recently — that Earth’s power grids are already strained, and future data centers may eventually move off-planet to take advantage of constant solar energy and natural cooling conditions. Musk also has a built-in advantage in that race: he owns the rockets. Axios has reported that he could potentially use a future SpaceX IPO to help fund Starship launches that would put compute infrastructure into orbit.
Even if you ignore the space angle for a moment, the Dojo restart still points to something important: Tesla doesn’t see AI as just a feature upgrade — it sees it as the foundation for everything it wants to become, including self-driving vehicles and humanoid robots like Optimus. Musk has claimed AI4 alone could eventually exceed human driving safety, and that AI5 could make Tesla vehicles “almost perfect” while significantly enhancing Optimus. Those are huge claims, but they explain why Tesla might be stubborn about building its own chips and training systems instead of relying entirely on Nvidia’s roadmap.
The timing is also interesting. At CES 2026, Nvidia unveiled Alpamayo, an open-source autonomous driving model that directly challenges Tesla’s Full Self-Driving narrative. Musk responded by acknowledging that solving rare “long tail” edge cases in driving is incredibly difficult, and even said he hopes Nvidia succeeds. It was a friendly comment, but it also underscores the competitive reality: Tesla is no longer the only serious player pushing hard on autonomy, and the next phase of competition is increasingly about compute, training scale, and who can iterate the fastest.
Of course, restarting Dojo doesn’t guarantee Tesla can deliver. Manufacturing custom chips at scale is brutally hard, and the gap between a working design and production-ready hardware is where plenty of ambitious projects die. Tesla also has a history of aggressive timelines and bold promises, especially around Full Self-Driving, that haven’t always matched reality. Still, the Dojo restart makes one thing clear: Musk isn’t backing away from the idea that the future belongs to companies that control the full AI pipeline — from the data to the models to the hardware. Whether Tesla’s version of that future ends up in a terrestrial data center or floating in orbit is another question entirely.
This analysis is based on reporting from Interesting Engineering.
Image courtesy of Unsplash.
This article was generated with AI assistance and reviewed for accuracy and quality.
