Elon Musk announced on Sunday that Tesla is restarting work on Dojo3, the company's third-generation supercomputer project. This comes after the team behind it was disbanded last year. Now, the effort will focus on space-based AI computing rather than training self-driving car models on Earth.
Background
Tesla first started the Dojo project to build its own supercomputers for handling massive amounts of video data from its cars. The goal was to train AI systems for self-driving features. Dojo1 launched with custom chips designed to process that video quickly. But progress hit roadblocks. Competition from companies like Nvidia made it hard for Dojo to keep up in raw power.
Last year, Tesla paused the project. The Dojo team broke up after key leaders left. Peter Bannon, who led Dojo, departed. Then Ganesh Venkataramanan, another top figure, started a new company called DensityAI with about 20 former Tesla workers. Tesla shifted resources to its vehicle AI chips and leaned more on partners like Nvidia for computing power. Bloomberg reported at the time that Tesla planned to buy more from Nvidia and AMD, and use Samsung for manufacturing.
Musk explained the pause by saying it did not make sense to split focus across too many chip types. He noted that clustering lots of the company's AI5 or AI6 chips could do the same job as a full Dojo system, while cutting down on wiring and costs.
Key Details
The restart ties directly to Tesla's progress on its own chips. Musk said the AI5 chip design is now in good shape. This chip, made by TSMC, powers self-driving features in cars and the Optimus robot. It matches the performance of Nvidia's Hopper chip in one unit, or Blackwell level with two units together, all while using less power at about 250 watts per chip.
Tesla signed a $16.5 billion deal with Samsung last summer for AI6 chips. Those will go into vehicles, Optimus, and data centers. Musk laid out a fast pace: new chips every nine months. AI6 comes next for Optimus and data centers. Then AI7, paired with Dojo3, heads to space.
“Now that the AI5 chip design is in good shape, Tesla will restart work on Dojo3… AI7/Dojo3 will be for space-based AI compute.” – Elon Musk
Musk made the announcement in a post on X. He also put out a call for engineers. People interested should email AI_Chips@Tesla.com with three bullet points on their hardest technical challenges solved. He called these the highest volume chips in the world.
Dojo3 marks a change. Past versions mixed Tesla chips with Nvidia GPUs. This one aims to use only in-house hardware. Tesla had canceled the wafer-level processor part of Dojo late last year. Now, with AI5 stable, resources shift back.
Team and Timing
Building the team starts fresh. Musk wants to rebuild after the departures. The timing lines up with other events. At CES 2026, Nvidia showed Alpamayo, an open AI model for self-driving that challenges Tesla's system. Musk said solving rare driving cases is very hard but hoped they succeed.
What This Means
Space-based AI compute means putting data centers in orbit. Musk and others see big upsides. Space offers constant sunlight for solar power, no night cycles. Cold temperatures cut cooling needs. Earth's power grids strain under current data centers. Orbital ones could run 24/7 without those limits.
Musk has an advantage with SpaceX. He controls launch rockets like Starship. Reports say he plans to use a future SpaceX IPO to fund compute satellites. These would stay in sunlight, pulling solar energy nonstop. Sam Altman from OpenAI shares the excitement for orbital data centers.
Challenges remain large. Cooling powerful chips in space vacuum is tough. No air means heat does not escape easily. Radiation and launch costs add hurdles. Experts question if it works at scale. Musk's ideas often start bold, then push forward with heavy effort.
For Tesla, this fits a pattern of linking cars, robots, and bigger AI goals. Dojo3 could train models beyond driving, maybe for Optimus or other uses. Success would cut reliance on Nvidia, saving money as chip needs grow. Tesla's chip roadmap speeds up to match AI demands.
The project could change supercomputing. If space proves better, others might follow. Power limits on Earth push innovators off-planet. Tesla's move tests if private companies can make it real. Watch for team hires and first tests. Musk's nine-month cycle means updates soon.
