What just happened? Tesla is rebuilding its in-house AI chip program five months after dismantling it. CEO Elon Musk just announced that work on Dojo3 – the company's third-generation AI training supercomputer – has quietly resumed. The relaunch marks a dramatic shift in purpose: the new iteration won't focus on training self-driving models on Earth but instead on powering "space-based AI compute."
The original Dojo initiative was abruptly shut down last year when Tesla's dedicated AI hardware team was disbanded after program lead Peter Bannon departed. About twenty engineers left for DensityAI, a startup founded by former Dojo head Ganesh Venkataramanan and other ex-Tesla technologists.
At that time, Tesla appeared to be backing away from custom silicon development altogether, opting to rely on partners like Nvidia, AMD, and Samsung to fill its compute needs.
Those plans now appear to have changed. Musk indicated on X that Tesla's internal chip roadmap has regained momentum, noting that the company's fifth-generation AI5 chip design is "in good shape."
Manufactured by TSMC, the AI5 was created to power both Tesla's autonomous driving systems and its Optimus humanoid robots. The follow-up AI6 chip, developed under a $16.5 billion partnership with Samsung, is expected to deliver the compute muscle needed for large-scale AI training in Tesla's data centers as well as onboard its vehicles.
AI4 by itself will achieve self-driving safety levels very far above human.
– Elon Musk (@elonmusk) January 18, 2026
AI5 will make the cars almost perfect and greatly enhance Optimus.
AI6 will be for Optimus and data centers.
AI7/Dojo3 will be space-based AI compute.
Dojo3, sometimes referred to by Musk as "AI7," extends that roadmap into orbit. The concept envisions autonomous computing hardware operating outside Earth's atmosphere – harnessing uninterrupted solar power and potentially easing the massive energy demands of terrestrial data centers.
While details remain limited, the reference to "space-based AI compute" suggests a push to integrate Tesla's chip expertise with SpaceX's orbital infrastructure. Space-launched AI nodes could, in theory, support continuous model training or serve as remote data-processing hubs for robotics and autonomous systems.
The technical challenge is immense. Cooling high-density compute hardware in the vacuum of space remains unsolved, and latency, repairability, and radiation shielding present additional hurdles.
Yet Musk's dual control over Tesla and SpaceX gives him considerable leverage: the latter's upcoming Starship system could serve as the deployment vehicle for experimental compute satellites. Axios recently reported that Musk plans to use proceeds from SpaceX's anticipated IPO – potentially valuing the company at $1.5 trillion – to help fund these space-based data center experiments.
Musk's timing coincides with growing competition in AI infrastructure. Nvidia used CES 2026 to unveil Alpamayo, an open-source AI model that mimics human-like reasoning in autonomous vehicles, a direct challenge to Tesla's Full Self-Driving software.
Well that's just exactly what Tesla is doing
– Elon Musk (@elonmusk) January 6, 2026
What they will find is that it's easy to get to 99% and then super hard to solve the long tail of the distribution.
Meanwhile, Tesla is moving to rebuild its Dojo team. Musk invited engineers to apply directly at [email protected], promising they would work on what he described as "the highest-volume chips in the world."