Tesla has reportedly disbanded its Dojo supercomputer team, marking a major retreat from its years-long effort to develop in-house AI training infrastructure. The move signals a deeper reliance on external partners like Nvidia, AMD, and Samsung as Tesla pivots away from its custom chip ambitions.
Peter Bannon, who led the Dojo initiative, is departing the company as part of the transition, according to a report from Bloomberg. Remaining members of the team are being reassigned to other data center and computing projects within Tesla, according to individuals familiar with the matter.
The decision follows the recent exit of about 20 Dojo engineers who left to launch a new AI venture, DensityAI, which aims to develop chips and hardware solutions for data centers and AI applications in robotics and autonomous systems.
Tesla’s Dojo project was first introduced at the company’s inaugural AI Day in 2021. Dojo was designed to rapidly train machine-learning models using vast volumes of video data collected by Tesla vehicles. At its peak, the project featured Tesla’s custom D1 chip, with a follow-up D2 version under development to improve performance and address bottlenecks.
But enthusiasm began to wane in mid-2024, when CEO Elon Musk began emphasizing a different internal effort—Cortex, a new AI training supercluster under construction at Tesla’s Austin headquarters. During Tesla’s most recent earnings call, Musk hinted that the company may consolidate its AI chip development with partner technologies, stating a desire to align future hardware like the Dojo 3 and AI6 inference chip into a unified architecture.
This strategic shift comes just weeks after Tesla secured a $16.5 billion semiconductor deal with Samsung Electronics to manufacture its AI6 chips through 2033. These chips are expected to support a wide range of applications, from Tesla’s Full Self-Driving (FSD) platform to its Optimus humanoid robots and large-scale AI training.
