Tesla Disbands Dojo Supercomputer Team, Shifts AI Strategy
In a significant strategic pivot, Tesla has officially disbanded the team behind its ambitious Dojo supercomputer, signaling an end to its in-house chip development program dedicated to autonomous driving. This move marks a substantial departure from a vision once touted as central to the automaker’s AI future, opting instead for a deeper reliance on external partnerships. The dissolution of the Dojo team also sees the departure of its lead, Peter Bannon, with remaining members being reassigned to other compute and data center projects within the company.
Dojo, first unveiled in 2021 and entering production in July 2023, was Tesla’s bespoke supercomputer designed to train the massive deep neural networks required for its Full Self-Driving (FSD) technology. Unlike traditional systems that might use LiDAR or radar, Tesla’s approach is vision-based, and Dojo was purpose-built to process petabytes of real-world video data collected from millions of Tesla vehicles, aiming to enhance the performance and safety of its autonomous systems. The goal was to create a highly optimized, efficient system that could accelerate AI training far beyond what general-purpose GPUs like Nvidia’s could offer. Tesla had committed over $1 billion to the Dojo project by the end of 2024.
However, the path for Dojo has been fraught with challenges. The project faced repeated delays and saw several key leadership changes, including the earlier departure of Ganesh Venkataramanan, who was instrumental in designing Dojo’s custom D1 chip. Elon Musk himself had previously acknowledged the formidable nature of the endeavor, describing Dojo as a “long shot” in January 2024, despite its potentially high payoff.
The shift in strategy is largely driven by a pragmatic decision to streamline resources. As Musk articulated on social media, “It doesn’t make sense for Tesla to divide its resources and scale two quite different AI chip designs.” Instead, Tesla will now concentrate its internal hardware efforts on its AI5 and AI6 inference chips, designed for real-time processing directly within vehicles and robots like Optimus, which Musk believes will be “excellent for inference and at least pretty good for training.” This focus on in-car computing is deemed critical for achieving robust FSD functionalities by reducing latency and improving practicality.
To compensate for the cessation of in-house training chip development, Tesla is significantly increasing its reliance on established external technology partners. The company has forged a substantial $16.5 billion deal with Samsung Electronics to secure AI semiconductors through 2033, with future AI6 chips expected to be manufactured at Samsung’s new plant in Taylor, Texas. Additionally, Tesla will lean on Nvidia and Advanced Micro Devices (AMD) for its broader computing needs, a move that allows Tesla to leverage advanced technologies from industry leaders without shouldering the full research and development burden itself.
Adding another layer to this evolving landscape, approximately 20 former Dojo engineers, including Ganesh Venkataramanan, have spun off to form a new startup named DensityAI. This new venture aims to build advanced AI chips, hardware, and software specifically for the automotive industry, targeting the very market Tesla once sought to dominate with Dojo. DensityAI’s emergence highlights the competitive nature of the AI hardware market and the ongoing talent mobility within the sector.
While the disbandment of the Dojo team represents a significant strategic realignment for Tesla, it underscores a broader industry trend where even well-resourced companies opt for collaboration and specialization in the capital-intensive and rapidly evolving field of AI hardware. This pivot could accelerate Tesla’s path to scaling its AI capabilities by tapping into external advancements, potentially allowing the company to maintain its significant presence in the AI-assisted automotive sector.