Tesla’s Dojo: A Complete Timeline and Breakdown
▼ Summary
– Elon Musk aimed to transform Tesla into an AI company, with a focus on achieving full self-driving capabilities through advanced technology.
– Dojo was Tesla’s custom-built supercomputer designed to train its Full Self-Driving neural networks, intended to process vast amounts of video data from Tesla vehicles.
– Despite years of development and investment, Tesla shifted focus from Dojo to a new supercomputer called Cortex, which uses Nvidia GPUs for AI training.
– Tesla disbanded the Dojo team and shut down the project in 2025, citing resource consolidation around its AI6 chip design for both inference and training.
– Cortex, deployed with roughly 50,000 Nvidia H100 GPUs, enabled advancements in Tesla’s Full Self-Driving software, reflecting a strategic pivot in their AI infrastructure.
Tesla’s vision extends far beyond manufacturing electric vehicles, positioning itself as a leader in artificial intelligence with a primary focus on achieving fully autonomous driving. Central to this ambition was Dojo, a custom supercomputer engineered to train the neural networks powering Tesla’s Full Self-Driving (FSD) system. While FSD currently operates as an advanced driver-assistance feature requiring human oversight, Tesla believed that immense computational power and vast datasets could eventually enable true autonomous operation.
The journey of Dojo began with Elon Musk’s early teases, evolving through years of development before ultimately giving way to a new system named Cortex. What follows is a detailed chronicle of Dojo’s lifespan, from its conceptual origins to its eventual phase-out.
In April 2019, during Tesla’s Autonomy Day event, Musk first hinted at a proprietary supercomputer designed for AI training. He emphasized that all Tesla vehicles produced at the time contained the necessary hardware for full self-driving, pending future software updates. This marked the initial promise of what would become Dojo.
By February 2020, Musk began publicly outlining Dojo’s potential, describing it as a system capable of processing enormous volumes of video data. He projected a launch within a year, though he also noted that Dojo wasn’t strictly essential, it was intended to push Autopilot’s safety far beyond human capabilities.
August 2021 brought Tesla’s first AI Day, where Dojo was formally introduced alongside the D1 chip. Tesla revealed plans to integrate these chips with Nvidia GPUs, aiming to create a powerful training cluster. A technical whitepaper released later that year detailed the floating-point arithmetic standards designed specifically for deep learning applications.
Progress continued throughout 2022. Tesla installed its first Dojo cabinet and demonstrated its capabilities by generating an AI image of a Cybertruck on Mars. The company announced intentions to construct multiple Exapod clusters, targeting completion by early 2023.
The following year, Musk characterized Dojo as a “long-shot bet” with transformative potential. He suggested it could eventually become a commercial service akin to Amazon Web Services. Production officially commenced in mid-2023, with Musk confirming that Dojo was already operational in Tesla data centers. The company projected massive compute expansion, aiming for world-leading capacity by 2024.
Entering 2024, Tesla pursued a dual strategy relying on both Nvidia hardware and Dojo. A $500 million investment was announced for a new Dojo supercomputer in Buffalo, though Musk clarified that Tesla’s spending on Nvidia would still be significantly higher. Technical advancements continued, with TSMC producing next-generation D2 tiles integrating entire systems onto single wafers.
Challenges emerged as Musk acknowledged difficulties in securing sufficient Nvidia GPUs, reinforcing the need to advance Dojo. By mid-2024, Tesla was accelerating its supercomputing efforts, with Musk sharing images of a stainless-steel-clad cluster reminiscent of the Cybertruck’s design.
A shift occurred in August 2024 when Musk introduced Cortex, a new supercomputing cluster based primarily on Nvidia GPUs. Located at Giga Texas, Cortex was designed to handle video training for both FSD and the Optimus humanoid robot. This signaled a strategic pivot away from Dojo’s proprietary architecture.
In early 2025, Tesla’s earnings reports made no mention of Dojo, instead highlighting Cortex’s role in enabling FSD V13. By mid-year, Musk indicated that future development would converge on the AI6 chip, designed for both inference and training. He confirmed the discontinuation of Dojo, noting that resources would focus on a unified chip architecture.
Key members of the Dojo team departed, some founding a new AI chip company. Musk explained the decision, stating that maintaining two separate AI chip designs was inefficient. The Master Plan Part IV, released in September 2025, emphasized “physical AI” but made no reference to Dojo or Cortex, closing the chapter on one of Tesla’s most ambitious technical projects.
(Source: TechCrunch)
