Tesla Dojo’s Ambitious Journey: Unraveling Elon Musk’s AI Supercomputer Vision

BitcoinWorld

Tesla Dojo’s Ambitious Journey: Unraveling Elon Musk’s AI Supercomputer Vision

In the rapidly evolving world of artificial intelligence and autonomous vehicles, Tesla has consistently pushed the boundaries, aiming to be more than just an automaker. Elon Musk’s audacious vision positioned Tesla as a leading AI company, with its sights set on perfecting self-driving technology. Central to this monumental ambition was Tesla Dojo, a custom-built supercomputer designed to revolutionize the training of its Full Self-Driving (FSD) neural networks. While FSD (Supervised) currently handles many automated driving tasks, it still requires a human driver’s vigilance. Tesla’s belief was simple: more data, more compute power, and more rigorous training could bridge the gap from ‘almost’ to ‘fully’ self-driving. This article chronicles the fascinating, often tumultuous, journey of Dojo, from its initial teasers to its eventual strategic pivot, and the emergence of Cortex.

The Genesis of Tesla Dojo: A Vision Unveiled (2019-2020)

The story of Tesla Dojo began to unfold in 2019, when Elon Musk first teased the concept during Tesla’s Autonomy Day. The company’s AI team took center stage, discussing the intricate AI powering Autopilot and Full Self-Driving. It was here that Musk revealed plans for a custom-built supercomputer specifically for training AI, hinting at Dojo’s crucial role. He also made the bold claim that all Tesla cars then in production possessed the necessary hardware for full self-driving, needing only a software update to unlock their full potential.

The following year, 2020, saw Musk ramp up discussions, effectively launching the ‘Dojo roadshow’. In February, he emphasized Tesla’s growing fleet of over a million connected vehicles, all equipped with sensors and compute power for self-driving. He touted Dojo’s anticipated capabilities:

• Processing vast amounts of video training data.
• Efficiently running hyperspace arrays with numerous parameters.
• Featuring ample memory and ultra-high bandwidth between cores.

By August 2020, Musk reiterated the plan to develop this neural network training computer, calling it “a beast” and projecting its first version “about a year away.” Interestingly, by December, he clarified that while Dojo wasn’t strictly necessary, it would significantly enhance self-driving capabilities, stating, “It isn’t enough to be safer than human drivers, Autopilot ultimately needs to be more than 10 times safer than human drivers.”

Making it Official: Tesla’s AI Day and the D1 Chip (2021)

2021 marked a significant milestone with Tesla Dojo officially taking center stage. At Tesla’s inaugural AI Day on August 19, the automaker made a concerted effort to attract top engineering talent to its burgeoning AI team. During this event, Tesla formally announced Dojo and introduced its custom-designed D1 chip. This chip was slated to power the Dojo supercomputer, working in conjunction with Nvidia’s GPUs. The ambitious plan included an AI cluster housing an impressive 3,000 D1 chips.

Further solidifying its commitment, Tesla released a Dojo Technology whitepaper on October 12. Titled “a guide to Tesla’s configurable floating point formats & arithmetic,” this technical document outlined a new standard for binary floating-point arithmetic. This innovation was specifically tailored for deep learning neural networks, designed for implementation “entirely in software, entirely in hardware, or in any combination of software and hardware,” showcasing Tesla’s deep dive into specialized hardware and software for AI acceleration.

Progress and Promises: The “Long-Shot Bet” (2022-2023)

As 2022 unfolded, Tesla continued to reveal progress on Tesla Dojo. In August, Musk stated that Tesla would “phase in Dojo. Won’t need to buy as many incremental GPUs next year,” signaling an increasing reliance on their in-house solution. At the second AI Day on September 30, the company showcased tangible advancements, announcing the installation of the first Dojo cabinet and successful load testing at 2.2 megawatts. Tesla also detailed its modular approach, building one tile per day, with each tile comprising 25 D1 chips. A memorable demonstration involved Dojo running a Stable Diffusion model to generate an AI image of a “Cybertruck on Mars,” highlighting its practical application.

The company set an ambitious target: a full Exapod cluster by Q1 2023, with plans for seven Exapods in Palo Alto. By April 2023, Musk framed Dojo as a “long-shot bet” with the “potential for an order of magnitude improvement in the cost of training.” He even hinted at Dojo becoming a “sellable service” akin to Amazon Web Services. Throughout 2023, Tesla continued to project aggressive growth in its compute power, claiming Dojo was already online and running tasks by June, and predicting Tesla’s compute would be among the top five globally by early 2024, aiming for 100 exaflops by October 2024. In July, Tesla confirmed Dojo production had begun, committing over $1 billion to the project through 2024, underscoring its strategic importance for FSD training.

Scaling Challenges and the Nvidia GPUs Factor (2024)

The year 2024 brought both ambitious scaling plans and significant challenges, particularly concerning Nvidia GPUs. In January, Musk reiterated Dojo as a high-risk, high-reward endeavor, confirming a “dual path of Nvidia and Dojo.” He stated Dojo was “working” and “doing training jobs,” with plans for future iterations like Dojo 1.5, 2, and 3. Tesla announced a $500 million investment to build a Dojo supercomputer in Buffalo, though Musk quickly contextualized this, noting it was “only equivalent to a 10k H100 system from Nvidia” and that Tesla would spend “more than that on Nvidia hardware this year.” This statement starkly highlighted the immense capital required to compete in the AI supercomputing race.

By April, TSMC confirmed that Dojo’s next-generation training tile, the D2 – which integrates the entire tile onto a single silicon wafer – was already in production. However, a major pivot emerged in June when a CNBC report alleged Musk diverted thousands of Nvidia chips meant for Tesla to X and xAI. Musk initially denied the report but later clarified that Tesla lacked a suitable location to deploy the chips due to ongoing construction at Giga Texas, stating they “would have just sat in a warehouse.” He then detailed Tesla’s substantial spending on AI, estimating “$3B to $4B” for Nvidia purchases that year, confirming the critical role of Nvidia GPUs in their immediate AI infrastructure.

Further complications arose in July when Musk hinted that current Tesla vehicles might lack the necessary hardware for the company’s next-gen AI model due to a “5x increase in parameter count.” During Tesla’s Q2 earnings call, Musk openly acknowledged the high demand for Nvidia hardware, stating it was “often difficult to get the GPUs.” This supply constraint further intensified Tesla’s focus on Dojo, with Musk stating, “I think this therefore requires that we put a lot more effort on Dojo in order to ensure that we’ve got the training capability that we need.” Tesla’s investor deck projected a ramp-up to 90,000 H100 equivalent GPUs by the end of 2024, with Dojo 1 expected to contribute “roughly 8k H100-equivalent of training online by end of year.”

The Pivotal Shift: From Dojo to Cortex and Beyond (2024-2025)

The narrative around Tesla Dojo took a dramatic turn in late 2024 and early 2025. By July 30, Musk began discussing AI5, Tesla’s next-gen inference chip, signaling a forward-looking shift. Just days later, on August 3, he revealed a walkthrough of “the Tesla supercompute cluster at Giga Texas (aka Cortex),” which he described as being made of “roughly 100,000 H100/H200 Nvidia GPUs with massive storage for video training of FSD & Optimus.” This marked the official introduction of Cortex as Tesla’s primary AI training supercluster.

The Q4 and full-year 2024 earnings call in January 2025 notably omitted any mention of Dojo. Instead, Cortex dominated the discussion, with Tesla confirming the deployment of Cortex, comprising approximately 50,000 H100 Nvidia GPUs. The shareholder deck highlighted Cortex’s role in enabling FSD (Supervised) V13, which boasted “major improvements in safety and comfort thanks to 4.2x increase in data, higher resolution video inputs … among other enhancements.” CFO Vaibhav Taneja confirmed that Tesla accelerated Cortex’s buildout to expedite FSD V13, with accumulated AI-related capital expenditures reaching “approximately $5 billion.”

The formal end of the Dojo project came swiftly. In July 2025, during Tesla’s Q2 earnings call, Musk still spoke of Dojo 2 operating at scale by 2026, but also hinted at “convergences” with future inference chips. Shortly after, Tesla signed a massive $16.5 billion deal with Samsung for its next-generation AI6 chips, a bet on a unified chip design for FSD, Optimus, and high-performance AI training. The final blow came in August: reports emerged of Dojo workers leaving to form a new company, DensityAI, followed by Bloomberg’s report that Tesla had disbanded its Dojo team and shut down the project, with lead Peter Bannon departing. Musk confirmed this on X, stating, “It doesn’t make sense for Tesla to divide its resources and scale two quite different AI chip designs. The Tesla AI5, AI6 and subsequent chips will be excellent for inference and at least pretty good for training. All effort is focused on that.” He further clarified, “Once it became clear that all paths converged to AI6, I had to shut down Dojo and make some tough personnel choices, as Dojo 2 was now an evolutionary dead end.” By September 2025, Tesla’s Master Plan Part IV made no mention of Dojo or Cortex, focusing instead on “physical AI.”

What Does This Mean for FSD Training?

The strategic shift from Dojo to Cortex and the focus on AI6 chips fundamentally redefines Tesla’s approach to FSD training. While Dojo was an ambitious in-house effort to control the entire AI stack, the pivot to Cortex – largely powered by Nvidia GPUs – and the development of the versatile AI6 chip suggest a pragmatic approach. This new strategy prioritizes immediate scalability and leverages proven, high-performance hardware while simultaneously developing a custom chip that can handle both inference (running AI in the car) and training (developing AI in data centers).

The successful deployment of Cortex and its role in enabling FSD V13 demonstrates that Tesla’s AI ambitions are far from diminished. Instead, the company has adapted its strategy to overcome hardware supply chain challenges and optimize its resource allocation. The goal remains the same: achieve truly autonomous driving, but the path to get there has evolved, embracing a hybrid model that combines external expertise with targeted internal chip development.

Elon Musk’s AI Vision: A Constant Evolution

Elon Musk’s AI vision for Tesla has always been characterized by audacious goals and a willingness to pivot when necessary. The Dojo saga is a testament to this dynamic leadership. His relentless pursuit of full self-driving and general-purpose AI, as evidenced by the Optimus robot, drives Tesla’s technological advancements. While Dojo represented a bold attempt to vertically integrate and innovate at the silicon level, the immense complexity and cost of building a competitive AI supercomputer from scratch, coupled with the rapid advancements in commercial off-the-shelf solutions like Nvidia’s, necessitated a re-evaluation.

Musk’s decision to consolidate resources around the AI6 chip and the Cortex supercluster reflects a strategic maturity. It acknowledges that even for a company with Tesla’s resources, focusing on core competencies and leveraging existing industry strengths can be more efficient. The ultimate aim – to solve “real-world AI” – remains unchanged, but the means to achieve it are subject to continuous optimization and adaptation.

The Future of Tesla’s AI Supercomputer Ambitions

With the shutdown of Tesla Dojo and the rise of Cortex and the AI6 chip, Tesla’s future in AI supercomputer development appears to be on a more streamlined, yet equally ambitious, trajectory. The $16.5 billion deal with Samsung for the AI6 chips signifies a long-term commitment to a unified, scalable hardware platform. This approach aims to reduce the fragmentation of resources and accelerate development across FSD, Optimus, and other AI initiatives.

The lessons from Dojo’s journey are invaluable: building a world-class AI supercomputer requires not just immense capital and engineering talent, but also strategic flexibility in a rapidly changing technological landscape. Tesla’s pivot demonstrates an understanding that innovation sometimes means adapting your internal projects to align with external realities and proven technologies. The focus on “physical AI” in Master Plan Part IV underscores Tesla’s commitment to bringing AI out of the data center and into the real world, powered by sophisticated training infrastructure like Cortex and advanced inference chips like AI6.

The journey of Tesla’s AI supercomputing ambitions, from the promising inception of Dojo to the pragmatic pivot to Cortex and the AI6 chip, illustrates the dynamic and challenging nature of cutting-edge AI development. While Dojo’s dedicated team was disbanded, its spirit of innovation undoubtedly contributed to the advancements now powering Tesla’s FSD and broader AI endeavors. Tesla continues its relentless pursuit of artificial intelligence, learning, adapting, and pushing the boundaries of what’s possible in autonomous technology and beyond. The strategic evolution underscores that in the high-stakes game of AI, adaptability is as crucial as ambition.

To learn more about the latest AI market trends, explore our article on key developments shaping AI models features, institutional adoption, etc.

This post Tesla Dojo’s Ambitious Journey: Unraveling Elon Musk’s AI Supercomputer Vision first appeared on BitcoinWorld and is written by Editorial Team