Tesla Dojo: The Ascent and Decline of Elon Musk’s AI Supercomputer

Elon Musk has consistently underscored the importance of Dojo, Tesla’s AI supercomputer, which is considered crucial for the company’s AI ambitions. In July 2024, Musk announced that Tesla’s AI team would ramp up its efforts on Dojo ahead of the highly anticipated robotaxi launch scheduled for October.

Yet, after six years of expectation, Tesla disclosed last month that it would cease operations on Dojo and disband its team by August 2025. Just weeks after unveiling plans for Dojo 2, a second supercluster utilizing proprietary D2 chips aimed at 2026, Musk abruptly dismissed the project as “an evolutionary dead end.”

Originally, this article sought to elaborate on Dojo’s role in advancing Tesla’s goals for fully autonomous driving, humanoid robots, semiconductor independence, and more. It now serves more as a farewell to a venture that led many investors and analysts to believe Tesla was transforming from an automaker into an AI trailblazer.

Dojo was Tesla’s customized supercomputer specifically designed to train its “Full Self-Driving” neural networks.

Enhancing Dojo was vital for Tesla’s objective of achieving full autonomy and launching a robotaxi service. FSD (Supervised) represents the advanced driver-assistance system present in many Tesla vehicles today, capable of performing some driving tasks autonomously but still requiring driver oversight. It facilitates the limited robotaxi service that launched in Austin in June, utilizing Model Y SUVs.

Although Dojo’s raison d’être was achieved, Tesla has not credited the supercomputer for its self-driving advancements—which have been debatable. Indeed, references to Dojo have dwindled over the past year. By August 2024, Tesla shifted its focus to Cortex, a new “massive AI training supercluster” being developed at Tesla HQ in Austin, intended for real-world AI applications, with ample storage for FSD and Optimus training videos.

In Tesla’s Q4 2024 shareholder presentation, updates about Cortex were shared, but none pertaining to Dojo, raising concerns about the implications of the shutdown for Cortex.

Responses to Dojo’s shutdown have varied. Some view it as another instance of Musk not meeting expectations, coinciding with dwindling EV sales and a sluggish rollout of the robotaxi service. Others argue that this termination signifies a strategic shift from a self-sustaining, high-risk hardware approach to a leaner model that relies on partnerships for chip production.

The narrative of Dojo underscores both the involved risks and the project’s limitations, as well as its implications for Tesla’s future trajectory.

A recap of Dojo’s shutdown

In mid-August 2025, Tesla dismantled the Dojo team and wrapped up the project. Peter Bannon, who led Dojo, left the company along with around 20 employees, who subsequently embarked on establishing their own AI chip and infrastructure venture named DensityAI.

Analysts observed that the loss of key talent can swiftly hinder specialized internal tech initiatives.

The shutdown followed Tesla’s $16.5 billion agreement to obtain next-gen AI6 chips from Samsung. The AI6 chip signifies Tesla’s ambitions for chip designs capable of scaling to power FSD and Tesla’s Optimus humanoid robots, as well as for high-performance AI training in data centers.

“Once it became evident that our paths converged on AI6, I had to discontinue Dojo and make tough staffing decisions, considering that Dojo 2 had become an evolutionary dead end,” Musk stated on X, the social media network he owns. “Dojo 3 arguably continues in the form of several AI6 [systems-on-a-chip] on a single board.”

Tesla’s Dojo backstory

Musk argued that Tesla is not just an automaker or solar technology provider but rather an AI enterprise that has unlocked the secrets to self-driving technology by emulating human perception.

Unlike most companies pursuing innovations in autonomous vehicles—utilizing a mix of sensors like lidar, radar, and cameras, along with highly accurate maps for localization—Tesla aims to achieve fully autonomous driving through solely camera data processed by advanced neural networks to guide swift driving decisions.

The vision entailed making the AI software trained on Dojo available to Tesla customers through over-the-air updates. The extensive scale of FSD allowed Tesla to gather millions of miles of video data to enhance its FSD capabilities, strengthening the belief that amassing more data would bring them closer to genuine full self-driving capabilities.

However, some industry insiders caution that there may be limitations in perpetually leveraging data to boost model intelligence.

“Economically, maintaining this approach could become overly costly,” remarked Anand Raghunathan, a Silicon Valley professor of electrical and computer engineering at Purdue University. He cautioned, “Experts warn that we might deplete meaningful data to train our models on. More data does not automatically equate to more knowledge; it depends on whether the data contains valuable information for enhancing the model and if the training process can distill that into a more effective model.”

Despite these apprehensions, the inclination to accumulate data seems steadfast, at least in the short term. More data necessitates increased computing power to manage and process everything for training Tesla’s AI models, where Dojo was expected to play an essential role.

What is a supercomputer?

Dojo was Tesla’s supercomputer framework designed as an AI training platform, specifically for FSD. The name reflects spaces for training in martial arts.

A supercomputer comprises thousands of smaller interconnected units called nodes, each featuring its own CPU (central processing unit) and GPU (graphics processing unit). The CPU controls the node’s operations, while the GPU handles intricate tasks, dividing them into multiple segments for simultaneous processing.

GPUs are crucial for machine learning tasks, such as training FSD in simulated environments, and they propel extensive language models. This is why the emergence of generative AI has significantly increased Nvidia’s valuation, making it the world’s most valuable company.

Tesla also employed Nvidia GPUs for AI training (more on that later).

Why did Tesla need a supercomputer?

Tesla’s commitment to a vision-only strategy demanded a supercomputer. The FSD neural networks require vast driving data to identify and classify surrounding objects for decision-making. Once active, these neural networks must continuously collect and evaluate visual data at speeds comparable to human recognition.

Fundamentally, Tesla aims to create a digital equivalent of the human visual cortex and cognitive functionality.

To achieve this, extensive storage and processing of substantial quantities of global video data is essential, coupled with executing millions of simulations to train its models effectively.

Tesla primarily relied on Nvidia to power its existing Dojo computer system but sought to mitigate over-reliance—especially given Nvidia’s costly chips. The company envisioned developing superior technology to enhance bandwidth and reduce latency, leading to the creation of tailored hardware programs meant to train AI models more efficiently than conventional setups.

At the heart of this effort were Tesla’s proprietary D1 chips, designed specifically for AI tasks.

Tell me more about these chips

Analogous to Apple, Tesla believes in the synergy of hardware and software. Consequently, Tesla aimed to transition from conventional GPU hardware to designing its own chips for Dojo.

During AI Day 2021, Tesla introduced the D1 chip, roughly palm-sized, which commenced production in July 2023.

Manufactured by Taiwan Semiconductor Manufacturing Company (TSMC) using 7-nanometer semiconductor technology, the D1 chip boasts 50 billion transistors and a die size of 645 mm², indicating it was engineered to be both powerful and efficient, adept at swiftly handling complex tasks.

However, the D1 does not outperform Nvidia’s A100 chip.

Tesla was in the process of refining the next-gen D2 chip, which aimed to eliminate bottlenecks in information flow. Rather than connecting individual chips, the D2 sought to integrate the entire Dojo tile onto a single silicon wafer.

Tesla has not disclosed how many D1 chips were ordered or delivered, nor has it provided a timeline for the deployment of Dojo supercomputers utilizing D1 chips.

What did Dojo mean for Tesla?

Tesla anticipated that by controlling its chip production, it could rapidly and cost-effectively scale AI training operations.

This strategy also reduced future reliance on Nvidia chips, which are becoming increasingly expensive and challenging to procure. Tesla is now focusing on collaborations with Nvidia, AMD, and Samsung, which is set to manufacture its next-gen AI6 chip.

During Tesla’s Q2 2024 earnings call, Musk indicated that the demand for Nvidia’s hardware was “so high that acquiring GPUs consistently has been a challenge.” He expressed “serious concerns about ensuring steady supplies of GPUs when needed,” leading to an intensified focus on Dojo’s training capabilities.

Dojo represented a high-stakes gamble that Musk recognized could fail.

In the long run, Tesla envisioned a business model deriving from its AI division. Musk indicated during the Q2 2024 earnings call that he anticipated “a pathway to competing with Nvidia through Dojo.” Although the D1 was primarily intended for tasks like computer vision and training for FSD and Optimus, Musk hinted that future iterations would need to accommodate general-purpose AI training.

Tesla may face obstacles as much of the current AI software is optimized to function on GPUs. Adapting Dojo chips for general-purpose AI model training might require substantial software modifications.

Alternatively, Tesla could consider a model similar to AWS and Azure, renting out its computing resources—an idea that intrigued analysts. A Morgan Stanley report from September 2023 estimated that Dojo could add $500 billion to Tesla’s market valuation by creating new revenue avenues via robotaxis and software services.

In conclusion, while Dojo chips served as a contingency for the automaker, they held potential for considerable returns.

How far did Tesla Dojo get?

Throughout its development, Musk frequently shared updates, but many of Dojo’s anticipated milestones went unfulfilled.

For instance, Musk proclaimed in June 2023 that Dojo had been operational for several months, contributing significantly. At the same time, Tesla anticipated that Dojo would rank among the top five most powerful supercomputers by February 2024, with a target processing power of 100 exaflops by October 2024—dependent on about 276,000 D1 chips or approximately 320,500 Nvidia A100 GPUs.

However, Tesla never offered any indications that supported these ambitious objectives being met.

Numerous pledges were made by Tesla and Musk regarding Dojo, including financial commitments. For example, in January 2024, the company allocated $500 million for the construction of a Dojo supercomputer at its Buffalo gigafactory, of which $314 million had already been spent as of 2024.

Soon after Tesla’s Q2 2024 earnings call, Musk uploaded images of Dojo 1 on X, claiming that it would achieve “around 8,000 H100-equivalent training capabilities by the year’s end. Moderate scale, though not insignificant either.”

Despite all these efforts—especially Musk’s communications on X and during earnings calls—discussions around Dojo abruptly halted in August 2024, shifting focus to Cortex.

In the Q4 2024 earnings call, Tesla announced the deployment of Cortex, “a ~50k H100 training cluster at Gigafactory Texas,” contributing to version 13 of supervised FSD.

In Q2 2025, Tesla reported an expansion of its AI training capacity with an additional 16,000 H200 GPUs at Gigafactory Texas, totaling 67,000 H100 equivalents for Cortex. During the same earnings call, Musk suggested that a second Dojo cluster might be operational “at scale” by 2026, while hinting at potential redundancies.

“Regarding Dojo 3 and the AI6 inference chip, it seems evident we wish to pinpoint a convergence—a shared chip,” Musk remarked.

Weeks later, he reversed his stance and ordered the disbandment of the Dojo team.

In late August 2025, TechCrunch confirmed that while Tesla will still dedicate $500 million for a supercomputer in Buffalo, it won’t be Dojo.

This article was first published on August 3, 2024, and last updated on September 2, 2025, regarding Tesla’s decision to discontinue Dojo.