Tesla Dojo: The Rise and Fall of Elon Musk’s AI Supercomputer

Elon Musk has long emphasized the significance of Dojo, Tesla’s AI supercomputer, deemed vital to the company’s AI aspirations. In July 2024, Musk declared that Tesla’s AI team would accelerate its efforts on Dojo before the eagerly awaited robotaxi introduction in October.

However, after six years of anticipation, Tesla revealed last month that they would terminate operations on Dojo and dissolve its team by August 2025. Just weeks after unveiling plans for Dojo 2, the second supercluster built on proprietary D2 chips targeting a 2026 scale, Musk abruptly termed it “an evolutionary dead end.”

Originally, this article aimed to detail Dojo’s role in facilitating Tesla’s ambitions for fully autonomous driving, humanoid robots, semiconductor independence, and beyond. Now, it serves more as a farewell to a project that led many investors and analysts to believe Tesla was evolving from an automaker to an AI innovator.

Dojo was Tesla’s custom supercomputer designed specifically for training its “Full Self-Driving” neural networks.

Enhancing Dojo was crucial to Tesla’s goal of attaining full autonomy and launching a robotaxi service. FSD (Supervised) is the advanced driver-assistance system available in numerous Tesla vehicles today, capable of some automated driving tasks but still requiring driver involvement. It supports the limited robotaxi service that began in Austin in June, using Model Y SUVs.

Although Dojo’s raison d’être was fulfilled, Tesla did not attribute its self-driving progress—which has been contentious—to the supercomputer. Indeed, references to Dojo diminished over the past year. By August 2024, Tesla redirected its focus to Cortex, a new “massive AI training supercluster” being developed at Tesla HQ in Austin, designed for real-world AI applications, with considerable storage for FSD and Optimus training videos.

In Tesla’s Q4 2024 shareholder presentation, updates were shared about Cortex but none regarding Dojo, raising concerns about whether the shutdown would affect Cortex.

Reactions to Dojo’s shutdown have varied. Some perceive it as another example of Musk failing to meet expectations, coinciding with declining EV sales and a sluggish rollout of the robotaxi service. Others contend that the termination reflects a strategic shift from a self-sufficient, high-risk hardware strategy to a leaner approach relying on partnerships for chip production.

The story of Dojo highlights the stakes involved, the project’s shortcomings, and its implications for Tesla’s future direction.

A recap of Dojo’s shutdown

In mid-August 2025, Tesla disbanded the Dojo team and concluded the project. Peter Bannon, the head of Dojo, departed the company along with around 20 employees, who then set out to establish their own AI chip and infrastructure company called DensityAI.

Analysts noted that losing key talent can swiftly impede specialized internal tech initiatives.

The shutdown followed Tesla’s agreement worth $16.5 billion to acquire next-generation AI6 chips from Samsung. The AI6 chip is Tesla’s endeavor 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 clear that our paths converged on AI6, I had to discontinue Dojo and make difficult staffing choices, given that Dojo 2 was now an evolutionary dead end,” Musk stated on X, the social media platform 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 contended that Tesla is not merely an auto manufacturer or a solar technology provider, but rather an AI enterprise that has unveiled the keys to self-driving technology by mimicking human perception.

In contrast to most companies pursuing advancements in autonomous vehicles, which employ a combination of sensors such as lidar, radar, and cameras along with highly precise maps for localization, Tesla aspires to attain fully autonomous driving through pure camera data processed by cutting-edge neural networks to inform rapid driving decisions.

The vision was for the AI software trained on Dojo to be made available to Tesla customers via over-the-air updates. The extensive scale of FSD enabled Tesla to amass millions of miles of video data to refine its FSD capacities, bolstering the belief that accumulating more data would bring them closer to genuine full self-driving abilities.

However, some industry insiders warn that there may be limitations to continually leveraging data to enhance model intelligence.

“Economically, continuing this approach may become too expensive,” noted Anand Raghunathan, a Silicon Valley professor of electrical and computer engineering at Purdue University. He also cautioned, “Experts warn that we might exhaust meaningful data to train our models on. More data doesn’t necessarily translate to more knowledge; it hinges on whether the data possesses useful information for improving the model and if the training process can distill that into a more effective model.”

Despite these concerns, the inclination toward accumulating data appears persistent, at least in the short term. More data necessitates heightened computing power to handle and process everything for training Tesla’s AI models, which was where Dojo was expected to play a key role.

What is a supercomputer?

Dojo was Tesla’s supercomputer system designed as an AI training platform, particularly for FSD. The name pays tribute to training spaces for martial arts.

A supercomputer consists of thousands of smaller interconnected units known as nodes, each featuring its own CPU (central processing unit) and GPU (graphics processing unit). The CPU manages the operations of the node, while the GPU handles complex tasks, breaking them into multiple segments for concurrent processing.

GPUs are vital for machine learning functions, such as training FSD in simulated settings, and they drive expansive language models. This is why the rise of generative AI has significantly elevated Nvidia’s value as the world’s most valuable corporation.

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

Why did Tesla need a supercomputer?

Tesla’s focus on a vision-only strategy required a supercomputer. The FSD neural networks need extensive driving data to recognize and classify surrounding objects for decision-making. When FSD is active, these neural networks must continually gather and analyze visual data at speeds comparable to human recognition abilities.

Essentially, Tesla seeks to create a digital counterpart of the human visual cortex and cognitive function.

Accomplishing this goal necessitates the storage and processing of vast amounts of video data collected globally, along with executing millions of simulations to effectively train its models.

Tesla primarily relied on Nvidia to power its existing Dojo computer system but aimed to avoid over-dependence—especially given Nvidia’s high-cost chips. Tesla envisioned developing superior technology to enhance bandwidth and reduce latency, leading to the creation of custom hardware programs intended to train AI models more efficiently than traditional configurations.

At the core of this initiative were Tesla’s proprietary D1 chips, specifically designed for AI workloads.

Tell me more about these chips

Tesla, similar to Apple, believes that hardware and software should work in harmony. Accordingly, Tesla aimed to transition from traditional GPU hardware to design its own chips for Dojo.

During AI Day 2021, Tesla unveiled the D1 chip, roughly the size of a palm. The D1 chip began production in July 2023.

Manufactured by Taiwan Semiconductor Manufacturing Company (TSMC) using 7-nanometer semiconductor nodes, the D1 chip features 50 billion transistors and a die size of 645 mm², according to Tesla, indicating it was built to be powerful and efficient, able to swiftly manage complex tasks.

Nevertheless, the D1 does not surpass Nvidia’s A100 chip in performance.

Tesla was in the midst of developing the next-gen D2 chip, designed to eliminate information flow bottlenecks. Rather than linking individual chips, the D2 aimed to integrate the whole Dojo tile onto a single silicon wafer.

Tesla has not disclosed how many D1 chips were ordered or received. They also did not provide a timeline for deploying Dojo supercomputers utilizing D1 chips.

What did Dojo mean for Tesla?

Tesla envisioned that by managing its chip production, it could potentially scale up AI training operations quickly and cost-effectively.

This move also diminished future reliance on Nvidia chips, which have become progressively expensive and harder to obtain. Tesla is now concentrating on partnerships with Nvidia, AMD, and Samsung, the latter slated to manufacture its next-gen AI6 chip.

During Tesla’s second-quarter 2024 earnings call, Musk noted that 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 required,” prompting an intensified focus on Dojo’s training capabilities.

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

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

Tesla potentially faced hurdles since much of the existing AI software is tailored to operate on GPUs. Adapting Dojo chips for general-purpose AI model training would require extensive software adjustments.

Alternatively, Tesla could explore a model similar to AWS and Azure, renting out its computing capabilities—an idea that fascinated analysts. A September 2023 Morgan Stanley report estimated that Dojo could contribute $500 billion to Tesla’s market value by unlocking new revenue streams through robotaxis and software services.

In summary, Dojo chips served as a backup for the automaker, although they possessed the potential for significant returns.

How far did Tesla Dojo get?

Throughout the journey, Musk frequently provided updates, yet many of Dojo’s projected milestones were not realized.

For instance, Musk asserted in June 2023 that Dojo had been operational for several months with significant contributions. At the same time, Tesla projected that Dojo would rank among the top five most powerful supercomputers by February 2024, aiming for total processing power to hit 100 exaflops by October 2024—relying on around 276,000 D1 chips or approximately 320,500 Nvidia A100 GPUs.

However, Tesla never provided any indication or information that suggested these ambitious targets were achieved.

Numerous commitments were made by Tesla and Musk regarding Dojo, including financial ones. For example, in January 2024, the company promised $500 million to build a Dojo supercomputer at its Buffalo gigafactory, with $314 million already spent as of 2024.

Shortly after Tesla’s Q2 2024 earnings call, Musk posted photos of Dojo 1 on X, asserting that it would reach “around 8,000 H100-equivalent training capabilities by the end of the year. Moderate scale, yet not insignificant either.”

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

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

In Q2 2025, Tesla reported an extension of its AI training capacity with an additional 16,000 H200 GPUs at Gigafactory Texas, culminating in a total of 67,000 H100 equivalents for Cortex. During the same earnings call, Musk indicated that a second Dojo cluster was expected to be operational “at scale” by 2026, while also hinting at possible redundancies.

“Considering Dojo 3 and the AI6 inference chip, it seems clear we want to find a convergence point—basically the same chip,” Musk remarked.

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

In late August 2025, TechCrunch confirmed that although Tesla will still allocate $500 million for a supercomputer in Buffalo, it will not be Dojo.

This story was initially published on August 3, 2024, and last updated on September 2, 2025, regarding Tesla’s decision to shut down Dojo.