Tesla reportedly pulls the plug on its ambitious Dojo supercomputer

In a huge strategy shift for a company that prides itself on doing things its own way, Tesla is reportedly shutting down its ambitious supercomputer project. According to reports, the automaker is winding down the Dojo program, reassigning its team, and switching to rely more heavily on established technology giants like Nvidia and AMD. As usual, Tesla has not made an official announcement, but the end of Dojo signals a major turnaround in its approach to developing the powerful artificial intelligence needed for its electric cars.

The Dojo project, officially launched in 2021, was a classic Tesla moonshot. Instead of buying all its computing hardware from outside suppliers, the company set out to design and build its own supercomputer from the ground up. The goal was to create a machine perfectly tailored to one task: training the AI that powers its Full Self-Driving (FSD) software and its Optimus humanoid robot.

CEO Elon Musk himself described the venture as a "long shot" but one "worth taking because the payoff is potentially very high." For years, Tesla has walked a "dual path," using both its custom Dojo hardware and high-performance GPUs from Nvidia to train its systems.

Building a world-class supercomputer from scratch is an immense challenge, possibly too much of a challenge even for Tesla. Reports suggest the Dojo program ran into serious roadblocks. The custom "wafer-scale" processors, which were meant to pack immense power onto a single, massive chip, were just too difficult to produce and had limitations with memory capacity.

The project's rollout was also slow. The company had projected its next-generation Dojo 2 cluster would be equivalent to 100,000 of Nvidia's powerful H100 GPUs by 2026 - a milestone that xAI is expected to hit earlier. Essentially, Tesla's long shot was falling behind in the race.

The reported shutdown has led to a shakeup within the company. Peter Bannon, the head of the Dojo project, is leaving Tesla, and he isn't alone. About 20 members of the specialized team have left to join DensityAI, a new startup founded by former Tesla executives. The remaining staff from the Dojo program are said to be moving to other computing and data center roles within Tesla, applying their skills to different projects.

With its in-house project being shelved, Tesla is not abandoning its AI ambitions but rather changing the strategy. The company now plans to deepen its partnership with Nvidia, the current leader in AI processors, and it wants to increase its reliance on AMD, bringing another major chipmaker into its high-performance computing fold. This new direction also extends to the hardware inside its EVs. Future Tesla vehicles are expected to use an AI5 processor produced by TSMC starting this year, with a successor AI6 chip to be manufactured by Samsung later in the decade.

In a bit of a twist, the Dojo story might not be over just yet, though. Elon Musk hinted at a new vision for Tesla's hardware: a "converged architecture." The idea is to develop a single chip design that could be used in small numbers in a car or robot, and in large numbers on a server board for AI training. Musk referred to this potential future as "Dojo 3 and the AI6," suggesting that while the original Dojo project may be over, the dream of a custom-designed Tesla chip may simply be evolving into a more practical, less exotic form.

Via