One vision of Tesla’s future was quite different from the current direction the company is taking. In that version, Tesla’s vehicles learned how to drive themselves without ever relying on Nvidia, AMD, or Samsung thanks to rows of specially designed D-class chips humming inside a proprietary supercomputer that processed millions of hours of road footage. Dojo was that. For six years, it was one of the most closely watched hardware bets in the entire technology industry, and then, in August 2025, Elon Musk posted a few sentences on X and it was over.
The shutdown was quiet. Musk referred to Dojo 2 as “an evolutionary dead end,” which carries unusual weight for a project he once described as essential to Tesla’s AI identity. It’s difficult to ignore how swiftly the story changed—from intensifying efforts on Dojo in advance of Tesla’s robotaxi announcement in late 2024 to dissolving the entire team in a matter of months. The timing seems more like a decision that had been quietly developing for a while before anyone made it public than a calculated retreat.
It wasn’t just the hardware that made Dojo appealing. The concept was that Tesla could train its own neural networks on its own silicon, independent of the chip supply chain that everyone else was battling over, because it had a fleet of more than four million vehicles producing real-world driving data at scale. It was a dream of vertically integrated AI. Investors reacted appropriately, viewing the supercomputer as proof that Tesla was developing something more akin to an AI platform company rather than just selling electric cars. Both in terms of market value and narrative momentum, that story was extremely valuable.
If you were paying attention, you could see the cracks prior to the announcement. By early 2025, Tesla’s shareholder communications had shifted focus to Cortex, the new AI training cluster being built at Tesla’s Austin headquarters, with barely a mention of Dojo. Important engineers had already started to depart. DensityAI, a new startup founded by former Dojo leader Ganesh Venkataramanan, eventually hired about twenty former Dojo employees. It is rarely possible to recover from losing that kind of specialized talent from a specialized internal project, and it’s possible that Musk’s conclusion was accelerated by the departures.
The pivot is significant. In order to supply its next-generation AI6 chips, Tesla and Samsung inked a $16.5 billion agreement that will last until 2033. Musk framed it with characteristic confidence, noting the fab’s proximity to his home as a logistical convenience while describing the strategic importance as nearly impossible to overstate. It’s still unclear if that confidence is justified. The AI6 chip is designed to handle both vehicle inference and data center training, which is an ambitious dual mandate — and AMD and Nvidia aren’t standing still while Tesla rearchitects its approach.
Some observers believe that the Dojo story sheds light on a more general aspect of the economics of custom AI silicon. At the frontier, developing proprietary hardware is costly, time-consuming, and punishing when the underlying strategy changes even a little. Tensor Processing Units are available at Google. Trainium is available on Amazon. However, both of those businesses have cloud operations that are capable of absorbing the investment and sharing the expense. The fact that Tesla is an automaker with AI aspirations is more significant than what the media portrays.
Musk maintains that Dojo 3 is conceptually still present in the architecture of the AI6 chip. In a technical sense, that might be true. However, the independent supercomputer program has been completed, the team has left, and Tesla is now purchasing chips from the same suppliers it previously appeared committed to outgrowing. It will take years to find out if the cloud approach eventually results in better autonomous cars than Dojo could have. For the time being, the hardware race continues without one of its more interesting competitors.
