Tesla Builds a New Home for ‘Dojo’ Supercomputer as AI Ambitions Rise
At Tesla’s headquarters in Austin, Texas, a new bunkerlike structure is under construction that could one day help move the company beyond electric vehicle manufacturing.
When it’s completed, the building will house part of a new supercomputer, known as Dojo, that Tesla is assembling to help run the artificial intelligence software behind the self-driving capabilities in its vehicles, according to two people familiar with the project. Eventually, Tesla could use Dojo to sell cloud services to other companies, the way Amazon Web Services does, the company’s CEO, Elon Musk, has said.
THE TAKEAWAY
• Tesla is building space in Austin to house its Dojo supercomputer
• Elon Musk has said Dojo could become an AWS-like cloud service
• Tesla is planning an assembly line to make robotaxis and a $25,000 car
The Austin Dojo project, details of which haven’t been previously reported, reflect an audacious plan by Musk to take greater control over the technology it needs to run the AI software at the heart of its products. Musk told investors in July that the company plans to invest “well over” $1 billion in developing Dojo by the end of 2024. The supercomputer—the first machines of which Tesla began using in a data center in Palo Alto, Calif., in the summer—is based on a chip designed in-house by Tesla. That could lessen Tesla’s dependence on Nvidia, the dominant supplier of the graphics processing units used to run AI applications, whose products the electric vehicle maker also uses.
Tesla already uses Nvidia AI chips to power the full self-driving software inside its fleet of Tesla vehicles, which feed video data from cameras on the cars back to the company. Tesla then uses that data to train the systems in its cars to better recognize their surroundings. But the company’s demands for AI computing power are expected to increase significantly as it adds new types of vehicles to its portfolio of products.
Inside Tesla's Texas gigafactory. Photo by Suzanne Cordeiro/AFP via Getty
For example, in the coming years, the company plans to build a robotaxi and a new entry-level Tesla expected to cost around $25,000, both of which it has long discussed. Tesla recently began detailed planning in Austin to build its next-generation assembly line—code-named NV9X—on which it plans to manufacture both the robotaxis and entry-level Teslas, according to people close to the project.
In April, Musk told Tesla investors: “Dojo also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon Web Services offers web services, even though it started out as a bookstore. So I really think that yes, the Dojo potential is very significant.”
Some Wall Street analysts have come to see Dojo as a big reason to be bullish on the company’s stock. Last month, Morgan Stanley analyst Adam Jonas predicted that Dojo could add up to $500 billion to Tesla’s enterprise value. Jonas based his forecast on an argument that Dojo could help Tesla realize its robotaxi ambitions and bring in new revenue from sources beyond the auto industry. He said Dojo customers could include companies that use vision-based computer learning, including businesses involved in robotics, aviation, railways and facial recognition, according to the report.
Chip Shortages
For Tesla, one wrinkle in those plans is that the kinds of chips required to run computers like Dojo are increasingly hard to come by, in large part because the primary supplier for them, Nvidia, is overwhelmed with demand.
”Everybody who is trying to do this is complaining that they can’t get enough of these chips,” said Glenn O’Donnell, a semiconductor researcher at technology research firm Forrester.
That’s a big reason why Tesla was motivated to design its own AI chip, known as D1. Other major tech players such as AWS and Microsoft have also begun designing their own AI chips to reduce their reliance on Nvidia. In July, Musk said Tesla would have continued to rely exclusively on Nvidia if the chip manufacturer’s supply could keep up with his company’s ambitions. In the future, the company plans to continue using Nvidia chips, along with the D1 chips that power Dojo, he said.
Press in Taiwan reported last month that Tesla had doubled its order for D1 chips with Taiwan Semiconductor Manufacturing Co., which is manufacturing them for Tesla.
Cost savings is also another important factor in Tesla’s decision to design its own chip. The D1 processor will give Tesla more control over how much energy it uses to run its AI software, Morgan Stanley’s Jonas wrote in his report. The company may also be able to process video data faster than with the Nvidia chips, he wrote. Tesla has said it expects Dojo to enable the company to train full self-driving workloads—discrete tasks or applications—in one week, a process that previously took a month, though those claims are yet to be proven. Morgan Stanley estimates that Tesla could save $6.5 billion over the next couple of years by designing its own chips rather than relying on Nvidia.
Tesla will need more and more physical space to accommodate Dojo’s growth. The company, which is headquartered in Austin and has engineers in California, has said it plans to use its Palo Alto data center to house seven ExaPods—a Tesla term for a cluster of 10 refrigerator-size cabinets brimming with Dojo computers running on D1 chips.
A year from now, Tesla has predicted, its Dojo computing capacity will increase to the equivalent of 100 ExaPods. That growth is likely one reason the company needs a second home for Dojo.
Long term, other Tesla businesses could drive additional demand for AI capacity. In July, Musk told Tesla investors the company is already in “early discussions with a major OEM,” an original equipment manufacturer, about licensing its full self-driving system to it. That could be another maker of commercial or consumer vehicles, though Musk didn’t specify what he meant by the term.
“We’re not trying to keep this to ourselves,” he said, referring to Tesla’s full self-driving software.
In his report on Dojo, Morgan Stanley’s Jonas also said Tesla could use Dojo to run the software behind a humanoid robot it is developing, Optimus. He speculated that other Musk companies, such as X and SpaceX, could buy services from Dojo.