For years, Elon Musk has talked about Dojo — the AI supercomputer that will be the cornerstone of Tesla’s AI ambitions. It’s an essential project for Musk, with plans to make it the foundation of Tesla’s self-driving tech. In July 2024, Musk emphasized how the company’s AI team would “double down” on Dojo as part of preparations for the launch of Tesla’s robotaxi service, which was revealed in October. But what exactly is Dojo? And why is it so crucial for Tesla’s long-term strategy?
What is Dojo?
In short, Dojo is Tesla’s custom-built supercomputer, designed to train the company’s Full Self-Driving (FSD) neural networks. These networks, which are integrated into hundreds of thousands of Tesla vehicles, can currently perform some driving tasks, but still require human supervision behind the wheel. Dojo aims to bring Tesla closer to full autonomy — meaning fully self-driving cars, with no human intervention needed.
The concept behind Dojo is rooted in Tesla’s vision of self-driving cars being trained solely with visual data. Unlike other companies that rely on sensors like lidar, radar, and cameras, Tesla believes FSD can be powered by cameras alone, using neural networks to interpret that data and make decisions — mimicking the way the human brain works. Dojo is integral to this vision because it processes vast amounts of driving data to train Tesla’s AI in real time.
Tesla’s Long-Term Strategy: The Road to Full Self-Driving
Tesla has positioned Dojo as the key to achieving full autonomy. The company’s goal is to deploy an autonomous ride-hailing service with its own fleet of self-driving vehicles. By June 2025, Tesla plans to launch unsupervised FSD in the U.S., marking a major milestone for the company and its ambitions to control the entire ecosystem of self-driving cars.
The rise of Dojo coincides with Tesla’s upcoming launch of robotaxis — a fully autonomous ride-hail service in Austin, Texas. Musk’s focus on AI and Dojo aligns with the company’s expansion into a broader AI-driven future, extending beyond just cars to include humanoid robots, as seen with Tesla Optimus.
Cortex and Dojo: What’s the Difference?
However, recently, much of the conversation around Tesla’s AI efforts has shifted toward Cortex, a new AI training supercluster designed to handle real-world AI problems. Cortex is being built at Tesla’s HQ in Austin, and Musk has referred to it as the next step in solving complex AI challenges related to FSD and Optimus. While Dojo was the primary focus in the past, Cortex now seems to be the focal point in terms of scaling Tesla’s AI capabilities. Musk has shared that Cortex is built to have massive storage and will handle video data from FSD and Tesla’s humanoid robot, Optimus.
Tesla’s AI Division: Pushing the Limits of Autonomous Driving
Tesla has always been vocal about AI-driven self-driving technology, and its investment in Dojo signals the company’s intention to lead the charge in AI-driven vehicles. This is not just about training FSD but revolutionizing how autonomous systems perceive and understand the world. The Dojo supercomputer is at the heart of this, processing enormous amounts of data from Tesla cars on the road.
For example, Tesla cars generate an incredible amount of video footage, and the company uses this data to refine its neural networks. The more data Tesla collects, the more it can push the boundaries of FSD technology. But as Anand Raghunathan, Purdue University professor, points out, there may be limits to this approach. Collecting more data does not necessarily mean a better model unless that data contains meaningful information that improves performance.
That’s where Dojo comes into play. Tesla is leveraging Dojo’s computing power to train its neural networks at unprecedented scale, allowing the company to process and analyze the data in real time. With Dojo, Tesla aims to bridge the gap between AI training and real-world performance, optimizing its system for fully autonomous driving.
Dojo’s Hardware: The Heart of Tesla’s AI Power
Dojo is built around Tesla’s custom D1 chips, designed for AI workloads. These chips are specialized to handle the intense demands of training Tesla’s neural networks. Unlike the GPUs used by other AI companies, the D1 chips are crafted for Tesla’s unique needs and offer a significant performance boost in machine learning tasks.
Image Source: Ganesh Venkataramanan, former senior director of Autopilot hardware, presenting the D1 training tile at Tesla’s 2021 AI Day.
Tesla’s choice to design its own chips is in line with the company’s philosophy of controlling its own hardware and software ecosystem. Much like Apple, which creates its own chips for its devices, Tesla believes that by controlling the hardware and software, it can create a more efficient and optimized system. The D1 chip is built by TSMC using a 7nm process and features 50 billion transistors, with a die size of 645 mm². Tesla has combined multiple D1 chips to create tiles that function as self-contained computers, each capable of 9 petaflops of computational power and 36 terabytes per second of bandwidth.
By scaling up its chip production and training infrastructure, Tesla is positioning itself to handle the growing demand for AI-powered vehicles and potentially unlock new revenue streams by offering AI as a service for other companies.
Dojo vs. Nvidia: Tesla’s Strategic Gamble
One of the significant challenges for Tesla has been its reliance on Nvidia’s GPUs, which are expensive and difficult to secure in large quantities. Although Tesla has used Nvidia’s A100 GPUs in the past to train its models, the company wants to shift towards its own hardware to reduce its dependency on third parties. This move comes with significant risks, as Dojo is still in its early stages, and there are concerns about whether it will be able to compete with the industry-standard Nvidia GPUs in terms of performance and availability.
Musk has emphasized the importance of Dojo to Tesla’s AI ambitions, stating that without Dojo, the company would not be able to process the massive amounts of data needed to train its models. Dojo is Tesla’s answer to the rising costs of Nvidia hardware, and the company is betting that its own chips will be able to handle the load more efficiently and at a lower cost in the long term.
What’s Next for Dojo?
As Tesla works to scale Dojo and continue refining its FSD capabilities, Musk’s vision of full autonomy is starting to become more tangible. Dojo is central to that vision, and as the company improves its supercomputer, it will continue to push the boundaries of what autonomous systems are capable of. In the coming years, Dojo will not only improve FSD but could also play a key role in the development of Tesla’s humanoid robots.
While the current version of Dojo may not yet be fully realized, Musk has laid out ambitious plans for its future. With Cortex now a major part of Tesla’s AI efforts, Dojo’s capabilities are only expected to expand. In the future, we may see Tesla offering AI as a service, renting out its AI supercomputing power to other companies, much like how cloud computing providers like AWS and Microsoft Azure operate.
As Dojo and Cortex evolve, Tesla is positioning itself to not only lead the autonomous driving revolution but also change the way the world thinks about AI and machine learning. With Dojo at the heart of Tesla’s plans, the company is betting that its custom-built supercomputer will be the key to achieving full autonomy and revolutionizing AI-powered vehicles and robots. Whether or not Dojo succeeds, it is clear that Tesla’s commitment to AI is not just about cars — it’s about reshaping entire industries.
