Tesla's Dojo Fails, Waymo Expands Globally: Autonomous Driving at a Crossroads

Tesla abruptly disbanded its Dojo supercomputer project, which had been painstakingly developed for six years, while Waymo steadily advanced its autonomous driving services in Denver and Seattle. The autonomous driving industry is now at a crossroads between two technological paths.

In August 2025, Tesla announced the disbandment of the Dojo team and the termination of the supercomputer project. Meanwhile, Waymo announced that it would start human - supervised tests in Denver and Seattle this week, preparing for the launch of autonomous taxi services in these two cities in 2025.

These two almost simultaneous events reveal the divergence and choices of different technological paths in the autonomous driving industry.

01 The Rise of Dojo

Elon Musk's Supercomputing Dream

In April 2019, Tesla first unveiled the Dojo supercomputer at the Autonomy Day event. Musk claimed at that time that all new cars were already equipped with the hardware for full - self driving, and only software upgrades were needed.

Dojo was designed as Tesla's proprietary supercomputer for training the FSD (Full Self - Driving) neural network. Its name comes from the Japanese word "dojo", symbolizing a training ground for AI technology.

Tesla had high hopes for Dojo. In August 2021, at Tesla's first AI Day, the company officially launched Dojo and showcased the D1 chip.

In July 2023, Tesla began producing Dojo and planned to invest over $1 billion before 2024. Musk boldly predicted that by February 2024, Dojo's computing power would rank among the top five in the world, and it would reach the goal of 100 exaflops by October.

02 The Vision - based Route

Tesla's Autonomous Driving Philosophy

The fundamental difference between Tesla and other autonomous driving companies lies in its pure vision - based technological route. Most companies rely on a combination of multiple sensors (such as lidar, radar, and cameras) and high - precision maps, while Tesla firmly believes that full - self driving can be achieved with just cameras.

Tesla's vision solution simulates the human visual perception system, using advanced neural networks to process visual data and make rapid driving decisions.

To support this technological route, Tesla needs to process a vast amount of video data. The millions of miles of video footage collected by Tesla's global fleet provide valuable resources for training the FSD system.

Dojo was precisely designed to handle this "truly massive amount of video data". Musk described Dojo as a "beast" in August 2020, capable of efficiently processing large - scale video training data.

03 Dojo's Failure

The End of an Ambitious Project

Despite huge investments and years of development, the Dojo project was ultimately terminated. In August 2025, Tesla disbanded the Dojo team, and the project leader, Peter Bannon, left the company.

About 20 Dojo employees left and founded the DensityAI company. On August 10, Musk posted an explanation for the decision to terminate Dojo 2, calling it an "evolutionary dead - end".

Musk said, "Since all paths lead to AI6, I had to terminate Dojo 2." He also revealed that Dojo 3 would essentially be "a large number of AI6 chips integrated on a single board".

The shutdown of Dojo was not entirely unexpected. In fact, Tesla started promoting Cortex as early as 2024, which is "a new super - training cluster for real - world AI R & D at the Austin headquarters".

04 Cortex Takes Over

Tesla's New Computing Strategy

According to Tesla's financial report in January 2025, the company had completed the deployment of Cortex, a training cluster consisting of 50,000 H100 GPUs. Cortex contributed to the performance improvement of FSD V13.

In the second quarter of 2025, Tesla further expanded its AI training computing power by adding 16,000 H200 GPUs, bringing Cortex's total computing power equivalent to 67,000 H100 GPUs.

Tesla also signed a $16.5 billion order with Samsung to purchase AI6 chips. These chips are planned to support FSD, Optimus, and high - performance AI training.

This series of actions indicates that Tesla is shifting from self - developed chips to a strategy of relying on partners.

05 Waymo's Expansion

Steady Progress of Autonomous Driving Services

While Tesla was adjusting its technological strategy, Waymo was steadily expanding its autonomous driving services. In September 2025, Waymo announced that it would launch autonomous taxi services in Denver and Seattle.

Waymo said it would start human - supervised tests in both cities this week. Up to 12 test vehicles will be deployed in each city, including fully electric Jaguar I - Pace and Zeekr autonomous driving models.

Waymo's current expansion plan includes entering ten new cities in 2025, launching services in Atlanta in cooperation with Uber in June, adding a teen account option in July, and obtaining a test permit in New York City in August.

Waymo currently provides paid autonomous taxi services in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta. The company also plans to launch commercial services in Miami, Washington D.C., and Dallas.

06 Global Landscape

Accelerated Competition in Autonomous Driving between China and the US

The competition in autonomous driving has long gone beyond the rivalry between Tesla and Waymo, presenting a global competition landscape, especially between China and the US.

Baidu Apollo Go opened its shared driverless car service to the public in the Huangpu Science City area of Guangzhou as early as July 2021. Users can take a free test ride through the Baidu Map App or the Apollo Go App. In April 2021, Baidu launched China's first paid fully driverless taxi service for the public at the Beijing Shougang Park. This is the second fully driverless ride - hailing service available to ordinary users globally after Waymo.

In September 2025, Chinese autonomous driving company WeRide announced that its first batch of the new - generation driverless taxis, Robotaxi GXR, had arrived in Singapore and started testing. This is the first time such autonomous driving models have landed in Southeast Asia.

In July 2025, the MogoMind large - scale model demonstrated by Mogu Auto at WAIC represents another technological direction. Through integrated devices for communication, sensing, and computing, it can capture massive amounts of heterogeneous data such as vehicle driving trajectories, speed changes, traffic flow, and pedestrian dynamics around the clock. The model has a parameter scale of 7 billion, with a perception accuracy and cognitive accuracy of over 90%, and has been applied in 8 cities.

Huawei also announced in September 2025 that its ADS 4.0 will have the pilot commercialization ability for high - speed L3 and the testing ability for urban L4 this year. This means that if regulations permit, we will be able to truly take our hands off the steering wheel on the highway in 2026.

07 Tesla Robotaxi

Service Expansion and Technological Challenges

Despite the termination of the Dojo project, Tesla's Robotaxi service is still advancing. On June 22, 2025, Tesla officially launched the Robotaxi pilot service in Austin, Texas, USA.

The first batch of passengers only need to pay a fixed fee of $4.20 to experience this service. About 10 to 20 vehicles were put into the pilot, which were modified based on the Model Y and equipped with Tesla's self - developed visual perception system and FSD software.

The service area of Tesla's Robotaxi expanded rapidly. It was launched in Austin on June 22, and within less than a month, it expanded its service area for the first time, covering an area larger than that of Waymo.

On August 3, the service area expanded again, with the geographical fence area nearly doubling to about 80 square miles. By August 27, Tesla expanded the geographical fence area of the Austin Robotaxi once more, covering 171 square miles, far exceeding Waymo's 90 square miles.

On September 3, Tesla adjusted the safety monitoring strategy for its Robotaxi service in Austin. As the service area expanded to include highways, Tesla decided to move the safety monitor's seat from the passenger seat to the driver's seat.

The Watershed Moment in Autonomous Driving

The termination of the Dojo project reflects the challenges Tesla faces in its technological route. Some industry experts believe that simply feeding more data to the model in the hope of making it smarter may have its limits.

Firstly, there are economic constraints, and this approach will soon become too expensive. Some claim that we may actually run out of meaningful training data.

More data does not necessarily mean more information. The key lies in whether the data contains information that can help create a better model and whether the training process can truly distill this information into a better model.

Nevertheless, in the short term, the trend of more data still seems to exist. And more data means more computing power is needed to store and process all this data to train Tesla's AI models.

This is exactly where the supercomputer Dojo was supposed to play a role.

------

The simultaneous termination of Tesla's Dojo project and the expansion of Waymo's services mark a watershed in the autonomous driving industry.

Three paths are becoming clear: on one hand, Tesla's vertical integration strategy is shifting from self - developed chips to partnerships with companies like Samsung and Nvidia; on the other hand, companies like Waymo and Baidu are pursuing gradual expansion, steadily advancing autonomous taxi services globally; the third path is the AI network route represented by Mogu Auto's MogoMind, achieving global optimization through a large - scale AI model in the physical world.

The future of autonomous driving is no longer a single - lane race on one path but a global competition with multiple technological routes coexisting.

This article is from the WeChat public account "Shanzzi", author: Rayking629, published by 36Kr with authorization.