Wall Street analysts are buying into Tesla’s in-development supercomputer Dojo, saying it could add US$500 billion of value to the company by expanding Tesla into the software and services market.
Morgan Stanley published an investor note earlier this week claiming that Dojo would be to Tesla what AWS is to Amazon: a way of breaking away from just the sale of real-world assets, like cars, into licensing the AI underlying its semi-autonomous abilities, or by selling compute time on custom supercomputers.
“Longer term, we see scope for Dojo to provide avenues for Tesla’s software and hardware capabilities to extend well beyond the auto industry,” Morgan Stanley’s analysts said.
“If Dojo can help make cars ‘see’ and ‘react’, what other markets could open up? Think of any device at the edge with a camera that makes real-time decisions based on its visual field.”
Morgan Stanley reportedly has close ties to Tesla and was intricately involved in Elon Musk’s acquisition of Twitter.
News of the firm’s analysis bumped Tesla’s share price up 10 per cent.
To be clear, Tesla already trains AI models on its own supercomputing clusters that, according to a 2021 presentation, boast using 5,760 Nvidia A100 GPUs to help process the immense amount of data sent home by its cars from around the world.
Tesla vehicles are computers on wheels. As such, they can download over-the-air updates that fix potentially fatal software errors but they can also upload data collected from external cameras and onboard telemetry.
Why would Tesla need to gather data from its cars? Because the company’s goal is to make a self-driving car powered by AI – and AI needs lots of data to train.
Customers using the Tesla’s Autopilot (smart cruise control) and its Full Self Driving (which isn’t full self-driving) features are beta testers.
If a driver intervenes on these somewhat automated systems – say, by slamming on the brakes because the car is about to run a stop sign – the car will send data from the incident back to Tesla headquarters where it can contribute to the next version of those features and, hopefully, reduce its likelihood of killing or maiming other road users and pedestrians.
Enter the Dojo
Here is where Morgan Stanley sees a, perhaps excessive, amount of value in Tesla.
Not only have its highly-connected electric cars been on the road for longer than its competitors, collecting and processing a lot more data, Tesla has also begun building its own infrastructure to turn that data into software features that other manufacturers may find valuable.
Tesla doesn’t plan on using Nvidia cards forever. Rather, it’s been designing proprietary supercomputing hardware: the D1 chip.
Back in 2021, Tesla offered a first look at this chip which it combined into a 9-petaflops tile. Add 12 of those tiles to a cabinet and you’ve got over 100 petaflops, Tesla claimed – combine 10 of those cabinets and you’ve got an ‘ExaPod’ with a remarkable one exaflop of compute power.
The US Department of Energy’s Oak Ridge National Laboratory built the world’s first exascale supercomputer last year, marking a new era of supercomputing.
Tesla claims it will reach the exascale milestone, without using out-of-the-box parts, by the end of 2024.
In fact, in Tesla’s July quarterly report, it claimed to have started production on Dojo.
According to Tesla, and parroted by Morgan Stanley, in-house production of supercomputing hardware will save US$6.5 billion “over the next couple of years” at a crucial time when businesses around the world are ramping up AI training.
“Although Dojo is still early in its development, we believe that its application long-term can extend beyond the auto industry,” Morgan Stanley said.
“Dojo is designed to process visual data which can lay the foundation for vision-based AI models such as robotics, healthcare, and security.
“In our view, once Tesla makes headway on autonomy and software, third-party Dojo services can offer investors the next leg of Tesla’s growth story.”