Tesla builds a supercomputer since he can’t get Nvidia chips

Tesla CEO Elon Musk is jealous of Nvidia customers getting the advanced AI training chips he so desperately wants. Nathan Laine Bloomberg/Getty Images

Think chip designer Nvidia’s trillion-plus-dollar valuation is mostly hot air thanks to the hype around generative AI?

Elon Musk would disagree. The Tesla CEO is willing to invest well over $1 billion to build an AI supercomputer he has dubbed Dojojust because he can’t get his hands on enough of Nvidia CEO Jensen Huang’s advanced A100 tensor core GPU clusters.

The EV visionary needs them to train both his Full Self-Driving (FSD) software currently retailing for $15,000, and his humanoid Optimus robots under development to perform useful everyday tasks.

Frankly, if they could supply us with enough GPUs, we might not need Dojo, Musk told analysts on Wednesday during Tesla’s second-quarter earnings call. But they can’t, because they have so many customers.

In October, Musk said he wasn’t even sure his latest project would prove superior to buying Nvidia chips off the shelf, which is unusual given that he typically demands the best from his team. Instead he praised Huang and thanked the company for prioritizing some of his orders.

Well, you actually get Nvidia hardware as fast as Nvidia will deliver it to us, he said yesterday.

Tesla has voracious demand for AI chips that seemingly outstrip virtually any commercial competitor. It aims to achieve an internal processing capacity of 100 exaFLOPS (an extremely high level of supercomputer performance) by the end of next year, a truly staggering figure that would easily place it among theworld top fiveknown vendors currently, according to the company.

Autonomy to skyrocket car sales

Musk is in a hurry as he has no intention of being proved wrong yet again with his annual ritual of predicting that he will soon achieve full autonomy. (Wednesday jokingly called himself the boy who cried FSD.)

This increasingly vague term tossing around can mean anything from a Tesla that just navigates traffic better than the average driver to actually driving itself without any human supervision.

The latter predicts that he will introduce Tesla’s ChatGPT moment and create in his words on Wednesday perhaps the largest asset value increase possibility in history as overnight his cars will instantly become robotaxis. That’s why Musk told analysts on the call that he’s so willing to sacrifice profitability for volume at a time when incumbent rivals like Volkswagen and Mercedes-Benz are taking the exact opposite approach.

If we had more training calculations, we’d do it faster, he said Wednesday. For now, FSD is just a driver assistance system that can make mistakes at the worst time.

When it comes to solving the problem, Musk has taken a fundamentally different approach to the entire industry, which requires a large expenditure on computing power.

Instead of a sinfully expensive suite of multiple redundant sensor systems like those used by Waymo, Cruise, Mobileye and their partners and deployed in carefully mapped geo-fenced areas, Musk wants his cars to be smart enough that they can drive anywhere autonomously just by relying on camera data. To train them, he’s using 300 million miles and counting the real-world driving data he scrapes off his customers’ cars every day.

That data will now be fed to Dojo, which is designed to significantly reduce the cost of neural network training using custom silicon developed in-house at Tesla. It is optimized for processing video data rather than mimicking the large language model used by OpenAI or Google DeepMind.

In the long run, we think the range will just be driving [vehicle sales] volume through the roof to the next level, Musk told investors Wednesday.