Transcription
Tesla's Dojo supercomputer is dead, and Elon Musk killed it, but probably not for the reason you think. Dojo was first introduced at Tesla AI Day in 2021 as a super powerful, superefficient supercomputer that was built in a way that no one had done before. Tesla promised that unlike other computing clusters, this one was designed specifically for AI training, which allowed it to do more work with less power and that would help deliver on the company's grand vision for self-driving cars and humanoid robots, which were also announced on the same day.
Now, 4 years later, everyone who was working on Dojo has either quit or been fired, including the guy who designed the computer system in the first place. And just recently, the guy who was brought in to replace him as Dojo lead has also left the company. So things had been looking bad for a little while. And then Elon Musk himself put the final nail in the coffin by saying publicly that he has shut down Project Dojo. Here's a post that Elon made earlier this week. And some of this stuff might sound like gibberish right now, but we'll explain all of this in a way that should help you understand how Tesla got into this situation, and then you'll know where they are going in the future.
So Elon wrote, "Once it became clear that all paths converged to AI6, I had to shut down dojo and make some tough personnel choices as Dojo 2 was now an evolutionary dead end. Dojo 3 arguably lives on in the form of a large number of AI6 SOC's on a single board."
Okay, let's start with SOC. That means system on a chip. It's a mini computer that all exists on one piece of silicon: the CPU, the GPU, and the RAM. Everything is just shrunken down and stuck together. That's how smartphones work. That's how all modern Apple computers work. And it's what Tesla chose for their first Dojo chip, the D1. That's what this guy is holding here in his hand at the AI Day presentation. That's Ganesha. We'll get back to him. He's important.
Now, to make Dojo work, you actually need 25 D1 chips, and those are all put on one motherboard. So now we've got 25 mini computers all working together to create one big computer. That's called a tile. It's about the size of a dinner plate. Here's Ganesha holding one of those as well. Then you stack a couple of those tiles into a cabinet, and that's a supercomputer. That's the architecture that defined Dojo and made it different from other computer systems. And that's what Elon means by a large number of SOC's on a single board. And it sounds like Tesla still likes that design idea. They're just going to do it with something different: AI6.
So what's that? Long before Dojo was announced, Tesla had already begun to dabble in the world of chip design. They built a computer system specifically for their vehicles so that the cars could use advanced artificial intelligence for autonomous driving. This was called Hardware 3. For Hardware 1 and 2, Tesla had bought third-party self-driving computers, but they didn't work very well, which led the company to design their own.
Okay, imagine this. You're handed the keys to a brand new Tesla, or maybe a Rivian or a Lucid, or you just take the $50,000 in cold, hard cash. That is what's up for grabs in the 11th annual Tesla raffle from Chesset Chicago. You are not locked into just one prize. You choose your dream EV or the cash. And with only 9,999 tickets sold, the odds are actually pretty good. And the best part, every ticket helps fund over 80 life-changing programs from Chesset Chicago, providing food, furniture, job support, and so much more to families in crisis. So head to ccraffle.com, use the code Teslaspace, and get $25 off two tickets or $500 off 15. That's ccraffle.com promo code teslaspace. Take your shot at the EV of your dreams and do some real good while you're at it.
What the in-car computer does is something called inference compute. That's what allows the AI model to run in real time and control the vehicle. Think about the AIs that you might typically interact with like ChatGPT or Grok. When you ask it a question, that prompt gets sent off through the internet. The request gets processed in a giant data center, and the answer gets sent back through the internet. So none of the AI work was actually performed on your device. This process used to take some time, but now it happens a lot faster because the data centers have gotten a lot more powerful. We'll talk more about that in the next video, but even now, you can't do something like self-driving by sending prompts and responses through the web. The AI has to be running locally in the car. And just like those data centers, in order to get a fast response, you need a powerful computer. The big difference is that a data center can pull massive amounts of electricity to fuel that processing power. A car doesn't have that luxury. So, the inference computer has to be incredibly efficient. That's what Tesla was really able to unlock with Hardware 3.
In 2023, Tesla upgraded their in-car computer to Hardware 4, which Elon Musk also started calling AI4 because its real job is just to run the full self-driving AI. In about a year from now, Tesla will upgrade again to AI5. Last week, Elon Musk posted, "The difference in real world performance between AI4 and AI 5 is far more than any chip version I've ever heard of by a lot. It's real good." And then by 2028 or 2029, Tesla already has a roadmap to producing AI6.
Meanwhile, Tesla was still developing the Dojo computer specifically for AI training. But unlike Hardware 3 and AI4, which are used by millions of vehicles on a daily basis, Dojo never actually contributed anything of value. There were a few Dojo cabinets built, and they may have done some useful work, but Tesla already had a massive AI training computer up and running before Dojo that was powered by Nvidia GPU chips. And then even after Dojo was released, Tesla built another much bigger, much more powerful AI training computer, again powered by Nvidia chips.
And this kind of gets into the real reason why Dojo was cancelled. Tesla designed their own inference computer because no one else was making a product that did the job they required. But when it comes to the training side of AI, Nvidia already makes a product that does exactly what Tesla needs. The entire AI industry is powered by Nvidia. It's what propelled the chipmaker to a new status as the most valuable company in the world. They went from supplying video game nerds and crypto miners to providing the hardware behind every AI model in the world. Tesla was the only one trying to make their own training chip.
And even while they were doing that, Elon Musk became Nvidia's number one customer. Between Tesla and XAI, Elon has bought somewhere in the neighborhood of a million Nvidia GPUs, and he's got a million more on order right now. But this was not a secret. Elon wasn't buying chips in backroom deals. This is one of the few occasions where Elon has been very realistic with his public statements. He never said that Dojo would be the ultimate supercomputer in the world and that no one would ever be able to compete. He said that about other stuff he's made, but when it comes to Dojo, Elon was always pretty clear that this was a gamble, a hedged bet, meaning he was going to keep buying Nvidia, but he was also going to create an Nvidia competitor.
This makes sense if you remember the world in the year 2021. We were in the height of a supply chain breakdown, and computer chips were the biggest victim. Video game consoles, vehicles, microwaves, nothing was getting made because there were no microchips to put in them. So, there was no way to know for sure at that time if Nvidia could actually supply enough GPUs to satisfy the booming AI industry. In 2021, we had never heard of ChatGPT, but people on the inside knew it was coming, and they knew there was an arms race brewing behind the scenes. So, of course, Elon would be taking steps to ensure that Tesla gets what it needs one way or the other.
Turns out that betting on Nvidia was the right call. If you'd put your money into their stock just 4 years ago, you'd be up almost 10x right now. If you'd done the same with Tesla stock, you'd have basically the exact same amount of money that you started with.
Anyway, that's what Elon was getting at with his post last week, writing, "It doesn't make sense for Tesla to divide its resources and scale two quite different AI chip designs. The Tesla AI5, AI6, and subsequent chips will be excellent for inference and at least pretty good for training. All effort is focused on that." Basically, meaning that Tesla already has all of the training chips they need, and they work great. So there's no point pouring money into research and development and manufacturing for a new training chip that is honestly still not proven to be effective at doing anything.
And let's be honest, Tesla is not exactly in a great financial era at the moment. But they do need in-car inference chips. Without those, self-driving robo-taxis will not work, and that would take their money troubles from bad to much worse. So that's where the resources are going right now. And if there ever was a situation where they needed extra training resources, it looks like these inference chips could be useful in a pinch.
From the sounds of it, Elon wants to build a new supercomputer cluster based on the Dojo architecture, but using the new AI chip, and then Tesla could run them any way they like, either for processing AI or training AI. That's what he was getting at with this post. In a supercomputer cluster, it would make sense to put many AI5/AI6 chips on a board. Whether for inference or training, or simply to reduce network cabling complexity and cost by a few orders of magnitude, one could call that Dojo 3, I suppose.
So the physical body of Dojo is pretty much dead, but its spirit may live on, not just at Tesla, but also in a new AI company. Remember Ganesha? Now, he was the first father of Dojo, but he turned out to be another deadbeat dad. He left Tesla in 2023 to start his own AI company, which we now know as Destiny AI. Then, over the years, Ganesha has slowly siphoned off 20 of the best engineers from Project Dojo over to his new operation. Density AI is working on full-stack AI solutions, meaning chips, computers, data centers, and the software to run on it. All made specifically for high-performance AI applications like autonomous vehicles and robots. Sound familiar? He's basically rebuilding Tesla's AI division in a company that's not Tesla, which honestly is a genius idea.
But over at Tesla, this really doesn't change anything. Dojo was never really that big of a deal. The internet just hyped it up into one, myself included. We are among the most guilty in that regard because here are the facts. Tesla has already built a massive AI training computer at Giga Texas named Cortex and filled it with Nvidia chips. And that's what's driving progress for FSD right now. Tesla is already planning a Cortex 2 and had already planned to fill that with Nvidia chips. Meanwhile, Elon is building two of the largest AI supercomputers in the world for XAI and he's happily filling both of them with Nvidia chips. So, Dojo or no Dojo, there's really no difference. Everything is going to be just fine in terms of AI hardware, at least. Other things at Tesla right now.