radically

Elon Musk is one of the most important agents in the era of artificial intelligence. Meta, Alphabet, Microsoft with OpenAI and Oracle are prominent names when we talk about gigantic data centersbut if there is someone who cuts the mustard, it is Musk with his xAI company. His Colossus Memphis with 100,000 H100 from NVIDIA to train Grok surprised even Jensen HuangCEO of NVIDIA, but Musk’s goal is not to depend on others. NVIDIA leads the way in chips to train AI (so much so that even Chinese companies want to buy its H200, even if they don’t let them do it). But Musk, like China, wants independence and technological sovereignty, and That’s why he invested in Dojo. It was an ambitious plan to build a customized supercomputer to train the neural networks of the controversial autonomous driving (the FSD). After more than five years in development, 1,000 million dollars invested and key engineers who took the lead drainMusk hill the tap in August of last year. The future was in the AI5 and AI6 chips which were less specific, but could still be used to train the FSD system. However, there is a new twist to this tortilla chip and Musk has decided to relaunch the project. tesla reactive the development of Dojo 3, and it does so by burning bridges with the previous philosophy of this supercomputer. Dojo 3, the heart of Tesla’s autonomous driving Although Tesla has stopped more doubts than anything else these last few years regarding autonomous driving concerned, this continues to be one of the pillars in the company’s short-term strategy. Because they not only have the FSD in their cars, but also in the controversial ‘robotaxis’. Supposedly, it will be this 2026 when Cybercaps will begin to be manufacturedcars that, unlike the taxis that we already see in some cities, will arrive without pedals or a steering wheel. But he doesn’t just want to fuel his cars. Musk wants to make money with softwarebut to have that software, you need to train the system and make it more secure than now. That’s where Dojo came into play. This hardware depended on a very specialized and complex architecture. The D1 chip was the heart of it all, but to achieve high computing power a complex network of thousands of D1 chips mounted in physically separate cases and interconnected by Ethernet cables was needed. It was a very specialized system, but complex to scale without skyrocketing costs. When Tesla turned off the Dojo tap, it commented that its companies would continue investing in the creation of less specialized chips such as the AI5, AI6, AI7 and subsequent ones. More conventional and easier to scale chips. And, precisely, the advances in this architecture are the decisive factor for Musk to revive Dojo. Instead of requiring complex interconnected equipment, Dojo 3 will adopt a modular architecture in which several AI chips can be installed on a single board. Not only is wiring complexity reduced, but heat dissipation is facilitated and the space required for installation is reduced. And, the easier it is and the less space it requires, the more chips can be mounted and the greater computing power. It is not the only advantage. Grouping chips on a single board reduces latency within the chips and improves the power efficiency of the device. To give an example, although they are a headache for expansion, it is the same philosophy that laptops with SSD or RAM memory soldered to the board: Everything communicates faster, more fluidly and requiring less energy to operate. Furthermore, being less specific than D1, xAI’s AIs fulfill both training and inference functions (the Dojo only served for training), which represents cost savings for the company. Now, Dojo 3 will not be a reality immediately. In recent days, Musk has shared via Twitter X the roadmap for its semiconductors. The AI5 developed together with TSMC is “almost finished” and they are already in the early stages of AI6. Meanwhile, he hopes that there will be a new version every nine months, with the AI7 and subsequent ones in the company’s plans for 2027. And a big question is who will make these chips. We can immediately think of TSMC, a leading company in these fields that even is expanding in the United States and that already has clients like herself NVIDIA for its new AI training chips. But no: it will be Samsung. At least, of course, for an AI6 with which Tesla signed a $16.5 billion deal that was seen as a victory for the South Korean giant’s function. We will see how the plans evolve, since if something appears that they consider better, they have shown us that their hand does not tremble when it comes to swerving, but This strategy on less specialized chips is interesting taking into account the needs in autonomous driving, AI training and robotics that the company faces. Images | xAI, Steve Juvetson In Xataka | Elon Musk wants to turn xAI into an ultra-valuable company and he knows how to do it: using the SpaceX vault