one of the biggest gambles that Elon Musk and Tesla have made in recent years is the design and production of their own bespoke supercomputer Dojo musk has said repeatedly that Dojo is a risky move but it's also a technology that Tesla might not be able to live without let's talk about why that is what's really fascinating is that when Elon Musk started talking about Tesla's project Dojo back in 2019 the demand for graphics processing units or gpus was still limited primarily to the video game and digital video rendering industry supercomputers were pretty much only thought
of as machines that solved medical Mysteries and made climate predictions yet at Tesla's first autonomy day we saw the company's AI team standing on stage telling the world that they were going to build their very own supercomputer and use it to train artificial intelligence very few people at the time would have known how significant this was flashforward of 5 years and now gpus and supercomputer clusters have become the single most important piece of infrastructure in the AI Revolution that has swallowed the world the rapid shift in the current of the tech industry has catapulted Nvidia
to the top of the stock market making the world's biggest GPU supplier into the world's most valuable company ironically Elon Musk and Tesla have had a major role to play in the rise of Nvidia Tesla has spent billions and billions of dollars on Nvidia AI Hardware this is most evident in the company's latest construction project a gigantic addition built onto the south end of the Texas gigafactory that is now home to a supercomputer cluster that Elon Musk has named cortex about half of the chips that power this data center will be made by Nvidia amounting
to about 50,000 of their h100 gpus but a significant amount of the computing power in cortex will also be made by Tesla and somehow at the same time as cortex was being built Elon also found the time to construct an even bigger AI training computer for his xai startup company at a location in Memphis Tennessee that one is known to house up to 100,000 th of nvidia's current h100 and h200 GPU chips with Elon already threatening to buy an additional 300,000 units of nvidia's Next Generation chip the b200 so all that to say Elon Musk
knows the value of Nvidia Hardware better than most which is why he has been pushing his own staff at Tesla to develop something that can compete with what Nvidia is doing and this is about more than just trying to cut out some of nvidia's market share it's more about preserving Tesla's own ability to remain at the front of the AI race during Tesla's Q2 earnings called this year Elon said that demand for NVIDIA Hardware is so high that it's often difficult to get the gpus he continued by saying quote I am quite concerned about actually
being able to get steady gpus when we want them and I think this therefore requires that we put a lot more effort on dojo in order to ensure that we've about the training capability that we need so one significant reason that Tesla needs Dojo is simply because no matter how big and viia grows one company can't possibly quench the world's thirst for AI Hardware as Elon Musk would say the best supplier is no supplier of course this is also about money economics 101 says that when the demand is high and the supply is limited the
price will go up and when a company finds themselves in a position like Nvidia does today where the wealthiest companies in the world are demanding more and more Hardware than you could possibly produce then you can pretty much charge any price you want now that's not to say that designing and building your own bespoke supercomputer from the ground up is going to be any cheaper than buying Hardware on the open market not in the short term at least but Dojo is a long game going back to tesas q1 earnings call 2023 Elon told investors that
Dojo has the potential for an order of magnitude Improvement in the cost of training and also has the potential to become a sellable service that we would offer to other companies in the same way that Amazon web services offers web services he followed that up by saying quote look at Dojo as kind of a Longshot bet but a bet worth making in the summer of 2023 Tesla's Dojo D1 chip goes into mass production at tsmc in Taiwan it's probably not a coincidence that this timeline matches up with a sudden rise of nvidia's stock price on
June 21st Elon Musk confirms that an early version of the dojo computer rack has already been online and running useful tasks at the company for a few months in January 2024 we get our first indication of what a functional Dojo training cluster will look like while Elon also reiterates that Dojo is a high-risk high-reward project he also says that Tesla was pursuing the Dual path of Nvidia and Dojo that Dojo is working and is doing training jobs he notes Tesla is scaling up and has plans for Dojo 1.5 Dojo 2 Dojo 3 and whatnot now
Dojo V1 is a $500 million installation at Tesla's production facility in Buffalo New York we are not exactly sure why Tesla chose an obscure location but as someone who lives in the same part of the world we know that it gets cold during the winter which probably helps with the cooling and there's also Al an abundant stable supply of green energy from hydroelectric dams so in many ways Buffalo is probably a better location for a data center than Austin or Fremont either way Elon continues to temper expectations on dojo reminding us that $500 million is
actually not very much money saying it's only equivalent to a 10K h100 system from Nvidia Tesla will spend more than that on Nvidia Hardware this year year the table Stakes for being competitive in AI are at least several billion dollars per year at this point in June 2024 Elon puts these numbers into an even more specific context writing on X of the roughly 10 billion in AI related expenditures I said Tesla would make this year about half is internal primarily the Tesla designed AI interface computer and sensors present in all of our Cars Plus dojo
for building the AI training superclusters and Nidia Hardware is about 2/3 of the cost my current best guess for NVIDIA purchases by Tesla are 3 to 4 billion this year he follows that up by reiterating there is a path for Dojo to exceed Nvidia it's a long shot as I've said before but success is one of the possible outcomes so even among Tesla's own AI Hardware spending Dojo is a lower level priority than inference Computing Hardware this is the autopilot computer that is integrated into every test TLA vehicle it is the hardware that runs the
incar hardware that takes input from the eight cameras and uses that to make real time driving decisions Tesla is currently producing version four of their inference AI computer with version five expected to arrive Sometime Late next year or maybe in 2026 given Elon timelines but we know now that this incar processing power is quickly becoming more important than the data center training hardware and that the autopilot V5 chip will be pretty much necessary to achieving full autonomous driving Elon wrote on X our next gen AI model after this has a lot of Promise around 5x
increase in parameter count which is very difficult to achieve without upgrading the vehicle inference computer okay so what does Dojo actually do that's the tricky bit to explain because AI training is about the closest thing to digital alchemy that we've seen so far but Elon recently specified that Dojo 1 which is the Buffalo cluster that's being installed now will have the equivalent to roughly 8,000 Nvidia h100s of training power online by the end of this year Elon shared a couple of pictures that show the front and back of the dojo cabinets but no specific details
so we can gather that Dojo V1 is not massive but not trivial either it's less than 150 power of Cortex at Giga Texas and less than one10 the power of the X a gigafactory of compute in Memphis but in the case of Dojo raw power power shouldn't matter much keeping in mind that this is a custom designed AI training system that was developed entirely by Tesla so the hardware is custom tailored to Tesla specific applications which should make it significantly more efficient at those particular tasks than a generic multi-purpose chip like those from Nvidia musk
has said that the first version of Dojo will be tailored for Tesla computer vision labeling and training which is great for FSD and for training Optimus Tesla's humanoid robot but it wouldn't be useful for much else and for right now that's totally fine we know that the company's main push right now is getting the robo taxi vehicle ready to be unveiled in October of this year and it would make sense that the acceleration of AI training Hardware with cortex and Dojo is being done specifically to support the robotaxi project this is one of the rare
vehicles that will be judged entirely on its functionality its ability to drive autonomously so the software absolutely needs to deliver now that does leave an open-ended question on what the future of Dojo will actually be if Tesla has already decided that they are going to solve full self-driving within the next year or at least in the relatively short term then what do they really need dojo for well many things really so first off Tesla is pushing out the robotaxi in America and potentially China but for FSD to work in every part of the world the
AI will need to be trained on roow data that Tesla hasn't even collected yet if you think about the quantity of video that Tesla has collected from the US and China their two biggest markets they need equivalent data from Brazil India Thailand every language every different variety of road sign or lack of road sign there is no end in sight for Tesla's driving Ai and beyond that we know that Tesla will move into a more generalized AI model approach for their humanoid robot project right now they are training the robots Vision navigation and object detection
networks in pretty much the same way that they train FSD but for the Tesla bot to break out into the consumer market and fulfill elon's bold expectations of bots outnumbering people then they're going to need a pretty staggering variety of AI models that go Way Beyond Vision which is where dojo is currently con strained all we have to go on is a very limited reply from Elon he wrote in June 2023 Dojo V2 will address these limitations so stay tuned thank you to drift for partnering with us on this video check out the link Down
Below in the description
