For years, Elon Musk has spoken of the promise of Dojo, the AI supercomputer that was once intended to be the cornerstone of Tesla’s AI ambitions. It was once vital sufficient to Musk that during July 2024, he stated the corporate’s AI workforce would “double down” on Dojo within the lead-up to Tesla’s robotaxi disclose, which came about in October.
After six years of hype, Tesla made up our minds ultimate month to close down Dojo and disband the workforce in the back of the supercomputer in August 2025. Inside of weeks of projecting that Dojo 2, Tesla’s moment supercluster that was once intended to be constructed at the corporate’s in-house D2 chips, would achieve scale through 2026, Musk reversed path, stating it “an evolutionary useless finish.”
This newsletter at first set out to provide an explanation for what Dojo was once and the way it will lend a hand Tesla succeed in full-self riding, self reliant humanoid robots, semiconductor autonomy, and extra. Now, you’ll call to mind it extra as an obituary of a challenge that satisfied such a lot of analysts and traders that Tesla wasn’t an automaker – it was once an AI corporate.
Dojo was once Tesla’s custom-built supercomputer that was once designed to coach its “Complete Self-Riding” neural networks.
Beefing up Dojo went hand-in-hand with Tesla’s objective to succeed in complete self-driving and produce a robotaxi to marketplace. FSD (Supervised) is Tesla’s complex motive force help device that’s on loads of 1000’s of Tesla cars these days and will carry out some computerized riding duties, however nonetheless calls for a human to be attentive in the back of the wheel. It’s additionally the root of an identical era powering Tesla’s restricted robotaxi provider that the corporate introduced in Austin this June the use of Fashion Y SUVs.
Whilst Dojo’s raison d’être began to return to lifestyles, Tesla did not characteristic its self-driving successes – arguable as they had been – to the supercomputer. Actually, Musk and Tesla had slightly discussed Dojo at far and wide the previous yr. In August 2024, Tesla started selling Cortex, the corporate’s “large new AI coaching supercluster being constructed at Tesla HQ in Austin to unravel real-world AI,” which Musk has stated would have “large garage for video coaching of FSD & Optimus.”
In Tesla’s This fall 2024 shareholder deck, the corporate shared updates on Cortex, however not anything on Dojo. It’s no longer transparent if Tesla’s shutdown of Dojo impacts Cortex.
Techcrunch tournament
San Francisco
|
October 27-29, 2025
The reaction to Dojo’s disbanding has been blended. Some see it as any other instance of Musk making guarantees he can’t ship on that comes at a time of falling EV gross sales and a lackluster robotaxi rollout. Others say the shutdown wasn’t a failure, however a strategic pivot from a high-risk, self-reliant {hardware} to a streamlined trail that is determined by companions for chip building.
Dojo’s tale unearths what was once at the line, the place the challenge fell brief, and what its shutdown indicators for Tesla’s long term.
A recap of Dojo’s shutdown
Tesla disbanded its Dojo workforce and close down the challenge in mid-August 2025. Dojo’s lead, Peter Bannon, left the corporate as smartly, following the departure of round 20 employees who left to begin their very own AI chip and infrastructure corporate referred to as DensityAI.
Analysts have identified that shedding key skill can briefly derail a challenge, particularly a extremely specialised, inside tech challenge.
The shutdown got here a few weeks after Tesla signed a $16.5 billion deal to get its next-generation AI6 chips from Samsung. The AI6 chip is Tesla’s wager on a chip design that may scale from powering FSD and Tesla’s Optimus humanoid robots to high-performance AI coaching in knowledge facilities.
“As soon as it was transparent that every one paths converged to AI6, I needed to close down Dojo and make some difficult workforce alternatives, as Dojo 2 was once now an evolutionary useless finish,” Musk posted on X, the social media platform he owns. “Dojo 3 arguably lives on within the type of a lot of AI6 [systems-on-a-chip] on a unmarried board.”
Tesla’s Dojo backstory
Musk has insisted that Tesla isn’t simply an automaker, or perhaps a purveyor of sun panels and effort garage techniques. As a substitute, he has pitched Tesla as an AI corporate, one who has cracked the code to self-driving automobiles through mimicking human belief.
Maximum different firms construction self reliant car era depend on a mix of sensors to understand the arena — like lidar, radar and cameras — in addition to high-definition maps to localize the car. Tesla believes it could actually succeed in totally self reliant riding through depending on cameras on my own to seize visible knowledge after which use complex neural networks to procedure that knowledge and make fast selections about how the automobile will have to behave.
The pitch has been that Dojo-trained AI device will sooner or later be driven out to Tesla shoppers by means of over-the-air updates. The dimensions of FSD additionally method Tesla has been ready to rake in hundreds of thousands of miles price of video pictures that it makes use of to coach FSD. The speculation there may be that the extra knowledge Tesla can accumulate, the nearer the automaker can get to in reality reaching complete self-driving.
Alternatively, some business professionals say there may well be a prohibit to the brute power manner of throwing extra knowledge at a type and anticipating it to get smarter.
“To start with, there’s an financial constraint, and shortly it’s going to simply get too dear to do this,” Anand Raghunathan, Purdue College’s Silicon Valley professor {of electrical} and pc engineering, advised TechCrunch. Additional, he stated, “Some folks declare that we’d in reality run out of significant knowledge to coach the fashions on. Extra knowledge doesn’t essentially imply additional info, so it is dependent upon whether or not that knowledge has data that comes in handy to create a greater type, and if the learning procedure is in a position to in reality distill that data into a greater type.”
Raghunathan stated regardless of those doubts, the craze of extra knowledge seems to be right here for the non permanent no less than. And extra knowledge method extra compute energy had to retailer and procedure all of it to coach Tesla’s AI fashions. That was once the place Dojo, the supercomputer, got here in.
What’s a supercomputer?
Dojo was once Tesla’s supercomputer device that was once designed to serve as as a coaching flooring for AI, in particular FSD. The identify is a nod to the gap the place martial arts are practiced.
A supercomputer is made up of 1000’s of smaller computer systems referred to as nodes. Every of the ones nodes has its personal CPU (central processing unit) and GPU (graphics processing unit). The previous handles general control of the node, and the latter does the advanced stuff, like splitting duties into more than one portions and dealing on them concurrently.
GPUs are crucial for system studying operations like those who energy FSD coaching in simulation. In addition they energy massive language fashions, which is why the upward thrust of generative AI has made Nvidia essentially the most treasured corporate in the world.
Even Tesla buys Nvidia GPUs to coach its AI (extra on that later).
Why did Tesla want a supercomputer?
Tesla’s vision-only manner was once the primary explanation why Tesla wanted a supercomputer. The neural networks in the back of FSD are skilled on huge quantities of riding knowledge to acknowledge and classify items across the car after which make riding selections. That signifies that when FSD is engaged, the neural nets have to gather and procedure visible knowledge frequently at speeds that fit the intensity and pace reputation functions of a human.
In different phrases, Tesla method to create a virtual replica of the human visible cortex and mind serve as.
To get there, Tesla must retailer and procedure the entire video knowledge gathered from its automobiles world wide and run hundreds of thousands of simulations to coach its type at the knowledge.
Tesla relied basically on Nvidia to energy its present Dojo coaching pc, nevertheless it didn’t need to have all its eggs in a single basket — no longer least as a result of Nvidia chips are dear. Tesla had was hoping to make one thing higher that higher bandwidth and diminished latencies. That’s why the automaker’s AI department made up our minds to get a hold of its personal tradition {hardware} program that aimed to coach AI fashions extra successfully than conventional techniques.
At that program’s core was once Tesla’s proprietary D1 chips, which the corporate stated had been optimized for AI workloads.
Inform me extra about those chips
Tesla, like Apple, believes {hardware} and device will have to be designed to paintings in combination. That’s why Tesla was once running to transport clear of the usual GPU {hardware} and design its personal chips to energy Dojo.
Tesla unveiled its D1 chip, a silicon sq. the dimensions of a palm, on AI Day in 2021. The D1 chip entered into manufacturing round July 2023.
The Taiwan Semiconductor Production Corporate (TSMC) manufactured the chips the use of 7 nanometer semiconductor nodes. The D1 has 50 billion transistors and a big die measurement of 645 millimeters squared, in keeping with Tesla. That is all to mention that the D1 guarantees to be extraordinarily robust and environment friendly and to take care of advanced duties briefly.
The D1 wasn’t as robust as Nvidia’s A100 chip, although.
Tesla were running on a next-gen D2 chip that aimed to unravel data drift bottlenecks. As a substitute of connecting the person chips, the D2 would have put all the Dojo tile onto a unmarried wafer of silicon.
Tesla by no means showed what number of D1 chips it ordered or won. The corporate additionally by no means equipped a timeline for a way lengthy it could have taken to get Dojo supercomputers operating on D1 chips.
What did Dojo imply for Tesla?
Tesla’s hope was once that through taking keep an eye on of its personal chip manufacturing, it will at some point have the ability to briefly upload massive quantities of compute energy to AI coaching techniques at a low value.
It additionally intended no longer having to depend on Nvidia’s chips one day, that are increasingly more dear and difficult to safe. Now, Tesla goes all-in on partnerships – with Nvidia, AMD, and Samsung, which is able to construct its next-gen AI6 chip.
All the way through Tesla’s second-quarter 2024 profits name, Musk stated call for for Nvidia {hardware} was once “so excessive that it’s incessantly tough to get the GPUs.” He stated he was once “rather curious about in reality with the ability to get stable GPUs when we wish them, and I feel this due to this fact calls for that we put much more effort on Dojo with a view to make certain that we’ve were given the learning capacity that we’d like.”
Dojo was once a dangerous wager, one who Musk hedged a number of instances through pronouncing that Tesla may no longer be triumphant.
Ultimately, Tesla toyed with the theory of making a brand new trade type in line with its AI department, with Musk even pronouncing throughout a Q2 2024 profits name that he noticed “a trail to being aggressive with Nvidia with Dojo.” Whilst D1 was once extra adapted for Tesla pc imaginative and prescient labeling and coaching – helpful for FSD and Optimus coaching – it wouldn’t were helpful for a lot else. Long run variations would must be extra adapted to general-purpose AI coaching, stated Musk.
The issue that Tesla may have arise in opposition to is that the majority AI device in the market has been written to paintings with GPUs. The use of Dojo chips to coach general-purpose AI fashions would have required rewriting the device.
This is, except Tesla rented out its compute, very similar to how AWS and Azure hire out cloud computing functions – an concept that excited analysts. A September 2023 file from Morgan Stanley predicted that Dojo may just upload $500 billion to Tesla’s marketplace worth through unlocking new earnings streams within the type of robotaxis and device services and products.
Briefly, Dojo chips had been an insurance plans for the automaker, however one who may have paid dividends.
How some distance did Tesla Dojo get?
Musk incessantly equipped development reviews, however lots of his targets for Dojo had been by no means reached.
For example, Musk advised in June 2023 that Dojo were on-line and operating helpful duties for a couple of months.” Round the similar time, Tesla stated it anticipated Dojo to be probably the most most sensible 5 maximum robust supercomputers through February 2024 and had deliberate for overall compute to succeed in 100 exaflops in October 2024, which might have required kind of 276,000 D1s, or round 320,500 Nvidia A100 GPUs.
Tesla by no means equipped an replace or any data that might counsel it ever reached those targets.
Tesla and Musk made a lot of different pledges for Dojo, together with monetary ones. For example, Tesla dedicated in January 2024 to spend $500 million to construct a Dojo supercomputer at its gigafactory in Buffalo, New York, and has already spent $314 million of that, according to a 2024 file.
Simply after Tesla’s second-quarter 2024 profits name, Musk posted pictures of Dojo 1 on X, pronouncing that it could have “kind of 8k H100-equivalent of coaching on-line through finish of yr. Now not large, however no longer trivial both.”
Regardless of all of this task — in particular through Musk on X and in profits calls — point out of Dojo swiftly ended August 2024. And communicate switched to Cortex.
All the way through the corporate’s fourth-quarter 2024 profits name, Tesla stated it finished the deployment of Cortex, “a ~50k H100 coaching cluster at Gigafactory Texas” and that Cortex helped allow V13 of supervised FSD.
In Q2 2025, Tesla famous it “expanded AI coaching compute with an extra 16k H200 GPUs at Gigafactory Texas, bringing Cortex to a complete of 67k H100 equivalents.” All the way through that very same profits name, Musk stated he anticipated to have a moment Dojo cluster working “at scale” in 2026. He additionally hinted at attainable redundancies.
“Interested by Dojo 3 and the AI6 inference chip, it sort of feels like intuitively, we wish to check out to seek out convergence there, the place it’s mainly the similar chip,” Musk stated.
A couple of weeks later, he reversed path and disbanded the Dojo workforce.
TechCrunch showed in past due August 2025 that Tesla nonetheless plans to dedicate $500 million to a supercomputer in Buffalo – it simply gained’t be Dojo.
This tale at first printed August 3, 2024. The thing was once up to date for a last time September 2, 2025 with new details about Tesla’s resolution to close down Dojo.
