Tesla Dojo: The Rise And Fall Of Elon Musk’s Ai Supercomputer

2 days ago

For years, Elon Musk has spoken of nan committedness of Dojo, nan AI supercomputer that was expected to beryllium nan cornerstone of Tesla’s AI ambitions. It was important capable to Musk that successful July 2024, he said nan company’s AI squad would “double down” connected Dojo successful nan lead-up to Tesla’s robotaxi reveal, which happened successful October.

After six years of hype, Tesla decided past period to shut down Dojo and disband nan squad down nan supercomputer successful August 2025. Within weeks of projecting that Dojo 2, Tesla’s 2nd supercluster that was meant to beryllium built connected nan company’s in-house D2 chips, would scope standard by 2026, Musk reversed course, declaring it “an evolutionary dormant end.”

This article primitively group retired to explicate what Dojo was and really it could thief Tesla execute full-self driving, autonomous humanoid robots, semiconductor autonomy, and more. Now, you tin deliberation of it much arsenic an obituary of a task that convinced truthful galore analysts and investors that Tesla wasn’t an automaker – it was an AI company.

Dojo was Tesla’s custom-built supercomputer that was designed to train its “Full Self-Driving” neural networks.

Beefing up Dojo went hand-in-hand pinch Tesla’s extremity to scope afloat self-driving and bring a robotaxi to market. FSD (Supervised) is Tesla’s precocious driver assistance strategy that’s connected hundreds of thousands of Tesla vehicles coming and tin execute immoderate automated driving tasks, but still requires a quality to beryllium attentive down nan wheel. It’s besides nan ground of akin exertion powering Tesla’s constricted robotaxi work that nan institution launched successful Austin this June utilizing Model Y SUVs.

Even arsenic Dojo’s raison d’être started to travel to life, Tesla grounded to property its self-driving successes – controversial arsenic they were – to nan supercomputer. In fact, Musk and Tesla had hardly mentioned Dojo astatine each complete nan past year. In August 2024, Tesla began promoting Cortex, nan company’s “giant caller AI training supercluster being built astatine Tesla HQ successful Austin to lick real-world AI,” which Musk has said would person “massive retention for video training of FSD & Optimus.”

In Tesla’s Q4 2024 shareholder deck, nan institution shared updates connected Cortex, but thing connected Dojo. It’s not clear if Tesla’s shutdown of Dojo affects Cortex.

Techcrunch event

San Francisco | October 27-29, 2025

The consequence to Dojo’s disbanding has been mixed. Some spot it arsenic different illustration of Musk making promises he can’t present connected that comes astatine a clip of falling EV income and a lackluster robotaxi rollout. Others opportunity nan shutdown wasn’t a failure, but a strategical pivot from a high-risk, self-reliant hardware to a streamlined way that relies connected partners for spot development.

Dojo’s communicative reveals what was connected nan line, wherever nan task fell short, and what its shutdown signals for Tesla’s future.

A recap of Dojo’s shutdown

Tesla disbanded its Dojo squad and unopen down nan task successful mid-August 2025. Dojo’s lead, Peter Bannon, near nan institution arsenic well, pursuing nan departure of astir 20 workers who near to commencement their ain AI spot and infrastructure institution called DensityAI.

Analysts person pointed retired that losing cardinal talent tin quickly derail a project, particularly a highly specialized, soul tech project.

The shutdown came a mates of weeks aft Tesla signed a $16.5 cardinal deal to get its next-generation AI6 chips from Samsung. The AI6 spot is Tesla’s stake connected a spot creation that tin standard from powering FSD and Tesla’s Optimus humanoid robots to high-performance AI training successful information centers.

“Once it became clear that each paths converged to AI6, I had to unopen down Dojo and make immoderate reliable unit choices, arsenic Dojo 2 was now an evolutionary dormant end,” Musk posted connected X, nan societal media level he owns. “Dojo 3 arguably lives connected successful nan shape of a ample number of AI6 [systems-on-a-chip] connected a azygous board.”

Tesla’s Dojo backstory

Musk has insisted that Tesla isn’t conscionable an automaker, aliases moreover a purveyor of star panels and power retention systems. Instead, he has sounded Tesla arsenic an AI company, 1 that has cracked nan codification to self-driving cars by mimicking quality perception.

Most different companies building autonomous conveyance exertion trust connected a operation of sensors to comprehend nan world — for illustration lidar, radar and cameras — arsenic good arsenic high-definition maps to localize nan vehicle. Tesla believes it tin execute afloat autonomous driving by relying connected cameras unsocial to seizure ocular information and past usage precocious neural networks to process that information and make speedy decisions astir really nan car should behave.

The transportation has been that Dojo-trained AI package will yet beryllium pushed retired to Tesla customers via over-the-air updates. The standard of FSD besides intends Tesla has been capable to rake successful millions of miles worthy of video footage that it uses to train FSD. The thought location is that nan much information Tesla tin collect, nan person nan automaker tin get to really achieving afloat self-driving.

However, immoderate manufacture experts opportunity location mightiness beryllium a limit to nan brute unit attack of throwing much information astatine a exemplary and expecting it to get smarter.

“First of all, there’s an economical constraint, and soon it will conscionable get excessively costly to do that,” Anand Raghunathan, Purdue University’s Silicon Valley professor of electrical and machine engineering, told TechCrunch. Further, he said, “Some group declare that we mightiness really tally retired of meaningful information to train nan models on. More information doesn’t needfully mean much information, truthful it depends connected whether that information has accusation that is useful to create a amended model, and if nan training process is capable to really distill that accusation into a amended model.”

Raghunathan said contempt these doubts, nan inclination of much information appears to beryllium present for nan short-term astatine least. And much information intends much compute powerfulness needed to shop and process it each to train Tesla’s AI models. That was wherever Dojo, nan supercomputer, came in.

What is simply a supercomputer?

Dojo was Tesla’s supercomputer strategy that was designed to usability arsenic a training crushed for AI, specifically FSD. The sanction is simply a motion to nan abstraction wherever martial arts are practiced.

A supercomputer is made up of thousands of smaller computers called nodes. Each of those nodes has its ain CPU (central processing unit) and GPU (graphics processing unit). The erstwhile handles wide guidance of nan node, and nan second does nan analyzable stuff, for illustration splitting tasks into aggregate parts and moving connected them simultaneously.

GPUs are basal for instrumentality learning operations for illustration those that powerfulness FSD training successful simulation. They besides powerfulness ample connection models, which is why nan emergence of generative AI has made Nvidia nan astir valuable institution connected nan planet.

Even Tesla buys Nvidia GPUs to train its AI (more connected that later).

Why did Tesla request a supercomputer?

Tesla’s vision-only attack was nan main logic Tesla needed a supercomputer. The neural networks down FSD are trained connected immense amounts of driving information to admit and categorize objects astir nan conveyance and past make driving decisions. That intends that erstwhile FSD is engaged, nan neural nets person to cod and process ocular information continuously astatine speeds that lucifer nan extent and velocity nickname capabilities of a human.

In different words, Tesla intends to create a integer copy of nan quality ocular cortex and encephalon function.

To get there, Tesla needs to shop and process each nan video information collected from its cars astir nan world and tally millions of simulations to train its exemplary connected nan data.

Tesla relied chiefly connected Nvidia to powerfulness its existent Dojo training computer, but it didn’t want to person each its eggs successful 1 handbasket — not slightest because Nvidia chips are expensive. Tesla had hoped to make thing amended that accrued bandwidth and decreased latencies. That’s why nan automaker’s AI section decided to travel up pinch its ain civilization hardware programme that aimed to train AI models much efficiently than accepted systems.

At that program’s halfway was Tesla’s proprietary D1 chips, which nan institution said were optimized for AI workloads.

Tell maine much astir these chips

Ganesh Venkataramanan, erstwhile elder head of Autopilot hardware, presenting nan D1 training tile astatine Tesla’s 2021 AI Day.

Tesla, for illustration Apple, believes hardware and package should beryllium designed to activity together. That’s why Tesla was moving to move distant from nan modular GPU hardware and design its ain chips to powerfulness Dojo.

Tesla unveiled its D1 chip, a silicon quadrate nan size of a palm, connected AI Day successful 2021. The D1 spot entered into accumulation astir July 2023.

The Taiwan Semiconductor Manufacturing Company (TSMC) manufactured nan chips utilizing 7 nanometer semiconductor nodes. The D1 has 50 cardinal transistors and a ample dice size of 645 millimeters squared, according to Tesla. This is each to opportunity that nan D1 promises to beryllium highly powerful and businesslike and to grip analyzable tasks quickly.

The D1 wasn’t arsenic powerful arsenic Nvidia’s A100 chip, though.

Tesla had been moving connected a next-gen D2 spot that aimed to lick accusation travel bottlenecks. Instead of connecting nan individual chips, nan D2 would person put nan full Dojo tile onto a azygous wafer of silicon.

Tesla ne'er confirmed really galore D1 chips it ordered aliases received. The institution besides never provided a timeline for really agelong it would person taken to get Dojo supercomputers moving connected D1 chips.

What did Dojo mean for Tesla?

Tesla’s dream was that by taking power of its ain spot production, it mightiness 1 time beryllium capable to quickly adhd ample amounts of compute powerfulness to AI training programs astatine a debased cost.

It besides meant not having to trust connected Nvidia’s chips successful nan future, which are progressively costly and difficult to secure. Now, Tesla is going all-in connected partnerships – pinch Nvidia, AMD, and Samsung, which will build its next-gen AI6 chip.

During Tesla’s second-quarter 2024 net call, Musk said request for Nvidia hardware was “so precocious that it’s often difficult to get nan GPUs.” He said he was “quite concerned astir really being capable to get dependable GPUs erstwhile we want them, and I deliberation this truthful requires that we put a batch much effort connected Dojo successful bid to guarantee that we’ve sewage nan training capacity that we need.”

Dojo was a risky bet, 1 that Musk hedged respective times by saying that Tesla mightiness not succeed.

In nan agelong run, Tesla toyed pinch nan thought of creating a caller business exemplary based connected its AI division, pinch Musk moreover saying during a Q2 2024 net telephone that he saw “a way to being competitory pinch Nvidia pinch Dojo.” While D1 was much tailored for Tesla machine imagination labeling and training – useful for FSD and Optimus training – it wouldn’t person been useful for overmuch else. Future versions would person to beryllium much tailored to general-purpose AI training, said Musk.

The problem that Tesla mightiness person travel up against is that almost each AI package retired location has been written to activity pinch GPUs. Using Dojo chips to train general-purpose AI models would person required rewriting nan software.

That is, unless Tesla rented retired its compute, akin to really AWS and Azure rent retired unreality computing capabilities – an thought that excited analysts. A September 2023 study from Morgan Stanley predicted that Dojo could add $500 billion to Tesla’s marketplace worth by unlocking caller gross streams successful nan shape of robotaxis and package services.

In short, Dojo chips were an security argumentation for nan automaker, but 1 that mightiness person paid dividends.

How acold did Tesla Dojo get?

Musk often provided advancement reports, but galore of his goals for Dojo were ne'er reached.

For instance, Musk suggested successful June 2023 that Dojo had been online and moving useful tasks for a fewer months.” Around nan aforesaid time, Tesla said it expected Dojo to beryllium 1 of nan apical 5 astir powerful supercomputers by February 2024 and had planned for full compute to scope 100 exaflops successful October 2024, which would person required astir 276,000 D1s, aliases astir 320,500 Nvidia A100 GPUs.

Tesla ne'er provided an update aliases immoderate accusation that would propose it ever reached these goals.

Tesla and Musk made galore different pledges for Dojo, including financial ones. For instance, Tesla committed in January 2024 to spend $500 million to build a Dojo supercomputer astatine its gigafactory successful Buffalo, New York, and has already spent $314 cardinal of that, per a 2024 report.

Just aft Tesla’s second-quarter 2024 net call, Musk posted photos of Dojo 1 connected X, saying that it would person “roughly 8k H100-equivalent of training online by extremity of year. Not massive, but not trivial either.”

Despite each of this activity — peculiarly by Musk connected X and successful net calls — mention of Dojo abruptly ended August 2024. And talk switched to Cortex.

During nan company’s fourth-quarter 2024 net call, Tesla said it completed nan deployment of Cortex, “a ~50k H100 training cluster astatine Gigafactory Texas” and that Cortex helped alteration V13 of supervised FSD.

In Q2 2025, Tesla noted it “expanded AI training compute pinch an further 16k H200 GPUs astatine Gigafactory Texas, bringing Cortex to a full of 67k H100 equivalents.” During that aforesaid net call, Musk said he expected to person a 2nd Dojo cluster operating “at scale” successful 2026. He besides hinted astatine imaginable redundancies.

“Thinking astir Dojo 3 and nan AI6 conclusion chip, it seems for illustration intuitively, we want to effort to find convergence there, wherever it’s fundamentally nan aforesaid chip,” Musk said.

A fewer weeks later, he reversed people and disbanded nan Dojo team.

TechCrunch confirmed successful precocious August 2025 that Tesla still plans to perpetrate $500 cardinal to a supercomputer successful Buffalo – it conscionable won’t beryllium Dojo.

This communicative primitively published August 3, 2024. The article was updated for a last clip September 2, 2025 pinch caller accusation astir Tesla’s determination to unopen down Dojo.