Tesla’s Dojo, A Timeline

Trending 2 days ago

Elon Musk doesn’t want Tesla to beryllium conscionable an automaker. He wants Tesla to beryllium an AI company, 1 that’s figured retired really to make cars thrust themselves. 

Crucial to that ngo was Dojo, a custom-built supercomputer designed by Tesla to train its Full Self-Driving (FSD) neural networks. FSD isn’t really afloat self-driving; it tin execute immoderate automated driving tasks, but still requires an attentive quality down nan wheel. But Tesla thinks pinch much data, much compute powerfulness and much training, it tin transverse nan period from almost self-driving to afloat self-driving. 

And that’s wherever Dojo was expected to travel in.

Musk teased Dojo for years and ramped up discussions astir nan supercomputer passim 2024. But Dojo is now out, and different supercomputer called Cortex has entered nan chat.

Below is simply a timeline of Dojo mentions and promises. Check retired this explainer connected what Dojo for moreover much accusation connected what it was, why it mattered, and what comes next.

2019

First mentions of Dojo

April 22 – At Tesla’s Autonomy Day, nan automaker had its AI squad onstage to talk astir Autopilot and Full Self-Driving, and nan AI powering them both. The institution shares accusation astir Tesla’s custom-built chips that are designed specifically for neural networks and self-driving cars. 

During nan event, Musk teases Dojo, revealing that it’s a supercomputer for training AI. He besides notes that each Tesla cars being produced astatine nan clip would person each hardware basal for afloat self-driving and only needed a package update.

Techcrunch event

San Francisco | October 27-29, 2025

2020 

Musk originates nan Dojo roadshow

Feb 2 – Musk says Tesla will soon person much than a cardinal connected vehicles worldwide pinch sensors and compute needed for afloat self-driving — and touts Dojo’s capabilities. 

“Dojo, our training supercomputer, will beryllium capable to process immense amounts of video training information & efficiently tally hyperspace arrays pinch a immense number of parameters, plentifulness of representation & ultra-high bandwidth betwixt cores. More connected this later.”

August 14 Musk reiterates Tesla’s scheme to create a neural web training machine called Dojo “to process genuinely immense amounts of video data,” calling it “a beast.” He besides says nan first type of Dojo is “about a twelvemonth away,” which would put its motorboat day location astir August 2021.

December 31 Elon says Dojo isn’t needed, but it will make self-driving better. “It isn’t capable to beryllium safer than quality drivers, Autopilot yet needs to beryllium much than 10 times safer than quality drivers.”

2021

Tesla makes Dojo official

August 19 – The automaker officially announces Dojo astatine Tesla’s first AI Day, an arena meant to pull engineers to Tesla’s AI team. Tesla besides introduces its D1 chip, which nan automaker says it will usage — alongside Nvidia’s GPU — to powerfulness nan Dojo supercomputer. Tesla notes its AI cluster will location 3,000 D1 chips. 

October 12 – Tesla releases a Dojo Technology whitepaper, “a guideline to Tesla’s configurable floating constituent formats & arithmetic.” The whitepaper outlines a method modular for a caller type of binary floating-point arithmetic that’s utilized successful heavy learning neural networks and tin beryllium implemented “entirely successful software, wholly successful hardware, aliases successful immoderate operation of package and hardware.”

2022

Tesla reveals Dojo progress

August 12 – Musk says Tesla will “phase successful Dojo. Won’t request to bargain arsenic galore incremental GPUs adjacent year.”

September 30 – At Tesla’s second AI Day, nan institution reveals that it has installed nan first Dojo cabinet, testing 2.2 megawatts of load testing. Tesla says it was building 1 tile per time (which is made up of 25 D1 chips). Tesla demos Dojo onstage moving a Stable Diffusion exemplary to create an AI-generated image of a “Cybertruck connected Mars.”

Importantly, nan institution sets a target day of a afloat Exapod cluster to beryllium completed by Q1 2023, and says it plans to build a full of 7 Exapods successful Palo Alto. 

2023

A ‘long-shot bet

April 19 – Musk tells investors during Tesla’s first-quarter earnings that Dojo “has nan imaginable for an bid of magnitude betterment successful nan costs of training,” and besides “has nan imaginable to go a sellable work that we would connection to different companies successful nan aforesaid measurement that Amazon Web Services offers web services.”

Musk besides notes that he’d “look astatine Dojo arsenic benignant of a long-shot bet,” but a “bet worthy making.”

June 21 The Tesla AI X relationship posts that nan company’s neural networks are already successful customer vehicles. The thread includes a chart pinch a timeline of Tesla’s existent and projected compute power, which places nan commencement of Dojo accumulation astatine July 2023, though it’s not clear if this refers to nan D1 chips aliases nan supercomputer itself. Musk says that aforesaid time that Dojo was already online and moving tasks astatine Tesla information centers. 

The institution besides projects that Tesla’s compute will beryllium nan apical 5 successful nan full world by astir February 2024 (there are nary indications this was successful) and that Tesla would scope 100 exaflops by October 2024.

July 19 – Tesla notes successful its second-quarter net study it has started accumulation of Dojo. Musk besides says Tesla plans to walk much than $1 cardinal connected Dojo done 2024.  

September 6 – Musk posts on X that Tesla is constricted by AI training compute, but that Nvidia and Dojo will hole that. He says managing nan information from nan astir 160 cardinal frames of video Tesla gets from its cars per time is highly difficult. 

2024

Plans to scale

January 24 – During Tesla’s fourth-quarter and full-year net call, Musk acknowledges again that Dojo is simply a high-risk, high-reward project. He besides says that Tesla was pursuing “the dual way of Nvidia and Dojo,” that “Dojo is working” and is “doing training jobs.” He notes Tesla is scaling it up and has “plans for Dojo 1.5, Dojo 2, Dojo 3 and whatnot.”

January 26 – Tesla announced plans to walk $500 cardinal to build a Dojo supercomputer successful Buffalo. Musk past downplays nan finance somewhat, posting connected X that while $500 cardinal is simply a ample sum, it’s “only balanced to a 10k H100 strategy from Nvidia. Tesla will walk much than that connected Nvidia hardware this year. The array stakes for being competitory successful AI are astatine slightest respective cardinal dollars per twelvemonth astatine this point.”

April 30 – At TSMC’s North American Technology Symposium, nan institution says Dojo’s next-generation training tile — nan D2, which puts nan full Dojo tile onto a azygous silicon wafer, alternatively than connecting 25 chips to make 1 tile — is already successful production, according to IEEE Spectrum. 

May 20 – Musk notes that nan rear information of nan Giga Texas mill hold will see nan building of “a ace dense, water-cooled supercomputer cluster.”

June 4 – A CNBC report reveals Musk diverted thousands of Nvidia chips reserved for Tesla to X and xAI. After initially saying nan study was false, Musk posts connected X that Tesla didn’t person a location to nonstop nan Nvidia chips to move them on, owed to nan continued building connected nan southbound hold of Giga Texas, “so they would person conscionable sat successful a warehouse.” He noted nan hold will “house 50k H100s for FSD training.”   

He besides posts: 

“Of nan astir $10B successful AI-related expenditures I said Tesla would make this year, astir half is internal, chiefly nan Tesla-designed AI conclusion machine and sensors coming successful each of our cars, positive Dojo. For building nan AI training superclusters, NVidia hardware is astir 2/3 of nan cost. My existent champion conjecture for Nvidia purchases by Tesla are $3B to $4B this year.”

July 1 – Musk reveals connected X that existent Tesla vehicles whitethorn not person nan correct hardware for nan company’s next-gen AI model. He says that nan astir 5x summation successful parameter count pinch nan next-gen AI “is very difficult to execute without upgrading nan conveyance conclusion computer.”

Nvidia proviso challenges

July 23 – During Tesla’s second-quarter net call, Musk says request for Nvidia hardware is “so precocious that it’s often difficult to get nan GPUs.” 

“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,” Musk says. “And we do spot a way to being competitory pinch Nvidia pinch Dojo.”

A chart successful Tesla’s investor platform predicts that Tesla AI training capacity will ramp to astir 90,000 H100 balanced GPUs by nan extremity of 2024, up from astir 40,000 successful June. Later that time connected X, Musk posts that Dojo 1 will person “roughly 8k H100-equivalent of training online by extremity of year.” He besides posts photos of nan supercomputer, which appears to usage nan aforesaid fridge-like stainless alloy exterior arsenic Tesla’s Cybertrucks. 

From Dojo to Cortex

July 30 –  AI5 is ~18 months distant from high-volume production, Musk says successful a reply to a station from personification claiming to commencement a nine of “Tesla HW4/AI4 owners angry astir getting near down erstwhile AI5 comes out.” 

August 3 – Musk posts connected X that he did a walkthrough of “the Tesla supercompute cluster astatine Giga Texas (aka Cortex).” He notes that it would beryllium made astir of 100,000 H100/H200 Nvidia GPUs pinch “massive retention for video training of FSD & Optimus.”

August 26 – Musk posts connected X a video of Cortex, which he refers to arsenic “the elephantine caller AI training supercluster being built astatine Tesla HQ successful Austin to lick real-world AI.” 

2025

Dojo shutdown, its squad disbanded

January 29 – Tesla’s Q4 and full-year 2024 net call included nary mention of Dojo. Cortex, Tesla’s caller AI training supercluster astatine nan Austin gigafactory, did make an appearance, however. Tesla noted successful its shareholder deck that it completed nan deployment of Cortex, which is made up of astir 50,000 H100 Nvidia GPUs. 

“Cortex helped alteration V13 of FSD (Supervised), which boasts awesome improvements successful information and comfortableness acknowledgment to 4.2x summation successful data, higher solution video inputs … among different enhancements,” according to nan letter. 

During nan call, CFO Vaibhav Taneja noted that Tesla accelerated nan buildout of Cortex to velocity up nan rollout of FSD V13. He said that accumulated AI-related superior expenditures, including infrastructure, “so acold has been astir $5 billion.” In 2025, Taneja said he expects capex to beryllium level arsenic it relates to AI.

July 23 – During Tesla’s Q2 2025 net call, Musk said Dojo 2 was expected to beryllium “operating astatine scale” sometime successful 2026, pinch “scale being location astir 100k H100 equivalent.” In nan aforesaid breath, he 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. 

July 28 – 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. 

August 6 – Bloomberg reports that adjacent to 20 Dojo workers near to commencement their ain institution that builds AI chips, software, and hardware called DensityAI. 

August 7 – Bloomberg reports that Tesla has disbanded its Dojo squad and unopen down nan project. Peter Bannon, Dojo’s lead, near nan company, arsenic well. 

Musk responded to nan reports connected X, saying: “It doesn’t make consciousness for Tesla to disagreement its resources and standard 2 rather different AI spot designs. The Tesla AI5, AI6 and consequent chips will beryllium fantabulous for conclusion and astatine slightest beautiful bully for training. All effort is focused connected that.”

August 10 – “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.”

September 1 – Tesla shares its Master Plan Part IV connected societal media level X. There is nary mention of Dojo aliases Cortex, though AI, and much specifically, “physical AI” takes a cardinal role.

This communicative primitively published August 10, 2024. The last update of nan Tesla Dojo timeline published September 2, 2025.

More