Tesla Dojo
Tesla's custom supercomputer for training its self-driving vision models
Tesla Dojo is a series of supercomputers built by Tesla to process video from its cars and train the neural networks behind Full Self-Driving. Announced in 2021 and powered on for production in July 2023, it uses Tesla's own D1 chips rather than standard GPUs. The project was disbanded in 2025 but restarted in 2026.
- Entered production in July 2023 after first being announced in 2021
- Built around Tesla's own D1 chip, made by TSMC on a 7 nm process
- A single ExaPOD holds 1,062,000 cores and reaches 1 exaflop at BF16/CFloat8
- A 2022 test drew 2.3 MW and tripped a San Jose power substation
- Disbanded in 2025, then restarted in January 2026 with a new chip
What it is
Tesla Dojo is a family of supercomputers designed and built by Tesla to handle video recognition work. Its job is to chew through huge volumes of driving footage collected from Tesla vehicles and use it to train the machine learning models behind Full Self-Driving, the company's driver-assistance system. The design is unusual because the workload is millions of terabytes of real-world video rather than typical scientific simulation. Dojo entered production in July 2023. The largest planned unit, an ExaPOD, packs 1,062,000 computing cores and is rated at one exaflop for the low-precision number formats used in AI training.
How it came about
Elon Musk first mentioned Dojo in April 2019 at Tesla's Autonomy Investor Day, and in August 2020 said it was roughly a year away, held up by power and cooling problems. It was formally unveiled at AI Day on 19 August 2021, alongside the D1 chip and a plan for a datacenter of 3,000 of them. A whitepaper followed in October 2021. At the September 2022 AI Day Tesla showed System Trays and one Cabinet, and said a test drew 2.3 MW before tripping a San Jose substation. Production power-on came in August 2023. Bloomberg reported the project was disbanded in August 2025, then restarted in January 2026 with a new chip iteration.
The compute
The building block is Tesla's D1 chip, designed under ex-AMD engineer Ganesh Venkataramanan and made by TSMC on a 7 nm process with 50 billion transistors. Each die carries 354 usable 64-bit cores at 2 GHz and 440 MB of on-chip SRAM, delivering 376 teraflops at BF16 or CFloat8. Twenty-five chips form a water-cooled Training Tile rated at 9 petaflops, six tiles plus a host interface make a System Tray, and ten cabinets form an ExaPOD. Cores are linked by an on-chip mesh, with Dojo Interface Processor cards adding HBM memory and Ethernet. Software runs on PyTorch, and Tesla defined configurable CFloat8 and CFloat16 number formats.
Where the GPUs are
Dojo hardware is assembled and run by Tesla itself. During a 2022 test the system pulled 2.3 MW and knocked out a local power substation in San Jose, California. Tesla was producing one Training Tile per day at that point. Separately, construction at Gigafactory Texas in Austin is for a computing cluster mixing Tesla AI hardware with Nvidia and other chips, planned at 130 MW of thermal design power initially and eventually above 500 MW.
Customers and deals
Dojo is an internal Tesla system rather than a service sold to outside customers. Its purpose is training the neural networks that let Autopilot and Full Self-Driving understand roads, using video clips from Tesla cameras together with object labels, depth and velocity data.
Money and ownership
Musk described Dojo in January 2024 as a long shot worth taking because the potential payoff is high, while putting a low probability on success. No separate funding, revenue or valuation figures for the project are given.
Its place in AI
Dojo sits alongside Tesla's earlier GPU clusters, which used thousands of Nvidia A100 cards for training and automatic labelling. One of those clusters was claimed in 2021 to be around the fifth fastest supercomputer in the world at about 81.6 petaflops, though that figure was disputed because the floating-point precision used was unclear. Dojo itself represents an attempt to replace such GPU clusters with in-house silicon tuned for video workloads.
Questions and criticism
The project has had an uncertain path. Musk called it a low-probability bet, and Bloomberg reported in August 2025 that it had been disbanded before a restart in January 2026. Early problems included power and thermal limits, and a 2022 test overloaded a San Jose substation. The claimed ranking of Tesla's earlier GPU cluster was also questioned over measurement precision.
Questions people ask about Tesla Dojo
What is Tesla Dojo for?
It processes video captured by cameras in Tesla cars and trains the machine learning models behind Full Self-Driving. The aim is to handle millions of terabytes of real driving footage efficiently, which is why the architecture differs from a conventional supercomputer.
What chips does it use?
Dojo uses Tesla's own D1 chip rather than GPUs. Each D1 has 354 usable 64-bit cores, 440 MB of on-chip SRAM and is manufactured by TSMC on a 7 nm process with 50 billion transistors.
How big can a Dojo system get?
Tiles of 25 chips stack into trays and cabinets, and ten cabinets make an ExaPOD with 1,062,000 cores, 1.3 TB of on-tile SRAM and 13 TB of high bandwidth memory, rated at one exaflop for BF16 and CFloat8.
Is Dojo still running?
It entered production in July 2023. Bloomberg reported in August 2025 that the project had been disbanded, but it was restarted in January 2026 with a new chip iteration.
What software does Dojo run?
It supports PyTorch, and the SRAM appears as a single address space. Tesla also created configurable CFloat8 and CFloat16 formats so the compiler can trade precision for speed and smaller storage.
Good reading and sources
- Top four highlights of Elon Musk's Tesla AI DayTechCrunch, Article
- Inside Tesla’s Innovative And Homegrown “Dojo” AI SupercomputerThe Next Platform, Article
- Tesla unveils chip to train A.I. models inside its data centersCNBC, Article
The facts on this page come from the Wikipedia article "Tesla Dojo", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryNeocloud rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.