Tesla plans to restart work on its Dojo3 supercomputer project after making strong progress on its next-generation AI5 chip, according to recent comments from Chief Executive Elon Musk. The move comes months after the company halted its earlier Dojo effort and disbanded much of the original team.
Dojo project was halted In 2025
Tesla paused development of its initial Dojo supercomputer and D‑series chips in August 2025. At that time, the company had shut down the Dojo group, reassigned engineers, and stepped back from treating Dojo as a fully separate platform from its main vehicle and robotics hardware.
The original Dojo program centered on the custom D1 chip and large training cabinets aimed at processing vast video data from Tesla vehicles. That data supports neural networks used for Autopilot, Full Self-Driving and early work on the Optimus humanoid robot. Yet internal changes, staff departures, and rising investments in more flexible AI hardware appeared to shift management focus.
During and after the 2025 pause, Tesla increased its reliance on major chip manufacturers and high-volume AI hardware from outside vendors. Company disclosures pointed to large-scale deployments of Nvidia GPUs for training, along with deeper ties to foundry partners such as TSMC and Samsung.
Musk described Dojo’s original direction as an “evolutionary dead end” and indicated that future AI systems would converge around a new family of Tesla-designed chips rather than a one-off supercomputer line. As a result, many investors concluded in 2025 that Tesla’s AI future would sit on shared platforms using standard boards packed with its own chips and third-party accelerators.
AI5 emerges as core hardware platform
The picture shifted again as Tesla advanced work on its AI5 chip. AI5 is expected to succeed the company’s current AI4 (Hardware 4) platform and is described by Musk as delivering up to ten times the computing performance of AI4. The chip is intended to support future Full Self-Driving versions, Optimus robotics capabilities, and other internal AI workloads.
Tesla plans two process variants of AI5, and one produced by TSMC on a 3 nm node and another from Samsung on a 2 nm process. Both versions are meant to be functionally equivalent, even though they use different manufacturing technologies. Company timelines point to volume production in the second half of this decade.
Musk links Dojo3 restart to AI5 progress
Musk has now tied the restart of Dojo3 directly to the maturity of the AI5 design. In recent posts, he said the AI5 chip is “almost done” and “in good shape,” and he used the update to invite engineers to join Tesla’s AI hardware efforts.
Dojo3, in this new phase, appears less like a standalone experimental supercomputer and more like a dense cluster of standardized Tesla AI chips. Musk has spoken about building boards packed with large numbers of AI5 and future AI6 chips to cut networking and cabling overhead. That layout would serve as a practical successor to the original Dojo system but rely on the same chip families used in vehicles and robots.
Musk has outlined an aggressive roadmap that stretches beyond AI5. He has mentioned AI6 already in early development, and has spoken of future generations dubbed AI7, AI8 and AI9, with ambitions for fast design cycles between each version.
Even with the restart of Dojo3, key questions remain. Investors are watching to see how much of Tesla’s AI compute continues to run on Nvidia-based clusters and how quickly in-house systems can reach scale. Previous forecasts once imagined Dojo as a major profit driver that could sell AI compute services to outside customers, but that scenario never materialized before the 2025 pause. Now, expectations appear more cautious, with emphasis on supporting Tesla’s own software, autonomy, and robotics plans.
For Tesla, the renewed Dojo3 effort marks a return to a more vertically integrated AI hardware vision, anchored by AI5 and its successors but still linked to global chipmakers through foundry deals. Will the new approach deliver reliable, cost-effective compute at the scale the company seeks? For now, that remains an open question.
You may also like:
- Tesla patent offers new life for HW3 FSD computers »
- Musk sets aggressive 9‑month cycle for Tesla’s next AI chips »
- Tesla’s new Texas lithium refinery starts operations »
We’ve launched an official Tesla Owners Online Forum. Join early and claim your "Founding Fuel" badge.
