Tesla Dojo supercomputer restarts 2026 after 2025 shutdown

Tesla announced that its Dojo supercomputer project, which was reported as disbanded in August 2025, has been revived in January 2026 with a next‑generation D1 chip iteration [6][24]. The restart underscores Tesla’s commitment to building a proprietary AI training platform that can handle the massive video data streams from its fleet, a capability it claims will be essential for advancing Full Self‑Driving (FSD).

Production timeline and hardware refresh

Dojo entered production in July 2023, marking the first time Tesla’s custom AI hardware was used for live training of its neural networks [1]. The system’s backbone is the D1 chip, fabricated by TSMC on a 7 nm process, featuring 50 billion transistors on a 645 mm² die [1]. Tesla’s senior Autopilot hardware director has said the architecture will deliver more than an exaflop of performance, a scale that would rival the 1.8 exaflops of Nvidia’s earlier AI‑training cluster [1][30]. After a brief pause—Bloomberg reported the project was disbanded in August 2025—Tesla revived Dojo with a new chip iteration in early 2026, though details of the upgrade have not been disclosed [6][24].

Why the restart matters for Tesla’s AI roadmap

The Dojo revival signals that Tesla still views in‑house compute as a strategic differentiator for its FSD ambitions. By processing “millions of terabytes” of real‑world driving video, Dojo aims to accelerate model training beyond what external cloud providers can offer, potentially shortening the gap between “almost self‑driving” and true full autonomy [1][3]. Competitors such as Nvidia and other AI‑focused chip makers continue to dominate the external training market, but Tesla’s claim of an exaflop‑scale, custom‑designed system could give it a unique data‑centric edge if the hardware delivers the promised throughput.

Competitive context and prior performance

Tesla’s earlier GPU‑based cluster—5,760 Nvidia A100 GPUs—was touted in 2021 as roughly the world’s fifth‑largest supercomputer at about 81.6 petaflops, though the measurement precision (FP32 vs. FP64) was disputed [7][9]. The Dojo architecture diverges from conventional designs by eschewing virtual memory and coherence mechanisms in favor of a high‑speed SRAM mesh, a choice intended to scale more efficiently for video‑heavy workloads [1][4]. If the new D1 iteration improves on the original’s 354 cores per chip and 1.25 MB SRAM per core, Tesla could outpace its own prior GPU cluster and challenge external providers on both speed and energy efficiency.

What to watch

• Hardware rollout – Tesla’s next update on the number of D1 chips and total cabinet count will indicate whether the exaflop target is realistic.
• FSD training milestones – Any disclosed acceleration in neural‑network training cycles or improvements in on‑road performance could validate Dojo’s impact.
• Industry response – Announcements from Nvidia, AMD or emerging AI‑chip startups about custom training solutions may signal competitive pressure on Tesla’s in‑house approach.

Tesla’s decision to resurrect Dojo after a brief shutdown highlights the company’s belief that proprietary, high‑throughput AI training hardware is critical to achieving full self‑driving. Whether the new chip iteration can deliver the claimed performance—and translate it into measurable FSD gains—remains the key question for investors and the broader AI hardware market.