On January 18, 2026, Elon Musk announced that development of Tesla’s Dojo 3 supercomputer has officially resumed. The project had been paused through much of 2025 while Tesla concentrated its engineering resources on stabilizing the AI5 chip—an effort Musk described as “existential.” With AI5 now “in good shape,” Tesla is restarting Dojo with a far more ambitious and unified vision.
🧬 Dojo 3 Reborn: A Unified Silicon Strategy #
Dojo 3 represents a fundamental change in Tesla’s approach to AI hardware. Earlier generations split responsibilities between onboard inference chips and the custom D1 accelerator used exclusively for Dojo training. That separation is now gone.
- Single Chip Family: Dojo 3 will be built from clusters of AI5 and future AI6 chips, using the same architecture across vehicles, Optimus robots, and training data centers.
- Cost-Driven Scale: Musk claims AI5-class silicon costs a fraction of NVIDIA’s $30,000+ data center GPUs, enabling far denser and cheaper training clusters.
- Performance Parity: According to Musk, a single AI5 SoC approaches Hopper (H100)–class performance, while dual-chip configurations reach Blackwell-class levels at dramatically lower power.
This unification allows Tesla to optimize one architecture end to end, rather than maintaining separate inference and training ecosystems.
⚙️ AI5 and Dojo 3 at a Glance #
| Feature | AI4 | AI5 (HW5) | Dojo 3 |
|---|---|---|---|
| Performance | Baseline | ~50× AI4 | Exaflop-scale training |
| Process Node | Samsung 4nm | TSMC 3nm / Samsung 2nm | Dense 512-chip clusters |
| Efficiency | Conventional | Industry-leading | Low-latency interconnect |
| Primary Role | Level 2+ FSD | Near-perfect FSD, Optimus | End-to-end NN training |
Tesla positions Dojo 3 not as a niche accelerator, but as the backbone for training every neural network it deploys.
🚀 The Space-Based Compute Vision #
True to form, Musk paired the Dojo announcement with a long-term, unconventional idea: space-based AI compute.
- Orbital Data Centers: Musk argues that space offers near-limitless solar energy and natural heat dissipation, potentially bypassing Earth-bound power constraints.
- Starship Enablement: Such systems would rely on Starship launches, possibly funded by a future SpaceX IPO.
- Skepticism Remains: While vacuum cooling and maintenance pose extreme challenges, Musk views launch dominance as a decisive advantage.
This concept reframes Dojo not just as a terrestrial supercomputer, but as part of a broader off-world infrastructure vision.
🏭 Manufacturing Push and Talent Hunt #
Tesla’s ambitions extend well beyond automotive AI.
- Foundry Strategy: A recently signed $16.5B deal with Samsung positions AI6 production at a Texas fab using a 2nm process.
- Unconventional Hiring: Musk has publicly invited engineers to email Tesla directly with brief summaries of the hardest problems they’ve solved—skipping traditional resumes entirely.
Tesla’s stated goal is to produce the highest-volume AI chips in the world, surpassing even traditional GPU vendors in unit count.
🔎 Why Dojo 3 Matters #
The return of Dojo 3 signals that Tesla’s 2025 pause was not a retreat, but a consolidation. By aligning cars, robots, and supercomputers around a single AI silicon roadmap, Tesla is pursuing a level of vertical integration rarely seen in the semiconductor industry.
If successful, Dojo 3 could redefine how large-scale AI training is built—challenging NVIDIA not just on performance, but on cost, power efficiency, and architectural unity.