Tesla has restarted development of its Dojo3 supercomputer leveraging its new AI5 processor architecture that rivals Nvidia's H100 in single-chip performance while consuming 65% less power, marking the company's first fully vertically integrated supercomputing platform.
Tesla has resumed development of its Dojo3 supercomputer project following breakthrough progress on its AI5 processor design, CEO Elon Musk confirmed via social media platform X. The restart signals Tesla's strategic shift toward complete vertical integration in AI infrastructure, eliminating reliance on Nvidia hardware for the first time in its supercomputing initiatives.
According to Musk, the AI5 chip design has reached production-ready status ("good shape"), enabling redeployment of engineering resources to the Dojo3 platform. The AI5 represents Tesla's most advanced processor to date, delivering single-chip performance comparable to Nvidia's Hopper H100 GPU while operating at 250W power consumption—approximately 65% lower than the H100's 700W thermal design power. When configured in dual-chip arrangements, the architecture achieves computational throughput equivalent to Nvidia's next-generation Blackwell GPUs.
This power efficiency translates to significant operational cost reductions. Industry analysis indicates Nvidia's data center GPUs command $25,000-$50,000 per unit on the current market, while Tesla's in-house silicon potentially cuts both acquisition and operational expenses through reduced power and cooling requirements. The restart coincides with Tesla's newly established nine-month chip development cadence, with AI5 entering production imminently followed by AI6 in late 2026 and AI7 in mid-2027.
"Dojo3 will be Tesla's first supercomputer using purely in-house hardware," Musk stated, contrasting with previous Dojo iterations that incorporated Nvidia accelerators. The project's revival follows the cancellation of Dojo2 in 2025 and underwhelming market impact of the initial Dojo1 system, which failed to achieve projected performance benchmarks amid competition from Nvidia's ecosystem.
Technical documentation suggests Dojo3 will implement wafer-scale integration techniques previously explored in canceled initiatives, though Tesla hasn't confirmed specific packaging technology. Musk indicated the supercomputer will eventually support "space-based AI compute" applications, aligning with SpaceX integration possibilities. This positions Dojo3 as a dual-purpose infrastructure investment for both terrestrial AI training and future extraterrestrial computing needs.
Market implications are substantial: A successful Dojo3 deployment would establish Tesla as the first automaker with vertically integrated AI infrastructure spanning silicon design, system integration, and application deployment. This reduces exposure to Nvidia's supply constraints that have impacted global AI development, while creating potential for Tesla to monetize excess compute capacity. Semiconductor analysts note that Tesla's 250W per accelerator power envelope could redefine data center power budgets, potentially influencing next-generation processor designs across the industry.
Tesla is actively recruiting semiconductor engineers to accelerate development, publicly soliciting candidates with experience solving "toughest technical problems" via [email protected]. The company's ability to execute its aggressive nine-month architectural revisions—with AI7 slated for Dojo3 integration—will be closely watched as a benchmark for vertical integration viability in high-performance computing.
Aaron Klotz is a contributing writer for Tom’s Hardware covering semiconductor and computing technologies.
Comments
Please log in or register to join the discussion