Tesla aims to shorten its in-house AI chip design cycle to one generation every nine months, targeting rivals Nvidia and AMD. However, industry analysts highlight automotive safety verification and software stability as the biggest bottlenecks.
According to Bloomberg, Tesla CEO Elon Musk announced on X that the AI5 chip design is nearing completion, prompting Tesla to restart Dojo3 R&D. Before being abruptly terminated in 2025, the Dojo program had focused on building an in-house AI supercomputer to support the development of self-driving technology and was once regarded as a core, multibillion-dollar effort for Tesla to gain an edge in the AI race.
Early development of the AI6 chip has already begun, and Musk revealed plans for subsequent AI7, AI8, and AI9 chips, all following a nine-month design cadence. He disclosed in 2025 that Samsung Electronics will manufacture the next-generation AI6 chip under a US$16.5 billion contract signed between the two companies.
Musk emphasized Tesla's ongoing recruitment of engineering talent with the goal of producing the world's highest-volume AI chip shipments. Compared to data center chips, automotive-grade chips face higher safety thresholds, lengthier validation processes, and stricter regulatory requirements. This has historically slowed Tesla's chip development pace relative to Nvidia and AMD.
Autonomous driving and advanced driver-assistance systems (ADAS) must comply with functional safety standards such as ISO 26262, pass scenario-based testing, road permits, safety of the intended functionality (SOTIF), cybersecurity, and software update mandates.
Tom's Hardware analyzed that achieving a nine-month design cycle per generation is feasible only through a platform-based, incremental approach, reusing core architectures, programming models, memory hierarchies, and security frameworks. Adjustments would be limited to compute scaling, static random-access memory (SRAM) tuning, or process node transitions. Introducing new memory types, programming models, or security architectures would inevitably extend timelines.
Article edited by Jack Wu
