Tesla Expands In-House AI Chip Design to South Korea
Tesla is expanding its AI chip design team to South Korea, actively recruiting local engineers. The move supports the company's push for a fully end-to-end neural network approach for its Full Self-Driving (FSD) system. This expansion signals a deeper commitment to in-house silicon development and vertical integration for its autonomous systems.
- Tesla's vertical integration for autonomous driving began after its initial reliance on Mobileye for the HW1.0 platform and NVIDIA for HW2.0; the company shifted to fully in-house designed chips with HW3.0 in 2019 to gain comprehensive control over the entire system. This strategy now encompasses roughly 80% of its production value chain, a significantly higher portion than the approximate 55% for legacy automakers. - The in-house chip design initiative was spearheaded by microprocessor engineer Jim Keller, who joined Tesla in 2016 as Vice President of Autopilot Hardware Engineering. Keller, known for his influential work on the AMD K8 and Apple A4/A5 processors, led the development of Tesla's first custom chip before departing in 2018. - Beyond the in-vehicle inference chips, Tesla developed its "Dojo" supercomputer for training neural networks, utilizing its custom D1 chip. Manufactured by TSMC on a 7nm process, the D1 chip features 50 billion transistors and is designed to process the massive video data sets collected from Tesla's fleet for FSD model training. - In a strategic shift in August 2025, Tesla announced it was streamlining its chip development to focus on a unified architecture for both training and inference, rather than separate designs like Dojo. This led to the development of the upcoming AI5 and AI6 chips, which aim to provide optimal performance-per-watt for inference tasks. - South Korea faces a projected shortfall of 50,000 to 56,000 semiconductor workers by 2031, a challenge exacerbated by top students increasingly choosing medical schools over engineering. Despite government initiatives, the admission withdrawal rate for top semiconductor university departments reached 179% in 2024, as many admitted students opt for medical careers instead. - The upcoming AI5 chip is expected to be produced by TSMC, with production potentially shifting to its Arizona plant, while the AI6 chip will be manufactured by Samsung in the U.S. under a $16.5 billion deal. - Tesla's FSD software has evolved significantly, with the V12 update reducing the codebase by approximately 300,000 lines by transitioning control from engineers to the neural network itself. The latest vehicles with Hardware 4 (HW4) and AMD Ryzen processors support the most advanced FSD functionalities.
