Tesla Seeks to Double AI Chip Production
Tesla is in negotiations with Samsung to more than double production of its AI6 chips, requesting an additional 24,000 wafers per month. The move, reported to bring Tesla's total to 40,000 wafers monthly, reflects surging demand for AI accelerators. This massive expansion highlights the increasing competition for advanced chip foundry capacity as AI workloads scale across the automotive and manufacturing industries.
This production ramp-up is part of Tesla's broader vertical integration strategy, controlling roughly 80% of its supply chain to drive innovation and reduce reliance on external suppliers. This in-house approach extends from raw material sourcing to designing and manufacturing its own chips, batteries, and software. The "AI6" designation follows a line of custom silicon designed to power Tesla's Full Self-Driving (FSD) capabilities. Predecessors like the AI4 chip, found in Hardware 4.0, are built on Samsung's 7nm process, a more mature and cost-effective node compared to cutting-edge alternatives. This contrasts with competitors like NVIDIA, who utilize TSMC's more advanced 4N process for their "Thor" autonomous driving chip. Samsung's 7nm Low Power Plus (LPP) process, the first to incorporate extreme ultraviolet (EUV) lithography, offers significant advantages over older 10nm technology, including up to 20% higher performance or 50% lower power consumption. This efficiency is critical for in-vehicle computers that handle immense data loads from cameras and sensors. Tesla's custom chip development allows for hardware-software co-optimization, tailoring the silicon specifically for its neural network architecture. This is a key differentiator from the "catalog engineering" approach of many traditional automakers who rely on off-the-shelf components. By designing its own chips, Tesla aims to achieve better performance per watt and reduce costs by 20% compared to using third-party hardware. The increased wafer order also supports the immense computational needs of Tesla's Dojo supercomputer, used for training its AI models. The D1 chip, the foundation of Dojo, is a custom 7nm processor with 50 billion transistors, designed by a specialized Tesla team. An ExaPOD, containing 120 Dojo tiles, integrates over one million cores to process the vast amounts of video data collected from its vehicle fleet. Looking ahead, Tesla is already developing its AI5 chip, which is expected to offer a 40x improvement over the AI4 chip by some metrics. Production for the AI5 is slated to involve both Samsung and TSMC. The subsequent AI6 and even AI7 chips are on an ambitious nine-month design cycle, with future applications potentially extending to space-based computing.
