Tesla Dojo Supercomputer — Buffalo New York

Tesla’s Dojo supercomputer at the Gigafactory New York (formerly SolarCity factory) in Buffalo is a purpose-built AI training system designed specifically for training Tesla’s Autopilot and Full Self-Driving (FSD) neural networks on video data from Tesla’s global fleet of 5 million+ vehicles. Dojo uses Tesla’s custom D1 chip — a 7nm training chip with 362 TFLOPS of FP32 performance — packaged into 25-chip training tiles and assembled into ExaPOD configurations. Tesla aims to achieve 100 exaFLOPS of training capacity with Dojo by the end of 2024. Each Tesla vehicle generates roughly 1 terabyte of video per hour; ingesting and training on this data at scale is among the most ambitious AI training infrastructure projects outside the major hyperscalers. Morgan Stanley valued Dojo as a potential $500 billion value creation opportunity for Tesla if commercialized as an AI cloud service.