Tesla tapes out AI5 chip

Tesla has taped out its AI5 chip, moving the design from engineering files to a foundry run as Elon Musk said AI6 and Dojo3 are already in development. (electrek.co) In chipmaking, “tape-out” is the point when a company finishes the design and sends it to manufacturing for first silicon, not the point when parts are shipping in volume. Tesla’s next step is bring-up and validation, and Electrek reported high-volume production is still targeted for 2027. (electrek.co) Tesla has been building its own artificial intelligence hardware for years because the company runs neural networks in cars and in data centers. Tesla’s AI page says it develops “inference hardware” for Full Self-Driving and robotics, while current job listings describe custom chips for both vehicle and datacenter workloads. (tesla.com, tesla.com) That split matters because the car chip and the Dojo chip do different jobs. Tesla says its vehicle stack depends on efficient inference hardware, while Tesla’s Dojo materials describe a separate system designed for machine-learning compute in Tesla datacenters. (tesla.com, hc34.hotchips.org) Tesla made Dojo public at Artificial Intelligence Day in August 2021, when it showed the D1 training chip and said the system was built to train neural networks on large amounts of video. CNBC’s report from that event said Tesla had already been producing in-house vehicle AI chips for about two years by then. (cnbc.com) The company is still selling Full Self-Driving as a supervised feature, not a hands-off product. Tesla’s support page says Full Self-Driving (Supervised) can drive “almost anywhere” only under driver supervision, which sets the commercial backdrop for any new in-car chip generation. (tesla.com) The new AI5 milestone also arrives as Tesla keeps staffing up around custom silicon. Recent Tesla job posts call for pre-silicon verification, post-silicon validation, compilers, drivers and product test work for AI hardware and Dojo systems. (tesla.com, tesla.com) Musk’s post points to a faster cadence beyond AI5, but tape-out does not settle the harder questions about yield, cost, thermal limits or software readiness. For now, the clearest fact is that Tesla’s custom-silicon program has moved one chip into manufacturing and kept the next designs on the drawing board. (electrek.co, tesla.com)