Tesla tapes out AI5 chip
Tesla reported that its AI5 chip has taped out and promises roughly five times the compute of the AI4 chip, targeting FSD and Optimus workloads with production slated at TSMC and Samsung in 2027. The company framed the chip as a step to scale neural nets for autonomy and humanoid applications. (x.com, x.com)
A car’s driving computer is the part that turns camera video into steering, braking and acceleration in fractions of a second. Tesla said this week it has finished the design of its next-generation AI5 chip and sent it to manufacturers to build. (x.com, electrek.co) In chipmaking, “tape-out” means the blueprint is locked and handed to a foundry for fabrication. It is a milestone, but the chip still has to come back from the factory, be tested in silicon, validated and ramped for volume production. (electrek.co, tesla.com) Elon Musk said AI5 is aimed at Tesla’s Full Self-Driving and Optimus robot workloads, and he thanked both Taiwan Semiconductor Manufacturing Co. and Samsung in separate posts. Tesla’s AI and robotics page says the company is building custom inference chips for vehicles and “bi-pedal robotics and beyond.” (x.com, x.com, tesla.com) The immediate issue for Tesla is scale. Its own site says Full Self-Driving is still “Supervised,” not fully autonomous, and the company’s software stack runs on custom in-car computers that process neural networks from multiple cameras in real time. (tesla.com, tesla.com, tesla.com) Tesla says those neural networks are getting larger: the company says a full build of its self-driving networks involves 48 networks, 70,000 graphics-processing-unit hours of training, and 1,000 predictions at each time step. More capable in-car chips let Tesla run bigger models with tighter power and heat limits than a data-center server can tolerate. (tesla.com) That helps explain why Tesla is talking about AI5 now, before it is in customer cars. Tesla reported first-quarter 2026 deliveries of 358,023 vehicles on April 2, and it is scheduled to report full quarterly results on April 22, when investors are likely to ask how quickly its autonomy and robotics spending turns into products. (ir.tesla.com, nasdaq.com) The timeline has already slipped. Musk said in June 2024 that AI5 would arrive in vehicles in the second half of 2025, but Electrek reported on April 15 that volume production is still more than a year away and that Tesla had previously pushed that target to mid-2027. (electrek.co, electrek.co) That gap matters for Tesla’s current lineup. Electrek reported that Tesla’s Cybercab is set to launch on existing AI4 hardware, and Tesla’s support pages show the company is still managing earlier computer generations through upgrades for some owners who bought Full Self-Driving. (electrek.co, tesla.com) Tesla’s critics argue the company keeps moving deadlines on self-driving hardware and software. Supporters point to the tape-out as proof that Tesla is still investing in a custom silicon path instead of relying on off-the-shelf processors from suppliers such as Nvidia. (electrek.co, tesla.com) What happens next is less glamorous than the announcement: first silicon, validation, and a manufacturing ramp that has to meet automotive reliability standards. If that process stays on schedule, AI5 becomes the next test of whether Tesla can turn bigger neural networks into a wider real-world rollout of supervised driving and humanoid robots. (electrek.co, tesla.com, tesla.com)
