Tesla completes design review for next-gen AI chips

Elon Musk says Tesla has finished a long design review for the AI5 chip with engineers in California and Texas. This chip will be 40 times faster than the AI4 used in Tesla cars and robots. Production will start at Samsung’s plant in Texas and TSMC’s fab in Arizona. Mass production is planned for 2026 or early 2027. Right after AI5, AI6, AI7, and AI8 chips are coming soon. Musk says AI8 will be “out of this world”.​

AI5 chip

Tesla’s AI5 chip will make cars and robots work better by solving tasks faster and using less power. It has eight times more raw compute power and nine times more memory than the previous chip. Memory bandwidth is five times higher, helping Tesla’s Full Self-Driving system make decisions quickly. Since Tesla only needs the chip for its own products, the design skips extra parts, like separate GPUs, that aren’t needed. That choice makes the chip simpler and more focused. Musk said, “We know what the chip needs to do, and just as importantly, we know what the chip doesn’t need to do”.​

AI5 uses a mix of tensor accelerators, DSPs, and scalar controllers to manage lots of different AI workloads. The chip has a strong memory system, with fast caches and a memory controller that saves power when it can. Tesla uses lower-precision math in the chip for vision and perception, doubling the speed without hurting accuracy. The chip runs at up to 3 Teraflops per watt, and up to 2500 TOPS (trillion operations per second) is expected. That’s about five times more than before. It costs less and uses less power per inference than Nvidia chips. Musk said, for AI tasks, this chip could be “the lowest cost silicon and best performance per watt”.​

Chip manufacturing plan and timeline

Tesla will build the AI5 chip at both Samsung and TSMC. TSMC starts in Taiwan before switching to Arizona. Samsung’s Texas plant has slightly newer equipment. Making chips at two sites reduces risk and supports steady supply. The combined output is expected to cover needs for up to three million vehicles each year. Initial chips will be used in the Cybercab robotaxi. Later, cars and Optimus robots will run on these chips. Tesla wants to have more chips than it needs, using the extras in its data centers.​

Tesla signed a $16.5 billion contract for Samsung to make the AI6 chip, which will handle both inference and training. AI6 could beat AI5 on real-world tasks and cuts the need for separate AI hardware. While AI7 and AI8 specs are still secret, Musk says they’re coming quickly, hinting that AI8 may be used off-Earth or is just a huge leap.​

End of Dojo

In August, Tesla shut down its Dojo supercomputer team and moved those engineers to focus on AI5 and AI6 chips. The company decided it was better to make a chip that supports both inference and training, rather than building two different chips. Dojo’s lead left, and some workers joined a new startup, DensityAI. Tesla can update products faster and save money by working on one chip platform for everything.​

The chip powers both Tesla vehicles with Full Self-Driving and Optimus robots. Tesla will use the same processor for robots doing factory work and data center training tasks. That lets the company improve its neural network models for all systems at once and use its manufacturing strengths to get benefits from scale.​

Optimus robot update

Musk confirmed that the Optimus robot demo review will happen after the chip design review. The robot is advancing quickly. Tesla engineers have Optimus Gen 2.5 models working nonstop at their Palo Alto lab. Gen 3 will launch in early 2026 and be able to do more tasks with better hand movement. The robot learns from watching people and requires much less teaching time. It can now balance and move well without anyone controlling it directly. Tesla will use these robots at its own factories in 2025, and aims for larger production later. Musk expects the robot to bring big changes to healthcare and daily life.​​

Only a few companies, like Apple, design chips for both consumers and data centers. Tesla wants to rely less on Nvidia and more on its own hardware. But Musk said Tesla will still use Nvidia GPUs for model training and use its own chips mainly for real-time tasks in vehicles and robots. At least for now, Tesla combines custom chips and Nvidia hardware for the best results.​

Tesla is betting big on its own chip designs to make cars and robots smarter and cheaper. Musk says the chip performs “better per dollar than anything else,” especially for AI models under 250 billion parameters. The company is using a simple design approach focused only on its needs. Chips will be made in two American plants. With AI5 about to roll out, and AI6, AI7, and AI8 moving up next, Tesla is changing fast from making only cars to building advanced AI tools for robots and factories.

You may also like to read:

• Tesla to launch Optimus V3 robot in early 2026 »
• Tesla unveils AI5 chip, 40 times faster than AI4 »
• Tesla Roadster may launch with flying tech before 2026 »