Tesla AI5 FSD computer to run inference 10x cheaper than Nvidia AI chips

The next generation Tesla FSD computer has 10x the performance per dollar cost compared to Nvidia's AI chips and up to 3x the performance per watt, according to Elon Musk.

In some parameters, the Tesla AI5 chip beats its current AI4 computer, as found on the 2026 Model Y, by a factor of 40, and is 10x faster overall than Tesla's current hardware that runs FSD 14.

The AI5 chip will be manufactured for Tesla at the Samsung and TSMC foundries in Texas and Arizona, respectively, with production starting in 2026.

Previously leaked AI5 chip specs pegged its performance in the 2000–2500 TOPS (trillion operations per second) range, or about 5x that of the current HW4 computer, but Elon Musk said that he had since spent most weekends working with the AI team to advance the AI5 architecture.

Instead of producing incredibly complex AI chips like Nvidia does for end products like the DGX Spark AI desktop found on Amazon, Tesla bet on stripping unneeded components and features down, since it only has to satisfy the demands of one customer, as per Elon:

If you're not sure how much data is going to go between each logic block on the chip, then you kind of end up having giant highways going all over the place. It's a very -- like it becomes an almost impossibly difficult design problem. And NVIDIA has done an amazing job of dealing with almost an impossibly difficult set of requirements. But in our case, we're going for radical simplicity.

Instead of having a dedicated graphics subsystem, as in the HW4 computer on the 2026 Model Y, Tesla made the whole AI5 chip a GPU for running AI inference and managed to weed out stuff like the image signal processor (ISP). "As a result of these deletions, we can actually fit AI5 in a half reticle and with good margin for the traces from the memory to the Tesla Trip accelerators, the ARM CPU cores and the PCI-X sort of the PCI blocks," tips Elon.

In short, Tesla is doing with its AI5 chip with businesses have begun to do with AI implementation in general. Instead of going for expensive general language models, they are developing solutions tailored to specific tasks that don't need the billions of parameters in an omnipotent LLM like ChatGPT.

Since Tesla only has one customer to take of - itself - and knows that it will be using the AI5 chip only for its self-driving cars and autonomous humanoid Optimus robots, it can develop the whole hardware stack to optimize it for critical points in its software. Even with AI4, however, Tesla promised that cars with FSD 14.2 will feel effortless when it comes to self-driving, as the update will introduce reasoning.

So with reasoning, it's literally going to think about which parking spot to pick, so it's going to say, this is the entrance, but actually, probably there's not a parking spot right at the entrance if it's a full - if the parking lot is fairly full, the probability of an open parking spot right at the entrance is very low. But actually, what it will simply do is drop you off at the entrance of the store and then go find a parking spot. But it's going to get very smart about figuring out a parking spot. It's going to figure out -- it's going to spot empty spots much better than human. It's got a 360-degree vision. And it's going to - yes, like I said, just - it's going to use reasoning to solve things.

It is not clear if the FSD 14.2 version will be able to run on older vehicles, though. Tesla says it will release an FSD 14 Lite for cars with the HW3 computer in Q2, but stayed mum on the features it will exclude. By then, it may have already started work on scaling the AI5 chip production before the Cybercab launch in the April-June period. Any excess AI5 chips Tesla produces, tipped Elon, will go into its FSD data centers to join the Nvidia hardware there.