Tesla AI Inference Chips

A Tesla AI inference chip is the custom processor inside the car that runs Full Self-Driving (FSD), Tesla's camera-based driver-assistance system. Inference just means running a trained neural network to make decisions in real time, which is different from training that network in a data center in the first place. In 2026, Tesla finished designing AI5, its next inference chip, and Elon shared the numbers showing how far it jumps ahead of the chip before it.

What is the Tesla AI Inference Chip?

The chip is the car's on-board brain. It reads the camera feeds and decides steering, braking, and acceleration right there in the car, without phoning home to a cloud connection. The family now runs from Hardware 3 (HW3) to Hardware 4, also called AI4, then the freshly designed AI5, and the planned AI6. The generation a car ships with sets the ceiling on the FSD it can run.

Why the Tesla AI Inference Chip Exists

The effort began in February 2016, when Tesla put together a custom-chip team led by the chip architects Jim Keller and Pete Bannon. What set it off was a need to stand on its own. Mobileye, Tesla's vision-chip partner, ended the relationship in 2016, so Tesla shipped a stopgap Hardware 2 built on an Nvidia Drive PX 2 computer while it designed a chip of its own. Rather than rent a general-purpose chip, Tesla wanted one built just for running neural networks in a moving car. Designing it in-house meant Tesla controlled the cost, the roadmap, and how tightly the chip matched its own software.

How it Works

The first FSD chip went out to be manufactured after roughly eighteen months of work, came back as a working chip, and reached volume in the Model S and Model X in March 2019. Each HW3 board carries two backup FSD chips for safety, and each chip pairs twelve ARM Cortex-A72 processor cores with two systolic arrays, which are grids of multipliers tuned for the exact math that neural networks run on. The jump in raw speed was huge: HW3 processed roughly 2,300 camera frames per second against just 110 for the older HW2.5 board.

The clearest thread across every generation is the process node, which is the size of the smallest features etched onto the chip. Smaller features mean a denser, more efficient chip.

HW3 was built on a Samsung 14-nanometer process in Austin, Texas. AI4 moved to Samsung 7-nanometer. AI6 is set for 2-nanometer. Each shrink packs more computing power into every watt, which matters a lot when the chip runs off a car's battery rather than a wall socket.

The Economics of the Tesla AI Inference Chip

The money logic is simple even without a dollar sign: the number of units, meaning cars plus robots, times the number of generations, times a falling cost for every decision the chip makes. Every shrink bends Tesla's chip cost down while raising the amount of self-driving it can sell. The program anchors a $16.5 billion deal with Samsung, signed July 28, 2025, to build the AI6 chip on a 2-nanometer line at a new Taylor, Texas factory, and Tesla pitched it as one shared design for cars, robots, and AI training. Elon called that deal a floor, not a ceiling. The aim is to drive the cost per decision down generation after generation and lean less on outside chipmakers.

Current Status of Tesla AI Inference Chip

As of mid-2026, the chip actually shipping in new cars is AI4, which has been rolling out since 2023 with twenty Cortex-A72 cores per side. The story across generations is also about memory, which is the working space a chip uses to hold a model and the data it is chewing through.

HW3 carried 8 gigabytes of RAM, and AI4 doubled that to 16 along with four times the storage. More memory lets a car run heavier, smarter FSD models, which is why the hardware generation sets a ceiling on the software a given car can run. HW3 owners run lighter builds than AI4 owners, and that gap frustrates people when older hardware cannot be upgraded to match the newer software. There was also a change at the top: Pete Bannon, who led HW3 and HW4, left Tesla in August 2025.

What Comes Next

AI5 is designed but not yet in cars, and Tesla has not, through an allowed primary source, given a firm ship date for it. Tesla is having it built at two chipmakers at once, TSMC (first in Taiwan, then Arizona) and Samsung, which is a hedge in case any single factory stumbles. When the design was finished, Elon laid out exactly how much ground AI5 gains on AI4.

The biggest leap is in on-chip memory, with raw computing power close behind, and even the smaller gains in useful work and bandwidth add up to roughly five times the useful work of a two-chip AI4 board, all from a single piece of silicon. After AI5 comes AI6, and Elon described "a clear path to a doubling of performance on all metrics for AI6 within 10 to 12 months of AI5 shipping."

The Bottom Line

Tesla turned a supplier breakup into a decade of its own chips that now span cars, robots, and data centers on one family. The hardware in your driveway sets the ceiling on the self-driving you can run, and each new generation raises that ceiling while lowering the cost of every decision the car makes. That is the kind of long game only Elon plays, and keeps winning.

Related

Keep reading: Autopilot, Cybercab. Zoom out to the Tesla overview, or open the Tesla Lineup Explorer.