AI5

[A J]

Hey Folks,

It looks like Tesla is customizing its chips to handle smaller AI models.

I imagine this is useful for both robots and autonomous vehicles.

Does anybody know if the Tech Stack is open source?

AI Overview

Tesla's next-generation AI5 (HW5) chip targets 2,000 to 2,500 TOPS of total raw integer compute performance, representing an 8x raw compute increase over AI4. Rather than relying on traditional high-precision floating-point (FP) configurations, the architecture heavily optimizes for low-precision INT/FP mix formats—specifically leveraging dense 8-bit, 4-bit, and 2-bit Multiply-Accumulate (MAC) tensor operations to maximize edge-inference efficiency.

[Chris Albertson]

> On Aug 11, 2026, at 8:16 AM, A J <aj48...@gmail.com> wrote:
>
> Hey Folks,
>
> It looks like Tesla is customizing its chips to handle smaller AI models.
>
> I imagine this is useful for both robots and autonomous vehicles.
>
> Does anybody know if the Tech Stack is open source?

It is funny how today we call a 10 Billion parameter model “smaller”. It is tiny for an LLM but “over the top huge" for an RL trained robot. Cars and Humanoids to NOT use LLMs.


You can see the open source Tesla uses by looking at their GitHub repository. GPL requires them to provide source code, and they do. But if you look, exactly zero of what is there has to do with AI. All the AI stuff is custom C++ code they write in-house.

We do know a little from interviews. Tesla uses a “compiler”. All of us normal folks, including people who work at OpenAI or Google, write in Python and make calls to PyTorch, and then PyTorch makes calling into a library that is written in C++ and that inturn makes calls to some Linear Algebra libray and that in turn calls a CUDA or “whatever” back end. Tesla skips all that. They write PyTorch-like code that gets compiled directly to the binary code that runs on the chips. There are no layers or libraries. This is much more efficient but very inflexible, and you can only use it if you have a Tesla computer. (or whatever else the compiler targets)

The AI computer itself does NOT run Linux. It runs a custom Tesla-written real-time OS that is apparently written in the MISRA subset of C. MISRA says (1) no dynamic memory allocation, 2) no pointer arithmetic, 3) no unbounded loops, 4) every type cast is explicit, and quite a lot of other things that seriously dumb down C so that you can PROVE things about the code. Tesla did not invent MISRA; it is used universally in the automotive industry for things like the OS that runs on engine controllers.

The tablet the driver uses does run Debian Linux, the touch screen is driven by Wayland and there is even a BussyBox shell a technician can SSH into. But this computer does not drive the car.

The other stuff we know is because Tesla has filed some patents, and all patents are public. So they had to explain their “end-to-end neural network”. Basically, it is a handful of networks, some for reading road signs, some for vision, some for car control, but the description is VERY high level, just one simplified block diagram that is kind of vague with details.


I brought up this subject only to point out that “robot AI” is very different from “ChatGPT AI”. A car or humanoid never needs to answer random questions like “Compare Picasso to DaVinci”; it only has to output motor commands and needs 1,000 times less general knowledge. So it can be smaller.

Today, every hobby robot builder who wants to build a robot with arms or legs is going to have to know how to build these kinds of robot AIs, pretty much from scratch. Or if he can’t do that, he can only build “kit robots” from a very exact parts list and then install “black magic” software that he can never change. I don’t see other options. But the good news is that these robot AIs are not huge and can run on $500 computers that most people already own.

While the new AI5 computer will be 2,000 TOPS, it does not currently exist. The current Optimus humanoid robot can walk and move its hands using “4" that is only a few hundred TOPS. And the cars can drive every well now on even the older version 3.

My guess is that you don’t need even 10 TOPS for basic robot operation, but I can’t prove that yet.