Tesla Just Finished AI5 Chip--But It Won't be in your Car!

Added: 2026-05-05

[00:00:00]For the past few years, when we've talked about AI progress, we've mostly been talking about data centers. Bigger clusters, more GPUs, more power, more scale, intelligence in that framing lives somewhere far away from us inside massive server farms. And we just kind of tap into it when we need it. But what Elon Musk just revealed about AI5 and AI6 points in a completely different direction. These chips are not designed for data centers. Rather, they're designed to run intelligence at extremely high levels in real time under brutal power constraints inside cars, inside robots, inside systems that don't get to pause, don't get to offload to the cloud, and don't get a second chance if they're wrong. Once you start looking at Tesla's chip architecture through that lens, memory, bandwidth, SRAMM, LPDDR6, and trip accelerators, you realize Tesla isn't just building faster chips.

[00:00:48]They're building a completely different kind of intelligence system. But if you expect AI5 in your car next year, Elon just said you might be disappointed. So, let's take a look. Before we start, a quick shout out to my channel sponsor Joah. They make amazing accessories for your Tesla and other EVs and also have great warranties and customer service, too. In fact, I use their accessories daily. Be sure to check the link in the description to get 5% off a fan cooled phone charger, a portable tire inflator, a foldout lap table, and a lot of other items. Oh, and they make perfect gifts for you and your Eevee loving friends, too. So, be sure to check out Joah. And now, let's get back to it. Hey y'all, it's Dr. Noah. Also, yeah, I want to talk about a couple of recent posts from Elon Musk that kind of set things out and let us know a lot more about AI5, AI6, the status of where they are, some of the architecture of the chips themselves, and the fact that we probably won't see AI5 in a car anytime soon.

[00:01:41]So, starting with this post from yesterday, April 15th, as I record this, Elon Musk said, "Congrats to the Tesla AI chip design team on taping out AI5."

[00:01:50]That means that they have completed the design work. They're ready to ship it off to Samsung and TSMC to actually start to build these chips. Of course, there's an initial production run.

[00:01:59]There's a lot of stuff to ramp it up to full scale production, but this is a huge milestone because that means it's actually finished. The design is done, and now they can focus on ramping this up and producing a lot of chips.

[00:02:09]finishing his original post, which also has a picture of the taped out version of AI5 there. I assume that's what it is. Doesn't actually say that, but I'm assuming that's what it is. Anyway, it says AI5, Dojo 3, and other exciting chips are in the works. So, Dojo 3 is interesting because Dojo 3 is likely to be, I believe, if I'm remembering correctly. He said that was about AI7 when we're going to see the edge compute chips. In other words, the AI stuff that's in your car and is going to be an Optimus eventually. We'll see that converge with their dojo project, which is their large data center, their large training center, which obviously operates in a data center, not on your car, but they're going to use the same chips. We also just found out that there's going to be an AI 6.5 as well as AI6. So, there will be stages in between. We'll say AI5 this year in 2026. We'll likely see AI 6 in 2027, probably AI 6.5 in 2028 now that there's a new AI 6.5 that he's talking about, and then maybe 2029 for AI7, which is where dojo 3 comes in. Or maybe it'll come in with 6.5. Anyway, some point in the future, we will see these things converge.

[00:03:12]And then we get more details from Elon.

[00:03:13]This is the best post of all of them. In response to Tesla Economics saying, "What was the most awesome thing and least awesome thing about taping out AI 5?" Congrats, by the way. Huge milestone. Elon said, "Best was working with such a great team of AI hardware and software engineers." So, clearly, they're designing this together. They're co-designing this. The idea is not to use generic like Nvidia GPUs and things like that off the shelf and have to write their software to it, but to write the software and the hardware to engineer all of that together to get the most out of their chips. And we're going to see that in just a second as he continues talking here. It was more fun than going to parties on Saturday night by far. I don't know. I never go to parties on Saturdays anymore. I would probably enjoy that, too, honestly. And then he said, "The least awesome thing was that we had to make several design concessions to move fast, but we were able to finish tape out 45 days ahead of schedule. So that means we're probably looking at around June 1st as their goal, and they actually taped it out early." And then he gets on to the design concessions and what AI6 will do.

[00:04:06]So AI6 with LP DDR6 memory addresses those design concessions and has many new great ideas. It will deliver a true doubling of performance over AI5 in the same half reticle size. In other words, half of one of those wafers using the Samsung 2nanmter FAB in Texas AI 6.5, this is the new one, will further improve performance using TSMC's 2nmter process in Arizona. The Samsung factory, as far as I know, is actually already up and operational, at least quasi operational. The TSMC fab on the other hand, I don't know exactly what its status is, but that's probably why we're looking at 6.5. Note, both chips have about half of the trip AI computation accelerators dedicated to SRAMM. So effective memory bandwidth is an order of magnitude greater than DRAM bandwidth for any calculations in SRAMM cache.

[00:04:53]All right, so there's a lot of stuff in this one reply. So let's break things down. Let's start with LPDDR. That's low power double data rate. It's a type of synchronous dynamic random access memory or SD RAM that's designed to use less power than conventional memory and it's commonly used in things like smartphones, tablets, computers, laptops and apparently not in Tesla vehicles themselves. If he's talking about LPDDR, it means that it's not in AI4 and certainly not in hardware 3. So, it has two advantages. Number one, it's low power, which is great for portable devices. Memory is a huge issue. We're going to get back to that because Elon talks about using their trip accelerator for memory itself, which is a really interesting thing. But then the other aspect is it's double data rate, which means it's able to push a lot of data through it very power efficiently. And that's particularly important for navigating the real world. In other words, full self-driving or Optimus moving around in the world and such because what you're doing is taking in millions of pixels of information every single frame. In other words, 36 times a second at the refresh rate of the cameras. You're taking in all of this pixel information. That's a gigantic amount of memory bandwidth that you have to shove through the pipeline very very rapidly and very power efficiently. So LPDDR is going to be a really big unlock to be able to allow the system to work even faster and more power efficiently.

[00:06:07]And so as Elon says that's going to allow a true doubling of performance over AI5. So AI5 itself doesn't have LPDDR yet either. And that'll be in the same half reticle size. In other words, the same size of chip as AI5. And of course, it'll use Samsung's 2n fab in Texas and then eventually the TSMC 2nmter fab in Arizona. So these are the fabs that are literally pushing the limits of physics to get down to almost atomicized transistor gates and squeeze more and more and more of them into individual chips. So this is really going to push the limits.

[00:06:40]Then we get to something that was even more interesting to me. This revelation is really fascinating. Both the chips AI6 and AI 6.5 I assume will have about half of the trip AI computation accelerators dedicated to SRAMM. So effective memory bandwidth is an order of magnitude greater than DRAM bandwidth for any calculations in the SRAM cache.

[00:06:59]That is a lot of stuff in one sentence.

[00:07:01]So first of all, what is the trip AI computation accelerator? Well, you'd think it would be an acronym and probably it is, but I couldn't find it.

[00:07:08]I even asked Grock and Grock couldn't find what the acronym actually stood for, but it's Tesla's internal code name for their custom neural network accelerators. So, this is the same sort of thing as the tensor core in Google's TPUs or tensor processing units or the neural engine in Apple's iPhones and their M series chips for their laptops and their desktops. So, these are specialized hardware blocks that are optimized for the heavy math of neural networks. They're specifically to do that. In other words, a lot of matrix multiplication and addition. And here's the really interesting part. As Elon said, in the AI6 chips and potentially in the AI5 chips as well, it's a little bit unclear. Roughly half of these trip accelerators are not dedicated to do raw math. In other words, they're not getting something and they're doing the calculations, but they're dedicated to onchip SRAMM. And the SRAM is the really fast RAM that's actually on chip. In other words, close to the GPU and the CPU. all of the calculations that are going on. That's the memory that can be accessed really really quickly by CPUs and GPUs and they don't have to hang around waiting for it to come from DRAM from offchip. And so it appears what's happening here is that you're going to use these trip accelerators, these math accelerators almost to calculate things coming out of the SRAM before it even gets to the CPU or GPU. This is of course my speculation, but I think that's what's going on. So what you're doing is you're sort of cheating things.

[00:08:24]So rather than pulling all the memory into your processing unit like your trip unit that's actually a calculation unit that's sitting separately from SRAMM so you have to wait for it then you have to do the calculation then you have to push it back again that takes a lot of time.

[00:08:37]What you do is you can pull it off of SRAMM do some pre-calculations and then send it on to the main calculation step that has to happen. So you're actually actively calculating on things as you're pulling it out of RAM instead of waiting on that. And that as Elon says increases things about 10x an order of magnitude higher memory bandwidth than offchip DRAM. That is a huge unlock especially for what Tesla is doing which is incredibly high bandwidth. Again you're pulling in millions and millions of pixels of data from the external world.

[00:09:07]You're having to act on all of that really really quickly. So the speed with which you can pull and push data through your pipeline is where everything bottlenecks. For the case of navigating the world that's much more important than the raw calculations internally.

[00:09:19]It's how much memory bandwidth you have and how fast it can go. So between lower power double data rate memory in AI6 and also half of the trip AI computation accelerators dedicated to the SRAMM rather than to the raw calculations. It looks like what they're doing is they're pushing for more and more and more ability to calculate taking in all of the information from the real world pushing it through the memory doing the calculations in a very low power extremely low latency environment and then pushing it back out to the control system. In a car, the control system again is relatively easy. Steering wheel, brake, and accelerator. There's only three real control outputs. For something like Optimus, it's much more extensive because, of course, you've got a full body with a bunch of different things that interact with each other. In other words, if you reach your arm out, it changes your center of gravity, so you have to move your torso across. If you pick something up, it really changes your center of gravity. If you squat down, things are different. If you do a handstand, things are different.

[00:10:12]Whatever it is, there's a lot of stuff going on with a humanoid robot. And that actually leads to Elon's next comment.

[00:10:20]And we'll note Elon's replying to himself here above. He said, "Thank you to TSMC and Samsung for your support in bringing this chip to production.

[00:10:27]Interestingly enough, it will be one of the most produced AI chips ever." So, they're planning on scaling this out a lot. And that leads to the next thing, which is Tesla owner Silicon Valley said, "Will this go into the cars or the robots?" And he said, "Optimus and our supercomputer clusters, in other words, dojo. AI4 is enough to achieve much better than human safety for full self-driving. So that's why I said in the opener, don't expect this in your car anytime soon. Will they eventually put it in the car? Maybe. It sounds like they're really convinced that AI4 is enough for full self-driving to be significantly safer than a human driving without having to put in new hardware into the vehicles. So why wouldn't Tesla go ahead and put it in cars anyway?

[00:11:05]Well, number one, that creates another fork in the hardware layers. you've already got hardware 3 and hardware 4 and they're having a hard enough time supporting hardware 3 right now. AI4 is the standard. If they go to AI5, then they're going to have to support AI4 and hardware 3 and AI5 at the same time. So, they probably don't want to do that. And if AI4 is good enough for the cars to get like, let's say, 10 times safer than a human driver on the road without having to have new hardware, then why not stick with AI4? And also, we're looking at Optimus production eventually outscaling cars by an order of magnitude at least. So if they're selling somewhere around 1.5 2 million cars per year, they could be doing 15 or 20 million optimi per year. And that means that they need all of the chips that they can get to put them in their humanoid robots and also to put them in their dojo supercomputer clusters eventually to help with the training and probably also to run some of what Grock does and all of that. Elon has talked about that as well because these chips are much lower powered than Nvidia chips. So, they'll use Nvidia for like the real intelligence stuff, just rewriting an email or something like that. They can use these less expensive chips to get the job done, but they're going to need a lot a lot a lot of this AI5. Again, as he said, it will be one of the most produced AI chips ever. And I think he's not just talking about AI5, but AI 6 and AI 6.5 and AI7 as well. But it sounds like they've already got plans for all of those chips in something besides cars. And if they think that AI4 is by far good enough to do full self-driving because again the output state of full self-driving is really really simple. It's only three output controls and the goal is just not to hit anything. So it's much much simpler than a humanoid robot. And so therefore they can divert all of this chip production into Optimus robots and into their data centers and that's the direction they're heading. So if you were thinking about waiting for AI5 to purchase a Tesla Model Y or something, I wouldn't wait at this point. It sounds like we're going to have AI4 in our cars for quite a while. All righty, folks. That's what I've got for you today. Let me know in the comments what you think about all of this. What do you think the biggest advance is here? Do you think that Tesla's going to be able to produce these chips in these kinds of quantities, or are they going to get stuck in production held? Obviously, the Terapab factors into this as well, and we'll have to talk about that in another video. But for now, I'd love to know your thoughts about all of these advances that Elon revealed. And of course, while you're down there, if you don't mind liking the video, it really helps out with YouTube's algorithm. And if you want to help out the channel a ton, please consider subscribing and hitting that bell notification icon so we can get to 100,000 subscribers in 2026. Thank you so much and I will see you in the next video. Bye-bye.