.png)
Here's a question that trips up a lot of people: when Tesla talks about its next self-driving computer, is that the thing that "learns" to drive? Not quite. The chip going into the car, called AI5 (and formerly known as Hardware 5, or HW5), only uses what a neural network has already learned. The learning itself happens in giant GPU clusters in Austin, Texas.
That split is the key to understanding every headline about Tesla's AI hardware. Below, we sort out what Tesla has actually confirmed, what's only been reported, and what the much-quoted "exaflop scale" really means. We even did the math ourselves, since Tesla hasn't published it. Everything here was checked on October 28, 2026.
The Quick Answer
AI5 doesn't train neural networks. It runs them. Tesla described it in its first-quarter 2026 investor update as an inference processor, and said the final design was completed in April.
The heavy training work happens on Cortex. That's Tesla's pair of clusters at Gigafactory Texas, and by Tesla's own listing they're built from more than 230,000 H100-equivalent GPUs.
Plain-English version: Cortex is the classroom, and AI5 is the driver's seat. One teaches, the other applies the lesson, and neither is useful without the other.
Training vs. Inference, Without the Jargon
Training is the slow, power-hungry part. A model studies a mountain of examples, in Tesla's case video from its fleet, and adjusts itself again and again. Inference is the quick part: the finished model looks at camera feeds and decides what to do in a fraction of a second.
What AI5 Is Supposed to Bring
The claims keep shifting depending on who's talking and when. In September 2025, Elon Musk said on a podcast that AI5 would be up to 40 times better than AI4 in some metrics, and eight times better in raw compute. A later summary of Tesla's fourth-quarter 2025 update mentions a 50x target, built on 10x raw compute and 9x memory.
| Feature | AI4 (current) | AI5 (reported or claimed) |
|---|---|---|
| Job | Runs FSD in today's cars | Inference for cars and Optimus |
| Raw compute | Baseline | 8x to 10x, per Tesla-linked claims |
| Memory | 16 GB | Up to 144 GB (reported) |
| Made by | Samsung | Samsung (Texas) and TSMC (Arizona) |
| Availability | In production vehicles | Volume hoped for around April to mid-2027 |
Tesla hasn't published an official AI5 datasheet. Figures reflect Musk's statements and press reports. The 2,000 to 2,500 TOPS number that circulates online is a reported target, and TOPS (low-precision operations) can't be compared directly with data-center training numbers.
Why don't the multipliers match? Because they're counting different things. "Raw compute" is one yardstick. The bigger "system-level" claims also fold in memory size, memory speed and efficiency tricks. Until Tesla puts out a spec sheet, the safe reading is that AI5 is a big jump over AI4, with the exact size still unconfirmed.
Where AI5 Stands Today
The design is done on Tesla's side, and Samsung has reportedly confirmed its own tape-out too. The hard part now is volume. Back in November 2025, Musk wrote that AI5 wouldn't be plentiful enough to switch Tesla's production lines until mid-2027, because the company needs several hundred thousand finished boards sitting beside the line first.
On the October 21 earnings call, Musk said he hopes AI5 reaches volume around April 2027 and that Tesla wants to put it in its Optimus robot at about that time. He called it an order of magnitude more compute than AI4.
Manufacturing is split between Samsung in Taylor, Texas, and TSMC in Arizona. Musk added that the two foundries will turn out slightly different versions, though the software runs identically on both. One practical result so far: the Cybercab robotaxi is launching on AI4, not AI5.
Tesla is also hedging its supply. Its first-quarter update says it's widening its manufacturing scope to include semiconductor fabrication. No finished fab has been shown, so for now that's a plan, not a capability.
Where the Learning Happens: Cortex
Cortex sits at Gigafactory Texas. In Tesla's first-quarter 2026 filing, Cortex 1 is listed at more than 100,000 H100-equivalent GPUs and Cortex 2 at more than 130,000. Tesla said Cortex 2 was online and already running training jobs. ("H100-equivalent" is simply a way of converting different GPU generations into units of Nvidia's H100.)
Later updates switched units. A summary of the Q3 2026 update lists Cortex 1 above 105 megawatts and Cortex 2 above 120 megawatts, with a goal of 400 megawatts by the end of 2026. For perspective, the original Cortex build was about 50,000 H100s and, according to Tesla, helped enable FSD v13.
Doing the Exaflop Math
Tesla hasn't published an exaflop number that we could find, so here's a back-of-the-envelope version. Nvidia's H100 SXM is rated at about 989 teraflops of dense 16-bit (BF16) math, or 1,979 with sparsity, according to spec listings based on Nvidia's datasheet.
Read that as a ceiling, not a speedometer. It's our arithmetic, not a Tesla number, and it assumes "H100-equivalent" maps neatly onto that spec. Real training runs land well below peak because of data movement, networking and hardware failures. So "exaflop scale" is a fair description, but an exact exaflop figure isn't something Tesla has disclosed.
The Dojo Plot Twist
Tesla once planned to train on a chip of its own design, called Dojo. In August 2025, Bloomberg reported that the team had been disbanded, and Musk called Dojo 2 "an evolutionary dead end." He added that AI5 and AI6 would be excellent for inference and "at least pretty good" for training.
Then the story turned again. In January 2026, Musk said Tesla would restart work on Dojo 3 now that AI5's design was in good shape, and he later described AI6 as aimed at Optimus and data centers, with Dojo 3 focused on space-based compute. Tesla's April filing also says custom silicon work with Dojo 3 continues, with the goal of lowering training costs over time. There's no public performance or deployment data for Dojo 3 yet, so today's training still runs on GPUs.
Confirmed, Reported, Unknown
With this much hype around, it helps to sort the claims into piles.
- AI5 final design completed in April 2026
- Cortex 1 and 2 capacity listed in filings
- Cortex 2 online, running training
- 2,000 to 2,500 TOPS
- Up to 144 GB of memory
- Samsung's tape-out is finished
- An official AI5 spec sheet
- A real exaflop figure for Cortex
- Dojo 3 performance
What It Means for U.S. Drivers and Buyers
Tesla says it designs hardware and software together, and the logic is easy to follow. Cortex produces bigger models, and AI5 exists to carry them. A chip with roughly nine times the memory can hold a far larger network inside the car. Whether the real-world gains match the targets will only show once AI5 ships and Tesla releases software that uses it.
The practical takeaways:
- No AI5 car is for sale today. Based on Musk's statements, volume is a 2027 story at the earliest.
- AI4 is still the current platform. Musk has said AI4 is enough for better-than-human safety, a claim that hasn't been independently verified.
- FSD is still "Supervised." The driver stays responsible.
- The spending is real. Tesla's Q3 R&D rose 69% to $2.76 billion, by one analysis, reflecting AI and robotics work.
Quick Questions
We found no announced retrofit program, and Tesla hasn't publicly confirmed one.
Its filings describe Cortex capacity in H100-equivalents, which points to GPU-based training today.
The Bottom Line
AI5 and Cortex are two halves of one plan, but only one half is running at scale right now. The chip is real, designed and waiting on volume. The training muscle is already here, in the neighborhood of hundreds of peak exaflops by our estimate.
Watch the next couple of earnings updates for proof that AI5 volume is on schedule, and for any hard numbers on Dojo 3. Those two things will tell you far more than another round of "40x" headlines.
Sources
- Tesla Q1 2026 update (SEC filing): AI5 design, Cortex capacity, Dojo 3
- Tesla Q2 2026 update (SEC filing): Cortex in megawatts
- EVwire: Q3 2026 earnings call transcript
- FourWeekMBA: Tesla's 400 MW training compute plan (secondary summary)
- Electrek: AI5 delayed to mid-2027
- TrendForce: Samsung and TSMC both build AI5
- ArenaEV: Samsung tape-out reported
- Not a Tesla App: Musk's AI5 spec claims
- SocialPruf: summary of Tesla's Q4 2025 update (secondary)
- TechCrunch: Tesla's Dojo, a timeline
- TechNews: Dojo 3 restart
- Teslarati: AI5 tape-out and Musk's AI4 claim
- sec-api.io: Q3 2026 financial analysis
- GPU Smith: Nvidia H100 SXM specifications
- Images: free to use under the Unsplash License (photographers credited in the captions)
.png){kind=logo}
No comments:
Post a Comment