# AI5 targets a 40x gain on an AI4 bottleneck

Markdown copy of https://realalpha.app/s/ai5-targets-a-40x-gain-on-an-ai4-bottleneck/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Elon Musk](https://realalpha.app/builders/elon-musk/) · Elon Musk is a Tesla CEO.
- Host: [Chamath Palihapitiya](https://realalpha.app/people/chamath-palihapitiya/) · Founder & CEO, Social Capital · Led Facebook's growth team; his firm backed Slack and AI chipmaker Groq.
- Host: [Jason Calacanis](https://realalpha.app/people/jason-calacanis/) · Founder & CEO, LAUNCH · Early Uber angel; sold Weblogs Inc. to AOL; hosts This Week in Startups.
- Host: [David Sacks](https://realalpha.app/people/david-sacks/) · Co-founder & partner, Craft Ventures · PayPal's first COO, founded Yammer; White House AI & crypto czar 2025–26.
- Host: [David Friedberg](https://realalpha.app/people/david-friedberg/) · CEO, Ohalo · Founded The Climate Corporation, sold to Monsanto for about $1.1B in 2013.
- Show: All-In
- Episode: [Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding](https://realalpha.app/v/qeZqZBRA-6Q/), Sep 10, 2025
- Moment: 12:16 to 16:53 of the episode (5 min), playing on https://realalpha.app/s/ai5-targets-a-40x-gain-on-an-ai4-bottleneck/; the episode from there: https://realalpha.app/v/qeZqZBRA-6Q/?t=736

## Quote

> “By some metrics, the improvement in AI5 will be **40 times better than AI4**.”
> — [Elon Musk](https://realalpha.app/builders/elon-musk/) at [13:23](https://realalpha.app/v/qeZqZBRA-6Q/?t=803) · All-In

## Takeaways

1. Tesla's AI software and silicon teams co-design the chip.
2. AI5 has 8x raw compute; 40x refers to a specific bottleneck.
3. Musk expects current cars to reach 2–3x human driving safety.

## About this segment

Elon Musk says AI5 has 8x raw compute, while its 40x figure reflects fixing a specific AI4 bottleneck. He expects current cars may reach two to three times human driving safety.

From “Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding”, Sep 10, 2025. [Watch the full episode](https://realalpha.app/v/qeZqZBRA-6Q/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[12:20](https://realalpha.app/v/qeZqZBRA-6Q/?t=740) So there's the LLMs, there's the actuation, the hands, but also there's the silicon that runs it. And there was, you know, Dojo, I think you posted on X, AI5 and AI6, and it just seemed like you were incredibly excited about the direction in which the silicon layer was also going. Can you tell us about that and what that is? And what are we building here? What is being built? Is it a complement to everything that exists in the world? Is it a potential long-term competitor what is it um yeah so at tesla we basically had two different chip programs one dojo and one uh dojo on the training side and then what we call you know ai4 it's just our inference chip um uh that the ai4 is currently shipping in all vehicles and we're finalizing the design of AI5, which will be an immense jump from AI4. By some

[13:29](https://realalpha.app/v/qeZqZBRA-6Q/?t=809) metrics, the improvement in AI5 will be 40 times better than AI4. Wow. So 40%, 40 times. And this is because we work so closely at a very fine-grained level on the AI software and the AI hardware. So we know exactly where the limiting factors are. And so effectively, the AI hardware and software teams are co-designing the chip. So a 40x improvement in the silicon, I think then as everybody here in the audience experiences it, is that just almost like an order of magnitude increase in the quality of FSD and the safety that you experience as a Tesla driver and then the quality of the robot?

[14:15](https://realalpha.app/v/qeZqZBRA-6Q/?t=855) Where does it all manifest when you bring it up and actually get it into production? Yeah, to be precise, the 40X is on, if you say, compared to the worst limitation on AI4, which is running the Softmax operation. We currently have to run Softmax in around 40 steps in emulation mode, whereas that'll just be done in a few steps natively in AI5. um ai5 chip will also be uh easily handle mixed precision uh models so you don't have it'll dynamically handle mixed precision there's a bunch of sort of technical stuff that ai5 will do a lot better um in terms of of nominal sort of uh raw compute it's it's eight times more compute um about nine times more memory uh roughly five times more memory bandwidth So, but because we're addressing some core limitations in AI4, you multiply that by that 8x compute improvement by another 5x improvement because of optimization at a very fine-grained silicon level of things that are currently suboptimal in AI4.

[15:31](https://realalpha.app/v/qeZqZBRA-6Q/?t=931) That's where you get the 40x improvement. You had, oh, keep going. So now that said, I am confident that the current chips, AI4 chips that are in the cars, will achieve self-driving safety that is at least two to three times that of human and maybe even 10x. And the software that will be released for that is coming out over the next few months. months so version 14 will be the biggest uh upgrade in tesla software since version 12.

[16:08](https://realalpha.app/v/qeZqZBRA-6Q/?t=968) we are increasing the parameter count by an order of magnitude um the there's this there's a lot of uh reinforcement learning that's been used as um you can think of AI as a way of compressing reality and some of those compression steps were too lossy and we addressed the lossiness in the compression steps so these are all software updates that'll go out so just over there are updates your car is going to feel like it is sentient by the end of the year it feels that way already