Robotics & Physical AI
Musk Allocates AI5 to Optimus, Leaving Tesla Cars on AI4

Key PointsMusk reserves AI5 for Optimus while keeping Tesla cars on AI4 for driving.
View more
- Musk reserves AI5 for Optimus while keeping Tesla cars on AI4 for driving.
- Reality View's latency stems from tight compute allocation rather than model size alone.
- Musk's proposed fix interleaves newer vision features to cut roughly 35% of latency.
Elon Musk said in an X reply on November 22, 2026, that Tesla’s AI5 computer is intended for Optimus, cars will continue using AI4, and AI4 is sufficient for superhuman driving. He was responding to an owner who asked when AI5 would reach new cars and whether to wait before upgrading from AI3.
The reply followed an earlier technical post in the same discussion, where Musk explained how Tesla’s Reality View visualization fits within AI4. Two linked in-car workloads sit underneath that explanation: a model producing driving actions and a separate generative model producing the interface visualization. Together, the posts give the chip choice a concrete infrastructure context: Tesla is dividing limited compute among connected driving and visualization functions.
Reality View’s In-Car Workload
Musk described Tesla’s approximately pixel-perfect world model as critical for synthetic data generation. Its much smaller in-car version is a conditional mean-flow model based on a recent sequence of Tesla Universal Vision Encoder, or UVE, features. Those same features feed a 3B encoder model that produces driving actions.
Musk said Reality View branches from those features into a separate, roughly 1B-parameter diffusion model. The process starts with a noisy version of the preceding output and reconstructs the current visualization, creating an eight-image video chunk at 36 frames per second and the interface screen’s resolution. This work is performed every UVE frame at 18 hertz and spread across four cycles.
AI4 Constraints and AI5 Headroom
In a follow-up post, Musk said Tesla had moved its small decoder into SRAM, leaving only half of one TRIP out of three on SoC B for Reality View. TRIP basically handles only 8-bit integers, and softmax must run in emulation mode over 28 steps.
Those details apply specifically to the visualizer, not to a stated limit on driving actions. They show that Reality View’s practical budget depends on engine allocation and numerical execution as well as model size.
Musk attributed the current, somewhat high Reality View latency to compute constraints and proposed interleaving newer UVE features through the diffusion transformer layers to cut roughly 35% of it.
Musk described AI5 as a substantial improvement with roughly 12 times more compute. The post gives no baseline for that multiple, so it does not establish a 12-fold gain in visualization frame rate, latency or driving performance.
Musk’s account separates confidence in driving capability from the practical limits of rendering Reality View. AI4 is presented as sufficient for superhuman driving, yet the visualization branch running on it still has a narrow compute allocation and only a proposed route to lower latency. The immediate rendering improvement he described depends on reorganizing vision-feature inputs within that environment, making software efficiency central to the in-car workload he detailed.












