2026 09 15 A Brain TOO BIG TO Carry ON Device

Summary

SemiAnalysis (September 14, paid; free portion filed) restates the on-device vs datacenter split for generalist robots. Action/servo loops at ~100 Hz cannot leave the robot; planning below ~20 Hz can. Frontier robot models are still billions of parameters (π0 ~3B, DreamZero 14B needing two GB200s off-robot). Nvidia's Jetson is a rounding error vs datacenter wafers; silicon-efficiency crossover is argued at ~7 robots per shared GPU. Follow-on to the September 9/10 robot-cognition note already ingested — attach only, do not mint a robotics chain into the overweight AI-infrastructure cluster.

Article

Where should the brain of the robot go?

So far, AI has mostly lived behind a screen. Chatbots answered questions. Then agents started driving software and finishing multi-step tasks on their own. The next step is AI that acts in the physical world, and the biggest piece of that is robots. It’s early. Nobody has settled the hardware, the models, or the economics.

With LLMs, the hardware bends to the model. Robotics inverts this due to two constraints. The first is time. A robot runs real-time control loops and can’t miss a deadline. The second is cost. The manufacturer has to build both the compute and the robot itself and pay for it upfront, on every unit.

This is why frontier robot models remain far smaller than frontier LLMs. Today’s generalist robot models are in the billions of parameters: Physical Intelligence’s π0 class at ~3 billion, π0.7 at 5 billion, ByteDance’s GR-3 at 4 billion, Generalist’s models around 10 billion, and NVIDIA’s DreamZero at 14 billion.

NVIDIA’s DreamZero, a 14-billion-parameter “world action model” built on a video-diffusion backbone, needs two GB200 GPUs off-robot just to run in real time. Months after DreamZero, NVIDIA’s RoboTTT made the opposite bet, using a 3-billion-parameter policy that continually updates its own weights with test time training. The result is a model with minutes of usable context that is small enough to run onboard.

Frequency determines which layers need to run on the robot. The action and servo layers run at hundreds of Hz: a 100 Hz loop must produce its next output every 10 milliseconds. A typical wireless round trip costs 10 to 50 milliseconds on its own, so for these loops the network consumes almost the entire budget before the model has computed anything. This is why anything above roughly 100 Hz can never leave the robot.

Below roughly 20 Hz, where the planning layer lives, the picture changes. A planner running at 5 Hz gets 200 milliseconds per decision, so even a 10 to 50 millisecond round trip fits comfortably. Jitter, not latency, is the harder problem.

Jetson Thor is the best robot compute you can buy today, and it still delivers only about 1/10th the FLOPs of a GB200 and roughly 1/30th of its memory bandwidth.

Nvidia’s accelerator output is almost entirely datacenter silicon. Jetson, the line used in robots, is the thin red sliver along the bottom, barely visible even in 2026. A Blackwell datacenter GPU earns far higher gross margins than a Jetson module, on the order of mid-to-high 70s percent versus mid-60s.

Thor has closed onto TSMC N4, and the generation after Thor likely moves to N3 alongside Rubin, then N2 alongside Feynman — exactly the nodes the entire AI-accelerator roadmap is fighting over.

A million Jetson-class chips in 2030, each around 400mm², is only about ten thousand wafers for the year, a rounding error next to the datacenter GPU ramp. If you look at silicon efficiency as wafers per robot, the onboard chip and the shared datacenter GPU cross over at around 7 robots per GPU. Beyond that, serving cognition from a shared GPU takes less silicon per robot than putting a chip in every machine.

(Paywalled after this point. partial: true. Follow-on to https://newsletter.semianalysis.com/p/where-does-a-robot-think-on-device already ingested 2026-09-09/10.)