# Figure runs its robot intelligence on board to meet latency needs

Markdown copy of https://realalpha.app/s/figure-runs-its-robot-intelligence-on-board-to-meet-latency/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Brett Adcock](https://realalpha.app/builders/brett-adcock/) · Brett Adcock is founder and CEO of Figure AI. · Founded Figure AI to build general-purpose humanoid robots; previously co-founded Archer Aviation and Vettery.
- Host: [Peter Diamandis](https://realalpha.app/people/peter-diamandis/) · Founder, XPRIZE · Co-founded Singularity University and founded Zero-G; co-wrote the books Abundance and Bold.
- Host: [Dave Blundin](https://realalpha.app/people/dave-blundin/) · Founder & GP, Link Ventures · Has co-founded 23 companies; founding CEO of Vestmark and chairman of insurance marketplace EverQuote.
- Show: Moonshots
- Episode: [Robotic Deep-Dive: Beating China, the Race Against Tesla, & New Robotics Timeline With Brett Adcock](https://realalpha.app/v/E864U582FVY/), Nov 16, 2026
- Moment: 24:06 to 29:19 of the episode (5 min), playing on https://realalpha.app/s/figure-runs-its-robot-intelligence-on-board-to-meet-latency/; the episode from there: https://realalpha.app/v/E864U582FVY/?t=1446

## Quote

> “So, everything from the planning and the reasoning to even the low-level sending trajectories out and torques, all that's **done on the robot today**.”
> — [Brett Adcock](https://realalpha.app/builders/brett-adcock/) at [25:48](https://realalpha.app/v/E864U582FVY/?t=1548) · Moonshots

## Takeaways

1. Helix inference runs on the robot, not in the cloud.
2. Fast joint and sensor control needs millisecond-level response times.
3. Cloud services coordinate fleets and handle real-time speech.

## About this segment

Brett Adcock says Helix runs on the robot because fast movement requires millisecond-level responses. Cloud systems handle speech and coordinating fleets, but network delays make them unsuitable for real-time control.

From “Robotic Deep-Dive: Beating China, the Race Against Tesla, & New Robotics Timeline With Brett Adcock”, Nov 16, 2026. [Watch the full episode](https://realalpha.app/v/E864U582FVY/)

## Transcript

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

[24:10](https://realalpha.app/v/E864U582FVY/?t=1450) be very shocked if we exit 2027 and we haven't seen this whole picture. How much of the compute is running on the figure robot versus on Wi-Fi? Also, training versus inference. How much is training? How much is inference? Yeah, most of our compute today is in training. The dollar is going towards compute. We're doubling our compute right now internally next week. And then we did a big deal with N-Scale for 2027.

[24:39](https://realalpha.app/v/E864U582FVY/?t=1479) That deal, 100,000 GPUs, three plus billion dollars. When's that come online? Yeah, those are like second half of 2027. Those are VR2200s, the best computer in the world. Yeah. Wait, what's the ratio of training compute you'll have on that day versus, say, a month ago? Oh, man. Let me do the calculator. There are different type of GPUs. I figured the VRs are 2x, like, raw. Are you using V300s today or something like that?

[25:06](https://realalpha.app/v/E864U582FVY/?t=1506) Yes. So we, let's see, we have, they're roughly on order of 4. 4x? Yeah. So, man. Yeah, like, if we took the whole deal with NScale and size it all the way up, well over 100x, what we have now. Over 100x. Yeah. So going back to the question of... Yeah. I mean, we also spend like nine figures a year now on compute before this. Like we're like, you know, spending like a large amount of dollars on compute here.

[25:34](https://realalpha.app/v/E864U582FVY/?t=1534) That's pretty intimidating to everybody else. Inference on the robot, on Wi-Fi? We do all inference on the robots you're seeing here on board today. So you've got what on board in terms of compute? Yeah, we have all of Helix is fully on board the robot. So, everything from the planning and the reasoning to even the low-level sending trajectories out and torques, all that's done on the robot today.

[26:02](https://realalpha.app/v/E864U582FVY/?t=1562) And what gets done on the server? The only thing we do today is we do real-time speech-to-speech through servers. And we use the servers to help think about... We think of the robots like agents, like physical agents. If you think about if I had 100 robots building a home, or building in construction, how do we get the robots to... how do we command that, hopefully, minute-to-minute, hour-to-hour?

[26:30](https://realalpha.app/v/E864U582FVY/?t=1590) Because as a home is getting built, the job's changing. The friend of your org. Yeah, exactly. So, that's all done from our agent orchestration in the cloud today, which is the only way to do it. In this case, we have multiple... even on this conveyor here. Robots are subbing out. They're recharging. If a robot gets damaged or doesn't feel healthy, it walks to maintenance. And we have another robot comes in and flags the fleet. All this is being networked together. That has to be done in the cloud. That will always be done in the cloud. We're experimenting with onboard and off-board computers. It's really hard.

[27:06](https://realalpha.app/v/E864U582FVY/?t=1626) I think the struggle we have is, we can't get around the physics of having very high dynamics on the robot. We need to be able to talk to all the joints and sensors, and then communicate back what the new plans are in a really fast way. We don't have seconds we can wait, or even half a second. We need to get down and... in the millisecond level. There's no getting around that, as we do really fast dynamics of the world. In the best-case scenario, you have a big brain living in cloud, you have a small brain living on board. In an extreme example, you have everything living on board, which we're doing now. Jury's still out. We're still doing tons of experimentation.

[27:52](https://realalpha.app/v/E864U582FVY/?t=1672) But all the robots right now are all networked on board. I think it's nice, because we can lose internet and the robots still do work, which is nice. If you're even at home, you you have bad 5G service or something. You don't want your robot stopping mid-laundry. You want it to continue to do work. So there's a lot of different engineering trade-offs we have to go through to really understand this better.

[28:09](https://realalpha.app/v/E864U582FVY/?t=1689) So, milliseconds, is that even viable on the cloud? No. The home Wi-Fi is a bottleneck, but then getting milliseconds is never going to work. Really, the best 5G networks we see that you can do is you can get down to maybe 20 to 30 milliseconds per leg. So you're looking at almost 50 milliseconds just for a full round trip, then you're at 20 hertz just in the network itself. And then you have all the forward paths you need to do on inference. And there's a bunch of other stuff we need to do as it relates to the images and torques coming back. So it's not just as easy as pinging a cell tower and kidney back. We have a lot of other work to go do.

[28:47](https://realalpha.app/v/E864U582FVY/?t=1727) Do you think then every home is going to have some kind of a brain box in it in the basement where the cable box is? Our hope is that everything's done on board and we have the right architecture now. But if every home has like 40 of them, do you really want to... That's a good market projection, 40 per ohm. No, I mean, everything's done on board. Like, we do have it on board right now. Like, the robots still need to know.

[29:05](https://realalpha.app/v/E864U582FVY/?t=1745) Chips are scarce, though, if you can share it, you know, because you might have, like, a roof, you know, gutter cleaning bot up there. I don't think we ever, like, listen, there'll be some neural nets that no matter what, we'll have to live on the robot. Like, they're just running too fast.
