VC
Value Add VC
⚡HomePulse⚡Helpful Apps📝Blog🤝Partner
Illustration for: Physical AI's Real Bottleneck Is the Control Stack
Value Add VC/Pulse/AI

Physical AI's Real Bottleneck Is the Control Stack

Google DeepMind's Gemini Robotics 2 extends full-body, video-native control to an entire humanoid robot rather than just its upper body -- a sign the real competitive layer in robotics may be the shared foundation model, not any single hardware maker.

TC
Trace Cohen
Early-stage VC & angel · Founder, New York Venture Partners
August 3, 2026
2 min read
ShareXLinkedInEmail

THE RUNDOWN

1

Gemini Robotics 2 extends DeepMind's robotics model to full-body control -- walking, crouching, and manipulating objects -- where its predecessor only handled upper-body manipulation tasks

2

The model is video-native, learning to control a robot's movements from video input rather than requiring specialized robot-specific sensor data, closer to how one general-purpose architecture collapsed dozens of narrow NLP tasks in the LLM wave

3

It puts Google DeepMind in more direct competition with Tesla's Optimus program, Figure, and China's Unitree -- all racing to prove a single foundation model can generalize across an entire robot's body rather than stitching together separate locomotion and manipulation systems

4

If a small number of foundation models end up powering most humanoid robots the way a small number of LLMs now power most AI applications, hardware makers may end up with less differentiated economics than the model layer sitting on top of them

TC

The VC Read · Trace's Take

Trace Cohen

The margin question every phone-and-PC investor already knows -- does the assembler or the platform layer capture the value -- is now the humanoid-robotics question, and Gemini Robotics 2 is DeepMind's opening bid to be the platform layer. If you're underwriting a humanoid-hardware company's valuation without asking whether it owns its own full-body control stack or is licensing someone else's, you're pricing the wrong asset.

AI Valuations Tracker →

Analysis

Google DeepMind's Gemini Robotics 2 extends the company's robotics AI model to full-body control of a humanoid robot -- walking, crouching, and manipulating objects together -- where its predecessor handled only upper-body manipulation tasks. The model is video-native, meaning it learns to control a robot's movements from video input rather than requiring specialized, robot-specific sensor data.

That's a meaningful technical jump, not an incremental one. Most humanoid-robot AI to date has stitched together separate systems for locomotion and manipulation; a single model handling both end-to-end is closer to how the large language model wave collapsed dozens of narrow NLP tasks into one general-purpose architecture. It puts Google DeepMind in more direct competition with Tesla's Optimus program, Figure, and China's Unitree, all racing to prove a single foundation model can generalize across an entire robot's body instead of licensing or building separate subsystems.

“For robotics investors, the stakes are about which layer actually captures value over the next several years.”

For robotics investors, the stakes are about which layer actually captures value over the next several years. If a small number of foundation models end up powering most humanoid robots the way a small number of LLMs now power most AI applications, the hardware companies building on top of those models could end up with less differentiated economics than the model layer itself -- the same margin question that's already playing out in phones and PCs, where the OS and chip vendor often capture more value than the device assembler.

That question sits directly underneath the humanoid-robotics valuation debate playing out elsewhere on this page: whether Unitree's public listing or Figure's private mark holds up depends partly on whether either company controls its own full-body control stack or is ultimately a hardware integrator running someone else's model. A foundation-model provider with genuine full-body generalization has real leverage over every hardware maker that licenses it.

What to watch: whether Gemini Robotics 2 gets licensed to third-party humanoid-robot makers beyond DeepMind's own hardware partners, and how it performs head-to-head against Tesla's and Figure's in-house foundation models on real-world, unscripted tasks rather than choreographed demos.

ShareXLinkedInEmail
More onGoogle →

Analysis and editorial commentary by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.

Read Next

AI· Aug 1, 2026

OpenAI's Astra Cracks 10 Unsolved Math Problems

Illustration for: OpenAI's Astra Cracks 10 Unsolved Math Problems
AI

OpenAI's Astra Cracks 10 Unsolved Math Problems

An internal version of OpenAI's next model, Astra, solved ten previously-open math and computer-science problems for roughly $2,000 in compute -- proofs Fields Medalist Timothy Gowers says he'd back for a top journal.

AI· Aug 3, 2026

AI's Dual-Use Risk Just Became a Live Security Problem

Illustration for: AI's Dual-Use Risk Just Became a Live Security Problem
AI

AI's Dual-Use Risk Just Became a Live Security Problem

A scoping error that let Claude models breach real production systems, and a US-China robot-ban standoff threatening rare-earth retaliation, broke days apart -- proof AI's dual-use risk is now operational, not hypothetical.

AI· Aug 3, 2026

DeepSeek's Offensive Use Forces a Red-Team Rethink

Illustration for: DeepSeek's Offensive Use Forces a Red-Team Rethink
AI

DeepSeek's Offensive Use Forces a Red-Team Rethink

Palo Alto Networks caught an operator using DeepSeek to autonomously attack 460+ systems after Claude and OpenAI's models refused the same job -- a live test of whether model-level refusal is a real safety layer or just a routing problem.

@Trace_Cohen·t@nyvp.com