Carbon Must Outperform Gemini on Contact Tasks
Sanctuary AI
The real risk is that general robot brains become cheap and swappable, which pushes Sanctuary to win on the part of the stack that actually touches the world. Google DeepMind is already showing Gemini Robotics across multiple robot bodies and hands, with Apptronik as a partner and other OEMs as testers. That means Carbon cannot just be a smarter controller in demos. It has to prove faster picks, fewer failed grasps, and better handling when the robot must feel contact, adjust grip, and finish the task cleanly.
-
Sanctuary is selling both Phoenix humanoids and Carbon control software, but its biggest expansion path is putting Physical AI onto existing industrial robots. That makes software differentiation central. If OEMs can buy a foundation model elsewhere, Carbon needs measurable edge in real shop floor workflows, not just broad reasoning.
-
Gemini is moving down market from research into distribution. Google launched Gemini Robotics in March 2025, added an on device version in June 2025, and by July 2026 was showing one checkpoint controlling different embodiments and end effectors. That is the shape of a licensable model layer that can compress standalone robot software margins.
-
Meanwhile hardware pricing is being reset from below. Unitree lists the G1 from $13,500, while Apptronik sits at a $5.5B valuation, Physical Intelligence at $5.6B, and Skild AI at $14B. As capital floods into hardware agnostic model companies and low cost robot makers, the premium shifts to whoever can deliver the best contact rich throughput per hour.
The next phase favors companies that can turn embodied AI into hard operating metrics. Sanctuary is best positioned if Carbon becomes the layer customers trust for tight insertions, irregular picks, bag handling, and other messy jobs where touch matters more than general chatty intelligence. In that market, the winner is the system that finishes more tasks per shift at lower supervision cost.