{{ Robotics + AI + Edge/Decentralized Compute + Blockchain + Crypto }}
Physical AI is just another term for autonomous robots sensing, deciding and acting within a specified environment. That eventually looks like autonomous lawn mowers, humanoid home robots that do your laundry, change the oil in your car, and walk your dog. Before that, it looks like AI-enabled robotics for industrial use cases in factories, warehouses, oil rigs, and other places where strength in harsh environments translates to profit. Such robots demand power, safety, low-latency communications, spare parts, skilled maintenance, and the ability to learn by watching. Learning is a major constraint.
What Is Already True
- Robotics in factories is not new. Autonomous robotics is very new. Deployment of autonomous robotics is already happening in China, suggesting that the US and Europe will catch up eventually. The unit economics matter. Prices per robot need to be equal or less than current assembly line robots for factories, and less than or equal to one year salary for replacing human workers such as oil rig workers.
- Robots need training data too and it is the type of data that is not sitting on a server ready to be scraped. Teleoperation to learn could work near term but only in industrial settings.
- Edge compute here looks like compute on the robot itself for minimal latency. That constrains the robot’s abilities but high enough compute exists equal to most tasks.
- Decentralized compute is in the testing phase and could result in the kind of on-site compute that lowers network communication latency for industrial robots.
- Blockchain offers timestamped accountability and tracking of robot decisions and actions while crypto offers agentic payments for robots who may need to make deliveries, or otherwise pay for things in the course of their work.
Assumptions To Be Tested
- AI will significantly accelerate robotics when we get to the level of AI operating systems so that robots can learn by watching and then write an app to do the job. SpaceXai’s concept for AI operating systems “Macrohard” would help that and I suspect it was one of the main points behind the idea (but only Elon could verify that).
- Consumer robotics will likely trail industrial use simply due to distrust and people being creeped out by humanoid robots. Distrust of teleoperated robots is real so expect pushback until home robots are actually autonomous and people are not scared of them.
- When we get to the point of a robot in every home, decentralized applications will allow individuals to rent out their robots to neighbors for specific skills (mechanic, chef, etc).
- Self-reinforced learning for autonomous robots has a long way to go, and must be able to learn from mistakes as well as from watching or experimentation.
- Manufacturing components for such robotics requires the same kinds of raw materials in demand from data centers. Will there be a clash of demand? Or new chip designs?
- Decentralized manufacturing is a real possibility and one worth backing, but is yet un-prototyped and therefore unproven.
- Decentralized manufacturing would start as on-site capabilities for industrial use cases and then translate to your neighbor having a full manufacturing capability in their garage, serving the neighborhood.
How I Would Underwrite It
When looking at a Physical AI play, focus on the use case, the manufacturing plan, and the training plan.
| UNDERWRITE | AVOID |
| Robot training projects, components and robot manufacturing. Industrial FIRST: Use cases where strength in dangerous environments equals profit. Consumer SECOND: Yard robots. DePIN for sensing, networking and compute. | Consumer robots without privacy planning. Teleoperated robots anywhere. Robot tokens. Robotics in high trust and legal-risk scenarios. |