TL;DR: In Part 1 we mapped the machines. Part 2 maps the brains: 21 startups building the foundation models, training infrastructure and applied labs and farms behind physical AI. Together the two parts cover Bessemer's full 50-company map.

A robot is only as smart as the model that runs it and the tools that train it. If Part 1 was the bodies, humanoids, drones and self-driving vehicles, this is the intelligence layer: the models, data and simulation stacks that turn motors into autonomy. Using Bessemer's atlas again, we grouped 21 more private leaders into four categories. Every name links to the company.

Foundation Models

The models giving robots a mind of their own.

World Labs: Fei-Fei Li's spatial-intelligence startup building world models that generate and reason over interactive 3D worlds.

Physical Intelligence: vision-language-action robot foundation models (the pi series) built to control any robot on any task.

Skild AI: the Skild Brain, an omni-bodied general-purpose robot foundation model out of Pittsburgh.

Generalist: hardware-agnostic robot foundation models (GEN-0/GEN-1) for general-purpose manipulation.

Dyna Robotics: DYNA-1, a commercial-ready robot foundation model for high-dexterity task automation.

Perceptron AI: foundation models for real-time multimodal perception and embodied reasoning.

Eka Robotics: vision-force-action foundation models using tactile sensing and sim-to-real for dexterity.

Field AI: Field Foundation Models, a general robot brain for GPS-free autonomy in unstructured environments.

Robotics Infrastructure

The plumbing that trains, tests and ships the machines.

Foxglove: a data, visualization and observability platform for robotics and autonomous systems.

Point One Navigation: centimeter-level GNSS positioning for autonomous systems.

Voxel51: FiftyOne, a data-curation platform for computer-vision and physical-AI datasets.

Zeromatter: a simulation platform for robotics, autonomy and aerospace.

Theseus: a GPS-denied visual navigation system that lets drones fly without GPS.

Health & Life Sciences

Robots and AI moving into the lab and the operating room.

Periodic Labs: AI scientists plus autonomous labs to automate materials discovery, from ex-OpenAI and DeepMind researchers.

Medra: AI robots that automate repetitive wet-lab bench work.

Radical AI: a self-driving autonomous lab pairing AI with robotic experimentation for new materials.

Mendaera: Focalist, a handheld robotic system pairing robotics with real-time ultrasound for needle-guided procedures.

Climate & Agriculture

Autonomy pointed at fields, orchards and wildfires.

Carbon Robotics: LaserWeeder AI robots and tractor autonomy for herbicide-free farming.

Orchard: FruitScope, an AI vision system for precision crop and fruit management.

Seneca: autonomous AI drones that detect and suppress wildfires.

Upside Robotics: solar-powered autonomous field robots for precise crop nutrient application.

Why it matters for crypto

If Part 1's machines are the demand side of a machine economy, Part 2 is its supply chain, and much of it maps cleanly onto crypto's open-infrastructure thesis. Training these robot brains needs enormous compute and data, exactly what decentralized compute and DePIN data networks aim to coordinate and pay for permissionlessly. The data that teaches robots to see and act becomes an asset worth owning, pricing and verifying, and onchain provenance is one way to prove where a model's training data and outputs came from. As autonomy spreads from labs to farms to city streets, the pipes that move value and trust between all these agents are the same rails crypto is already building.

This completes our two-part Physical AI map. See Part 1 for the machines themselves.

Source: Bessemer Venture Partners, "50 Startups Transforming Industries With Physical AI." Categorization adapted by CryptoDiffer. Not financial advice.