Why Physical AI Is the Next Frontier | Applied Intuition with a16z

In this talk

1. 1:40

   Physical AI spans many industries

   Applied Intuition builds software and AI for vehicles and other machines, with automotive making up about 30% of its business and the rest coming from other sectors. Gasser sees opportunities in areas such as defense, mining, agriculture, construction, and logistics, where automation could improve safety and productivity. He says the company has more than 1,000 engineers and has deployed models on about 50 platforms.

2. 6:02

   Autonomy must fit the real world

   Long-lived equipment cannot simply be replaced, so existing machines need to become more capable. Coordinating fleets across a port, mine, or quarry could help operators respond to breakdowns and anticipate maintenance. Gasser also points to labor shortages and dangerous working conditions in farming, trucking, and mining as reasons to automate.

3. 13:21

   Deployment, data, and economics shape progress

   Physical AI depends on proprietary data, safety validation, and models that can run within the timing and hardware limits of a machine. Applied Intuition says it has hundreds of petabytes of data and uses synthetic data and simulation to develop systems. Gasser describes commercial deployment as a major challenge, alongside hardware faults such as overheating and sensor problems.

4. 50:00

   Dana aims to open up autonomy development

   Applied Intuition introduced Dana, a platform for designing, training, and deploying autonomous systems. Gasser’s example is a student building a delivery robot: Dana would help define requirements, create scenarios, train the system, and learn from deployment problems. He says workflows that once took days or weeks can in some cases run in minutes.

5. 1:01:00

   Simulation and autonomy face practical constraints

   Gasser describes world models as simulations that represent an environment and respond to an autonomous agent, with approaches ranging from physics-based systems to neural video generation. He says realistic simulation could make training easier, but generated environments still need to match real-world behavior. On deployed machines, models must also meet real-time, safety, and determinism requirements.