Applied Intuition: A Billion Intelligent Machines
What this covers
This Business Breakdowns episode has Applied Intuition co-founders Qasar Younis and Peter Ludwig walking through how the company got from developer tools to licensing intelligence into other companies' machines. They frame it as a technology provider, closer to a chipmaker than to Waymo, which they say sells robotaxi rides rather than intelligence to manufacturers. The core of the discussion is why physical AI differs from chatbots: hard real-time decisions, affordable silicon, and safety-critical systems that share space with people. Peter argues the model is only about 1 percent of the full solution, and that the real obstacle is getting the whole stack to work, not chips or hardware.
The range runs from technical to commercial. On the technical side, Peter contends that end-to-end imitation learning alone won't produce productionizable autonomy, and that pairing it with reinforcement learning inside high-performance simulation is the way forward. Qasar's claim that specialized models are becoming general, as chatbots did after transformers, is flagged as contested. The pair also describe a data engine that collects proprietary real-world data, and say that data from drones and trucks in Japan improves models across domains.
The business discussion covers why they started with tools and then moved into the operating system. They also cover why they skipped robotaxis in the early 2010s, and how a horizontal company stays exposed to autonomy breakthroughs arriving roughly every two years. Mining is the most concrete use case: about 1% of the workforce but about 8% of work-related fatalities. Competition gets a distinctive framing: in large, fast-growing markets it is not zero-sum, and new hardware startups are customers rather than threats. They also say they have raised about $1 billion and spent none of it, and they treat capital as a variable to fix when it constrains the mission.
Applied Intuition argues physical AI, intelligence embedded in machines that move in the real world, will be far larger than digital AI, and that a horizontal technology provider selling tools and models across many industries, rather than building one vertical product, is the defensible way to capture it.
- Physical AI faces safety-criticality, hard real-time, and cost-envelope constraints that chatbots do not, so general-purpose LLMs are insufficient
- Starting with tools and building horizontally across automotive, defense, mining, and agriculture let the company reuse work and avoid being too early or too late
- Proprietary real-world data, a deep data-collection stack, and sticky embedded deployments form the claimed moat
A general-purpose model is roughly 1% of physical AI because deployment needs about 1,000 engineering problems solved.
- A general-purpose model is only a small fraction of a physical AI solution, roughly 1%, because deployment requires about 1,000 distinct engineering problems, including data pipelines, operating systems, simulation, and safety validation.
“Sometimes it gets trivialized and people think about, oh, it's just about the model, but really there's about 1,000 different problems you have to solve to make this all work.”