Key Takeaways

  • Applied Intuition, a $15B company in physical AI, operates on a stark reality: AI for data centers is fundamentally different from AI in physical machines.
  • "Offboard" AI runs in data centers, handling large models without real-time latency worries, ideal for training and general tasks.
  • "Onboard" AI, powering autonomous vehicles and robots, is severely constrained by milliseconds, power, and cost, demanding highly efficient, distilled models.
  • This means safety-critical elements of physical AI systems are developed 100% in-house, requiring deep hardware-software co-design, rather than relying on generalist foundation models like Gemma.
  • The limiting factor for physical AI isn't model intelligence, but the ability to deploy robustly on specific, constrained embedded hardware.

The Millisecond Trap of Physical AI

When Peter Lewig and Kassar Ounounas built Applied Intuition into a $15 billion powerhouse, they learned a hard lesson about AI: one size does not fit all. Especially not when you're building systems that move in the real world. Forget the vast server farms powering ChatGPT; physical AI lives in the unforgiving realm of embedded systems.

Lewig lays out the core challenge simply: “The great thing about offboard software is you don't have to care about time... But on board you don't have have any of those benefits.” He's talking about the brutal reality of autonomous machines. Your vehicle can't wait for an answer. Every fraction of a millisecond counts. It's a world where models must be so efficient they're almost unrecognizable from their offboard training counterparts. Think of it as a distillation process, where only the absolute essentials survive to run under extreme constraints of power, heat, and immediate responsiveness.

Why Off-the-Shelf Models Fail on the Edge

For many AI applications, big, generalist models like Gemma are impressive. They can handle a range of tasks, from voice assistants to content generation. But try to stick one into an autonomous vehicle's safety system, and you've got a problem. Kassar Ounounas explains that in physical AI, “we're not really constrained right now by like the intelligence of the models. It's actually what Peter's talking about is actually deploying them in… on the hardware you give you.” It's not about making a smarter model; it's about making a smart model run reliably when your life, or the safety of a multi-ton machine, depends on it.

This isn't a problem of research; it's a problem of engineering. The real bottleneck isn't capital or finding enough PhDs, it's the gritty work of making complex algorithms perform under conditions that would crash a data center. Peter Lewig acknowledges that while generalist models have their place for “generic use cases like voice or voice assistant,” anything that requires precise, low-latency, safety-critical performance is a different beast entirely.

Build Your Own Brain (For Safety)

This deep constraint leads to a clear strategic choice for Applied Intuition: for anything safety-critical, they build 100% in-house. You can't just download a neural network and expect it to guide a robot through a chaotic factory floor or navigate a self-driving car in unpredictable weather. It demands deep hardware-software co-design. Every chip, every sensor, every line of code is optimized together, not as separate components cobbled together.

Kassar Ounounas offers a helpful rule of thumb: “any system that you as an as a human would need special training to operate, you can think of a little bit differently.” For these complex, high-stakes scenarios, off-the-shelf general intelligence simply won't cut it. Founders building hard tech must understand that they're not just training an AI; they're crafting a new kind of embedded intelligence, purpose-built for its physical home.

What to Do With This

If you're building physical AI or hard tech, stop thinking of hardware as an afterthought or a commodity. Pull your engineering leads into hardware discussions tomorrow. Demand they explain exactly how your models will perform on target embedded systems, not just in simulation. Your path to $15 billion depends on optimizing for milliseconds, not just model size.