After the Agents: Building for the Physical World

Oct 15, 2026

Almost all modern models are built on top of the neural network architecture. The three pillars of neural network architecture are “data, algorithm and compute”.

Research shows these models achieve scaling laws which means, model performance tends to improve predictably with more data, more compute, and larger models (more parameters), provided the data is useful and the training process (algorithm) can effectively use these resources.

Models built by frontier labs (eg; astra, opus, etc.) have demonstrated extraordinary coding and agentic execution abilities, they possess potential to automate a significant portion of the knowledge work.

We are already seeing early signs that the next wave of AI applications will extend beyond knowledge work, with frontier labs increasingly focusing on scientific research, personal agents, physical AI, physical product design, engineering, manufacturing, etc.

There is a useful parallel between how LLMs learned and how models for physical AI and engineering might learn.

As frontier models became increasingly capable of coding and agentic execution, we saw a surge in companies building across the digital AI agent stack (developer tools, AI infrastructure, training data and evaluation platforms, horizontal agents, vertical AI companies, and enterprise agent platforms).

As models become increasingly capable of understanding the physical world, designing products, and performing physical tasks, we could see a similar surge in companies building across the physical AI stack.

This is how I am looking at it, just as the digital agent wave created companies around:

data → training → evaluation → tooling → applications

physical AI could create an even broader ecosystem around:

data → models → simulation → design → manufacturing → testing → deployment

The important distinction is that physical AI requires both software infrastructure and physical infrastructure.

That’s why I expect entirely new companies to emerge around hardware design and production.

Example of companies emerging across the physical AI stack could be, AI CAD, data collection, autonomous prototyping, engineering agents, design validation, AI-native manufacturing platforms and so many more.

I hesitate to draw direct analogies, since the underlying technologies and opportunities are quite different.

But if we look back in time, the early wave of internet companies was largely focused on software and digital services. As the internet matured, we saw the rise of companies like Uber, Airbnb, and DoorDash that used software to coordinate and deliver services in the physical world.

We might see a similar shift with AI, from automating digital work to enabling entirely new businesses in the physical world.

One distinction will matter more than ever in physical AI. As frontier models become increasingly capable, software-only model wrappers will become easier to replicate. Companies that combine proprietary data, physical infrastructure, and strong customer relationships will have more opportunities to build lasting advantages. Owning the loop from AI-generated design → manufacturing → testing → real-world feedback seems like an interesting opportunity.

Another important distinction between digital and physical AI is that, in the digital world, we can vibe code, experiment, and deploy products entirely through software.

In physical AI, we might be able to design products, run simulations, and test ideas digitally, but eventually, we still need to turn those designs into working physical products.

I’m particularly excited about this last step: bridging the gap between digital designs and physical reality.

This is me thinking out loud and putting my thoughts into words. Would love to hear where you think things are heading and learn from your perspectives :))