Everyone wants to know what the next big theme is going to be.
I was fortunate enough to be positioned in the optics super cycle before most and it has been life changing.
What if we can find the next theme of this magnitude and position ourselves early?
I believe we can.
And that is what this report is about.
After really digging into this one, I believe this is it.
So in this report I am going to lay out the theme, why it’s investible now, where it’s investible, all the layers + supply chain, and what my plan is going forward.
Optics / Photonics is still my main theme that I will be covering in depth on this page.
But there is opportunity out there just waiting for us.
And this theme I am going to lay out here is going to be the 2nd main coverage universe on my Substack over time.
We are still early and we have time to position ourselves.
Let’s figure out the best way to do that.
Physical AI
So much of my portfolio has been in AI Infrastructure over the past year. Which has obviously been great.
Now I increasingly want fresh dollars that I have to be spent on new opportunities in the Physical AI theme.
All the Infra companies that I invest in enable this new theme.
My goal here with this report is to introduce this theme at a very digestible level so we can all start off on a very grounded foundation.
When I think about Physical AI, I think about it being the point where machines become intelligent.
So we are talking about sight, movement, recognition, force, safety, feedback. Just to name a few of the features/characteristics of some end products.
The flow looks like seeing the world (through cameras and/or sensors), understanding this input, deciding what actions makes sense, creating movement or some kind of response, and then checking what happened and adjusting.
The feedback loop is a critical piece here. The machine has to learn from the result and this is a big differentiator from ‘non intelligent’ machines that just repeat the same pre-programmed motion.
As you are reading this post, I want you to take notes as you go. Whenever you read about an area that you think might be investible, stop and write it down. Try to create a branch of how many different ways you think this layer might be investible. Not companies, just different ways to monetize either with a product or a service. Once you finish that, continue to read. Rinse and repeat.
The Shift
From the screen to the real world.
A major of the use case for AI right now falls into Generative AI with text, image, code, video, research, analysis etc. Practically everyone reading this has used it, if not uses it daily. ChatGPT, Claude, Gemini etc.
The next wave that we are already seeing is Agentic where AI starts running tasks instead of solely giving responses. So an agent can search files, update spreadsheets, draft emails, work on your schedule etc. So from ‘answering’ to ‘doing’, but solely in the digital realm.
Then there is Physical AI where you have control of something int he real world. Rather than telling a machine operator what went wrong, it can adjust the machine.
I think that the visual example most people turn to when they hear about Physical AI is a humanoid. We just saw Figure put on a live side by side comparison of a humanoid and a human sorting through packages.
But in my opinion, that is one of the least practical forms of Physical AI today. What’s real are drones, robotic arms, self-driving tractors, warehouse systems, surgical system, defense systems and so on.
The deeper idea is that AI is moving closer to labor.
Tying it back, Gen AI helps with knowledge, Agentic helps with office/software, and Physical helps with…well…Physical work. Glad they kept it simple for us.
So, what’s the TAM here? How much money is really in Physical AI?
The answer is: ‘Yes’.
All of it. Everything.
Ok but seriously, the physical economy is massive.
We are talking farms, warehouses, mines, hospitals, defense systems, transportation networks, energy infra, construction. You name it.
Why This Is Hard
Let’s ground this report in realism.
When we are isolating AI inside a screen it is much easier to manage. Mistakes are not as critical. Edits are easy to make. The environment is relatively controlled.
Step into the real world and it gets messy.
Object placement is unpredictable. Humans are chaotic (no offense). You have environmental factors like rain and wind. Light inside a warehouse can vary.
So these systems have to UNDERSTAND uncertainty. Again, this is the ‘intelligence’ part. These systems need to work even when objects move unpredictably, sensors get dirty, lighting changes, people walk in their path etc.
Then there is the safety aspect. Yes there are safety concerns with software, especially if we let agents roam free in an enterprise, but we are talking about physical safety here. Inventory can get damaged, humans can get injured, equipment could break.
So these systems have to be accurate, reliable, fast, and safe all at the same time.
I think accuracy, reliability and safety are all common sense.
Speed could sound surprising to some people. But think about it. The clearest example is self driving. If a human steps in front of a car unpredictably, ti has to react near instantly to avoid catastrophe. There has to be minimal latency. This is where Edge AI comes in as you can’t wait for a cloud server in these response windows.
Then there is coordination between parts. The brain alone is insufficient. We need sensors to see, chips to process info, software to make decisions, motors to move, power systems to delivery energy, control systems to stay precise, and safety systems to prevent bad outcomes.
Sounds like a supply chain…right?
Good. Good.
So when we see demos, real deployment could look much different and further out. It’s one thing to have a machine work in isolation. It’s much different to being deployed every single day in real environments. So on top of all those things we just said these machines need to be able to do, they also need to be durable and last.
If this all sounds like its a major feat…good.
It is.
And that’s where the opportunity lies.
Because the companies that CAN do it. And CAN scale. Are positioned to make some real money.
The Core Loop
Practically every system we are discussing here has to go through a smilar cycle.
1 - See the world
2 - Understand what’s happening
3 - Decide what to do
4 - Check results
5 - Improve
Let’s take a practical example.
A warehouse robot.
First it has to see the aisle. It may use camera, lidar, depth sensors, or other sensors to understand where the shelves are, where the boxes are, and where other robots or people are moving.
Then it needs to take this and understand it. It needs to understand tahat one object is a box, one is a human etc. It needs to understand distance, speed, direction, and risk.
Then comes the decision. Should it keep moving? Slow down? Turn left? This is where perception starts turning into action.
Next up is movement. Motors spin, robotic arms adjust, gripper closes. Whatever it may be. We are now in the phyiscal.
Then it checks the result. Did it get to the right shelf? Did it grab the right item? This is a crucial step because the real world is unpredictable.
Lastly, it improves. It uses feedback from what happened to make a better decision next time. That’s basically how we operate as humans as well. Maybe it learns that a certain object is harder to grip or that a certain route gets congested at a certain time.
This feedback loop is what separates Physical AI from basic automation. Old automatic is a fixed instructive set. Physical AI adapts, watches reacts, adjusts.
All these steps is why the investible stack is so large. We’re talking about sensors, chips, software, motors, power, controls, safety systems, data. A robotic body alone is useless without the loop.
The Investible Stack
Now we can move into more specifics about what each layer contains.
We are not diving into companies yet. We need to lay the foundation.
I like the idea of comparing Physical AI to a person. Not in the humanoid/aesthetic aspect. But the functions & senses.
Humans need a brain, eyes, muscles, nerves, energy, balance, memory, and practice. A Physical AI system needs the machine version of all those things in one form or another.
The literal robot/machine that you see it just one part of the system. What lies behind the machine is much larger and more in depth. And where the public company exposure really lies.
Even before the stack starts, we have training. Before a machine can work int he real world, the AI has to learn. So we need data, models, examples of what good actions look like. So it may train on real-world data, synthetic data, simulation data, or human demos. It’s like a new employee learning how to perform a job before starting it.
Next up comes simulation. This is where the machine practices in a virtual world. So a warehouse robot can practice moving through digital aisles. A drone can practice flying through a digital battlefield etc. This is important because real-world practice is expensive, slower, and sometimes risky.
Next up is perception.
To read the rest of this report, click here:
I dive into the rest of the investible layers, where I can see this theme going, and what my process is going to be like.
Along with Optics, Physical AI is now the 2nd major theme I am covering in depth.
Full portfolio building
Unpacking all the layers
Presenting specific companies
Earnings coverage etc.
