A Secret Option in Robotic and Vision of Agents

Please before diving in, read this disclaimer.

If this is your first time here, welcome. I look for listed companies the market has mislabelled, businesses still filed under a sleepy legacy sector long after their technology quietly became critical infrastructure for one of the great megatrends. The really interesting money, I think, is rarely in the obvious name everyone already owns. It is in the company hiding in plain sight, sitting on a chokepoint nobody has bothered to value. Every issue, I find one, take it apart, and tell you exactly where the mispricing is and where I might be wrong.

These past few weeks I have been running a themed season, the Agentic Economy. AI agents are about to climb out of the chat box and start acting in the real world, and the question I keep asking is not which model wins, but which listed companies own the layers those agents will actually run on.

So far this season I have written about Snowflake, about Reddit as the ground the agents stand on, their layer of human trust, and about a third company I am keeping behind the paywall for now.

Today’s piece is the next layer down, the most physical one of all

Before an agent can act, it has to see

If you want the full thesis, the company, and the honest version of the trade, the rest is for paying subscribers. Subscribe below, and read on.

In the summer of 1966, Seymour Papert sat down at the Massachusetts Institute of Technology and wrote a memo that has since become one of the quietly humbling documents in the history of computing.

Papert, a South African mathematician who would go on to co-invent the LOGO programming language and shape how a generation of children met computers, set his students a task he and his colleague Marvin Minsky believed was tidy enough to finish over a single summer break. Connect a camera to a computer. Have the machine look at a scene, separate the objects from the background, and describe what it saw. A young Gerald Sussman (studying AI since 1964) coordinated the group.

The plan was elegant, the deadline was September, and the optimism was total. I would have loved witnessing it…

Sixty years later, university courses on computer vision are still organised around the walls those students hit that summer.

This is the oldest joke in artificial intelligence, and it is not really a joke. Hans Moravec gave it a name in 1988: “the things that feel effortless to a one-year-old, seeing a room, recognising a face, catching a ball, turn out to demand staggering amounts of computation, while the feats long ranked as the summit of human intelligence, chess, calculus, passing the bar, fall comparatively easily to a machine”.

The field poured its genius into the hard problem and assumed the easy one would solve itself over a summer. It did not because the hard problem was not thinking, but seeing.

I keep coming back to Papert’s memo because the entire investment world has, in 2026, forgotten it again.

Walk into any conversation about the agentic economy this year and you will hear about brains. The models that reason, plan, write code, and book your flights.

You will hear the numbers too, the ones that have made every adviser, every retail buyer, and every index committee a little drunk: the coming mega-listings, the capital being raised against projects that do not yet make money, the tokens sold below the cost of producing them. I wrote about some of this and my scepticism there has not softened.

But there is a quieter shift underneath the noise, and it is the one that matters for the next decade, in my opinion, of course. The agent is climbing out of the chat box and into the world. It is being asked to drive, to pick, to inspect, to walk across a warehouse floor and not crush the person standing in it…

Poor kid…

Earlier in this series I argued that Reddit was becoming the ground the agents stand on, the layer of human trust they reach for when they need to know what is real. Today I want to go lower in the stack, to the most physical layer of all. Before an agent can stand anywhere, it has to see. Like, you, like me. Not see in the poetic sense like ho what a beautiful sunset, more like in the sense of turning photons into a decision, fast enough and cheaply enough to matter, in a world that does not pause for the cloud to think or being out due to whatever reason AWS is giving us;

Let me just explain a bit why that is harder than it sounds, because the difficulty is the whole investment case.

A conventional camera, the kind in your phone or bolted onto most of today’s robots, works like a flip book.

It captures the entire scene at fixed intervals, thirty times a second, sixty, perhaps a hundred and twenty and between two frames it is blind. It records the wall behind a moving hand with exactly the same diligence as the hand, generating a flood of data that is mostly redundant and it carries a built-in delay, because the fastest it can react is the length of the gap between two snapshots. For a holiday photograph, none of this matters. For an agent reaching for a moving object, or a delivery robot reading a junction, or a cobot working a hand’s width from a human arm, all of it matters at once.

Don’t be fooled: the robot’s limit is not the cleverness of its model but the speed, the power budget, and the autonomy of its eyes.

A genius brain wired to slow, greedy, cloud-dependent eyes is a genius that arrives at every decision a beat too late.

With robots in factories, or with guns, it’s not possible.

So here is the question I find genuinely interesting, and it is not the question the market is asking. Who supplies these dawn eyes?

There is one listed company that, almost by accident of its history, sits on the physical layer of machine sight.

It does not build robots though.

And it does not train models either.

It does not appear on a single slide at the conferences where the agentic economy is debated... (You’ll understand soon enough why)

Most investors file it under a category that has nothing to do with artificial intelligence, and that act of misfiling, I will argue, could be the entire opportunity.

The market is pricing a legacy toy.

I think it owns a piece of the future that nobody has bothered to value, yet.

Before I name it, sit with the questions that pulled me into the filings in the first place. These are the question I write on a page when I dig the company. It easier for me after to create a thesis. Every investor has its methodology, this is mine. I am really curious.

• What if the most defensible position in the coming robotics boom belongs not to a maker of robots, but to the company that quietly supplies their eyes? (the starting point of the thesis).

• There is a listed company that already controls more than half of the world market for a component every agentic robot will need. Why does the market still price it as something else entirely? (the issue)

• Earlier this year, the most advanced demonstration ever published of a machine beating human professionals at a fast, physical game appeared on the cover of one of the world’s most prestigious scientific journals. What was the component hidden at the centre of that machine? (really incredible shit by the way)

• What changes for the economics of a factory floor the moment a robot can see a human and stop in roughly three thousandths of a second, rather than the hundreds of milliseconds an ordinary camera needs?

• One of this company’s sensors sends no pictures at all. It sends only change, pixel by pixel, in millionths of a second. Why might that single design choice matter more for robotics than any further leap in artificial intelligence?

• That sensor was built with a small French partner, and no rival has matched it at volume. What makes it so hard to copy? (moat)

• In May this year the company signed a preliminary agreement with the world’s largest chip foundry to build a dedicated plant for next-generation sensors, aimed explicitly at physical AI. Why might that be the opening move in a much longer sequence?

• The company’s robotics software does not lock its customers in commercially. It locks them in architecturally. What is the difference, and why does it compound? (it’s important you don’t miss this part)

• The thesis I am about to make is, I will admit, a strange one. So what is the real downside, and why would I want to own this anyway?

I have spent the past week inside this company’s annual report, (most of it was boring as fuck), its most recent filing with the American regulator, the datasheets for its sensors, and a paper published this spring in Nature.

I have come out the other side with a thesis unusual enough that I would rather make the case slowly than sell it quickly, below the line I name the company, show you how its components actually work and what they make possible, answer each of these nine questions in turn, and then set out the honest version of the trade, including the parts I do not like and the reasons it might not work and if you want to understand the one infrastructure layer of the agentic economy that almost nobody is pricing, this is the train, catch it.

This is the stock idea of this week.

The remainder of this letter is for paying subscribers.