Robotics

This isn’t a thesis about industrial robots

Industrial robots have been around for decades. They weld, paint, lift and place with a precision no human can match. They are also, almost without exception, doing one predefined thing in a space built specifically for them. Move the part two inches and the robot keeps swinging at where the part used to be. That’s a real industry with real economics, and it isn’t what interests me.

What I’m thinking about is general-purpose physical intelligence. A machine that can look at an environment it hasn’t seen before, work out what’s going on, decide what to do, do it, notice when it has got something wrong, and recover. A machine you give a new task rather than reprogram for one. In the long run I’m thinking about something much closer to general intelligence inside a body than to a better assembly line.

That distinction is the whole thesis.

Intelligence is one problem. Acting on the world is a different one.

T·01 covers how I think about intelligence becoming scalable, so I won’t repeat it. The link is simple enough: AI is the intelligence, and robotics is what happens when that intelligence can act on the physical world.

I don’t want to imply the second follows automatically from the first, because it doesn’t, and that assumption is where most robotics enthusiasm goes wrong.

The physical world is unforgiving in ways software isn’t. A model that produces a poor answer costs you a retry. A robot that misjudges a grip drops something, damages something, or hurts someone. Software can be wrong cheaply and often. Hardware can’t. And the things humans find easiest are the things machines find hardest. Abstract reasoning has turned out to be more tractable than picking up an unfamiliar object with the right amount of force. We have models that pass professional exams and robots that struggle with a bag of groceries.

So this isn’t AI with a body attached. It’s a harder problem that happens to depend on the first one being solved.

Can it actually do the work?

This is the centre of how I think about robotics, and where I part company with most of the coverage.

Tesla and others have talked about eventually making humanoid robots cheap enough for mass adoption, and those figures get a lot of attention. They aren’t where my interest sits, because price only becomes the question after capability is established.

A cheap clumsy robot is still a bad robot.

Walking across a stage isn’t the test. Moving a box in a controlled demonstration isn’t the test. A genuinely useful general-purpose robot has to work in environments that are messy, variable and not designed around it. It has to see, understand what it’s seeing, reason about it, act, and recover when something goes wrong. If an object isn’t where it expected, it adapts. If it meets a task nobody programmed, it works out an approach.

Reliability is the part people underrate. A robot that completes a task correctly most of the time is not most of the way to being useful. In a real operation the exceptions are what cost money, because every failure needs a human to notice and fix it, and that human is the cost you were trying to remove.

So the question isn’t whether someone can build an impressive humanoid. It’s whether it can do the work.

What actually exists today

The gap between demonstrations and deployments is where most of the money in this field will be made or lost.

Real humanoids are doing real work right now, in narrow ways. Agility Robotics reported in late 2025 that its Digit robot had moved more than 100,000 totes at a GXO facility under a multi-year robots-as-a-service agreement. Figure has reported an extended run on a BMW production line at Spartanburg. Apptronik has a pilot with Mercedes-Benz. Named customers, recurring schedules, money changing hands: a far higher bar than a video.

They are also constrained tasks in structured environments. Moving totes along known routes is genuinely valuable. It is not general-purpose physical intelligence, and the two get conflated constantly.

The most useful data point comes from the company I’m most positive about. On Tesla’s Q4 2025 earnings call in January 2026, Elon Musk said Optimus was still at an early stage and still in the R&D phase. Robots had done some basic tasks in Tesla’s factories, but weren’t in material use, because each new version of Optimus deprecates the one before it. On that call he didn’t expect significant production volume until roughly the end of 2026. A year earlier he had predicted Tesla would build around 10,000 of them during 2025. Whether the newer timeline held is one of the specific things I’ll be checking rather than assuming.

That deprecation detail is worth sitting with. It says the hardware is still changing fast enough that deploying it properly isn’t yet worth doing, which is a precise description of a technology that hasn’t arrived. I’d rather build a thesis on that than on a demo reel. It tells me the timeline is longer than the enthusiasm suggests, and that the people building it know it.

What happens if the capability problem gets solved

If general-purpose robots eventually perform a wide range of the physical work humans do today, the consequence isn’t a bigger robotics market. It’s a change in what labour is.

Human labour is scarce in a particular way. There are only so many people, they need rest, they need wages, and you can’t manufacture more of them on demand. Machine labour would be scarce differently. It could run long hours, do repetitive and dangerous work, and scale through manufacturing rather than population. That last point is the one I keep returning to, because it turns labour from something you recruit into something you produce.

The loop this creates is what I find most interesting. Today humans build the factories that build the machines. Machines capable enough to do general physical work could eventually help manufacture other machines, maintain the factories, build infrastructure, extract resources, move materials and operate energy systems. At that point productive capacity stops being limited by how many people are available to do the building.

Machine labour could become scalable productive capacity. That’s a different economic object from a robot you buy.

Human labour contributes substantially to the cost of a lot of goods and services, particularly labour-intensive ones. I won’t claim it dominates the cost of everything, because materials, energy, land, logistics and capital equipment are all still scarce. But if labour becomes far cheaper and far more scalable, an economy can produce considerably more with far less human input, and the cost of many things could fall. That’s the version of this I care about. Not how big the robot market gets, but what happens when labour itself stops being scarce.

Taken far enough it points somewhere unusual: a world where humans don’t have to work to survive. Not a world where humans don’t work. People will still build, create, compete, lead, research, entertain and chase status and meaning, because that’s what people do when survival stops being the constraint. But compulsory labour as the price of existing could become much less central. That’s a long-term scenario, not a forecast about this decade, and I hold it loosely.

The transition may not be smooth

I have no idea how society gets from here to there, and I don’t think anyone does.

Today’s economy runs on work producing wages producing consumption. If machines displace paid human labour meaningfully before the benefits of abundance are widely shared, the middle of that transition could be disruptive. There are proposed answers, from universal basic income to what Elon Musk has called universal high income, to models nobody has designed yet. I don’t have a view on which wins. I just think the current relationship between employment and income looks fragile if machines eventually perform a large share of productive labour. That’s a transition risk inside this thesis, not a political position.

If intelligence and labour become abundant, energy may not

I want to state this carefully, because it’s an inference rather than an established fact.

Robots, compute, factories, transport and infrastructure all consume power. If intelligence becomes cheap and machine labour becomes abundant, something physical still has to run all of it. Energy could become one of the defining constraints on how much an automated economy can actually produce.

There’s early evidence pointing that way, though it concerns data centres rather than robots. The IEA reported that global data centre electricity demand grew 17% in 2025, with AI-focused facilities growing far faster, and that more than 2,500 GW of projects sit stalled in grid connection queues worldwide. New grid infrastructure takes five to fifteen years to plan, permit and build, far slower than the things wanting to connect to it. That mismatch between how fast demand appears and how slowly power arrives is structural, not a feature of one year’s numbers.

Extending that to a future robot economy is my inference, not a proven claim. But it’s why I treat energy as part of this thesis rather than a separate topic.

Humanoid is probably the bridge

The humanoid form makes sense to me right now, for an unglamorous reason: we built this world around our own bodies. Stairs, doorways, tools, shelves, vehicles and workstations are all sized for human proportions. A machine with roughly human dimensions, arms and hands can work inside that world without anyone rebuilding it first.

That’s an argument about compatibility, not about the human form being optimal. In a future where machines increasingly design the factories and infrastructure that other machines work in, different forms may prove far more efficient. Wheels beat legs on flat floors. Four arms may beat two.

Humanoid may be the bridge into the world humans built, not the final form of robotics.

Why Tesla, and why that isn’t permanent

In T·01 I said plainly that I don’t know which company ultimately wins. In robotics I have a stronger view, and I want to be equally clear about what it rests on.

My reasoning is about vision-based physical intelligence. Humans operate mostly through sight. We look, interpret, reason, then act. Tesla has spent years on a structurally related problem in autonomous driving: perceive the environment, understand it, predict what happens next, decide, act. Tesla’s own stated position is that an approach built on AI for vision and planning, running on efficient inference hardware, is the path to a general solution across both self-driving and bipedal robotics. That’s the specific claim I find credible.

Solving self-driving does not solve robotics. Manipulation, hands, balance and contact forces are additional problems, and a car has the enormous advantage of moving through the world without touching it. So my belief is conditional, and I’d rather state the condition than hide it: if vision-based intelligence turns out to be one of the dominant paths to general-purpose physical AI, Tesla’s existing work in real-world perception and autonomy, combined with its manufacturing capability, gives it an important starting position. If that condition is false, most of the advantage evaporates.

As of August 2026, Tesla is behind on external deployment evidence. Agility and Figure have more documented customer work. What Tesla has is the perception stack, the fleet data and the ability to manufacture at scale, which are advantages that matter later rather than now.

That stronger view changes how I act, not just what I hold. Where my conviction is higher, I’m willing to size in more. That’s the practical difference between this thesis and T·01: there the uncertainty about eventual winners is wide enough that I spread the exposure, and here I don’t.

It’s also worth being clear about how the two theses relate, because Tesla could plausibly sit in either. It sits here. What I’m buying is physical AI, and Tesla is the vehicle for that. The language-model side of Musk’s world now lives somewhere else entirely: SpaceX acquired xAI and Grok in February 2026 and listed publicly in June, so it is a separate company with a separate set of questions. Grok appearing in Tesla vehicles is an integration, not common ownership.

None of this makes Tesla permanent. If another company demonstrates materially better reasoning, perception, dexterity, reliability, autonomous task completion, manufacturing capability or deployment economics, I’ll reassess. If it becomes investable and I think its probability of success is higher, I’m willing to move capital toward it.

My conviction is in general-purpose physical intelligence, not in a ticker.

Measuring useful work, not robot news

Because this field generates spectacular footage, I try to hold it to evidence footage can’t fake.

Producing thousands of robots doesn’t prove the investment case. A staged demonstration doesn’t prove it. A robot working in a heavily controlled environment doesn’t prove general capability. What I’m looking for over time is autonomous productive hours, task completion rates, how much human intervention is still needed, whether tasks are learned rather than hard-coded, whether real human labour is being substituted, what deployments cost to run, and whether there’s commercial demand from customers who aren’t also investors.

Most of that isn’t published yet, which is part of what makes this early.

What would make me wrong

The bear case I take most seriously isn’t that AI stalls. It’s that intelligence keeps improving fast while physical capability doesn’t.

That’s a real possibility, not a courtesy caveat. Hands are hard. Dexterity is hard. Balance under load is hard. Batteries are heavy and don’t last long. Hardware wears out. Safety around humans is a genuine constraint, not a formality. And training data for the physical world can’t be scraped the way text can, because someone has to generate it through actual physical interaction. Any one of these could stay stubborn for a long time.

The brain could get brilliant while the body stays clumsy.

If that happens, the transformation I’ve described either takes far longer than expected or arrives through specialised machines rather than general-purpose ones. There’s a serious argument that purpose-built automation keeps beating humanoids on cost and reliability for every task that can be specialised, leaving the general-purpose robot as a solution chasing the shrinking set of problems specialisation can’t reach. I don’t think that’s right, but I can’t dismiss it, and today’s deployments are more consistent with it than with my view.

Other things would reduce my conviction more slowly and less dramatically: robots continuing to need heavy human supervision, hardware economics failing to improve enough, energy requirements proving prohibitive, and deployments running for years without generating real returns for the companies operating them.

What wouldn’t shake me is the humanoid form losing. My thesis doesn’t require humanoids to win. If a different form proves far more efficient, that changes the implementation, not the argument.

The deeper failure would be this: AI becomes extraordinarily intelligent, and we never manage to turn that intelligence into scalable, economically useful physical work.

That’s what I’m actually watching for. Everything else is detail.