If you take the history of economic arbitrage decay seriously, the obvious follow-up is “what comes after AI commoditises cognition?” The defeatist answer is “we cannot know.” The more honest answer is that visionaries at each previous transition often did see what was next — Edison saw electricity, Gates saw computers in every home, Bezos saw e-commerce — by positioning at the frontier of the new capability.
A defensible candidate for the next arbitrage emerges if you split intelligence into two domains.
Knowledge IQ vs Spatial IQ
Most current AI work targets knowledge IQ: logical reasoning, data interpretation, language, code, synthesis. This is already very good and improving rapidly. Within 5–10 years, knowledge-IQ AI will likely saturate the way mental arithmetic saturated when calculators arrived. Spending years now sharpening “data analysis” or “critical thinking” looks a lot like the person in 1900 perfecting their sword technique.
The much less developed branch is spatial IQ: physical interaction with objects and environments. When knowledge AI tells you what to do, something still has to actually alter the physical world. That work has to be done by robots.
What is a robot, structurally?
Strip away the marketing and a humanoid robot is essentially legs + hands. The torso is plumbing. The head is sensors. Nearly all meaningful human interaction with the world happens through feet or hands. No one opens a door with their chest or pulls a chair with their thigh.
Of those two:
- Legs are largely a balance problem. Boston Dynamics, Optimus, and several Chinese humanoid projects have effectively solved bipedal locomotion. Walking, running, jumping, recovering from a push — all solvable, increasingly solved.
- Hands are not solved. They are combinatorially more complex. The same hand picks a pen with palm-down, a frying pan with palm-up (cylindrical handle, different grip), a wet glass with different friction tension, and a hammer with completely different posture. Babies learn this through trial and error; current robotics largely cannot.
Why hands are uniquely hard
A few reasons stack:
- Degrees of freedom. A human hand has ~27 degrees of freedom. The control space is enormous.
- Object generalisation. A baby that has never seen a frying pan will try to lift it like a pen, feel the slip, and adjust. Current robotic hands typically cannot recover from a misjudged grip on a novel object.
- Context entanglement. Folding a shirt isn’t a hand problem alone. It’s hand + vision + fabric physics + surface friction + the cluttered laundry basket. These elements don’t decompose cleanly.
- Sim-to-real gap. Training in simulation is cheap, but transferring to physical hands is brittle. Real-world data is expensive to collect — far harder than text or image data — and a handful of companies with millions of robots in the field will have compounding advantages.
Why humanoid form factor wins for shared spaces
A common pushback is that humanoid robots are unnecessary — washing machines don’t mimic hand-scrubbing, so why should a laundry robot? The answer comes from where the robot needs to operate.
Industrial robots don’t need to be humanoid. Factories already use conveyor belts, robotic arms on rails, and purpose-built machines optimized for one task. There is no humanoid robot in an Amazon fulfilment center, and there shouldn’t be.
Home robots almost have to be humanoid. The entire built environment — doors, drawers, counters, switches, faucets, stairs, knobs, hanging spaces — is designed for a bipedal creature with two arms. Trillions of dollars of infrastructure assumes the human form factor. A non-humanoid laundry robot would have to either:
- Build conveyor belts through every room (impossible in a shared living space), or
- Have a new house designed around it (impossible at scale)
The laundry example makes it concrete: take clothes out of the basket, walk to the washing machine, open its door, load the clothes, add detergent, close, start. Later: take wet clothes out, walk to the balcony, hang them on the line. Later: collect dry clothes, walk back, fold, store in wardrobe. That’s a sequence across multiple rooms requiring navigation of human-scale spaces and manipulation of human-designed objects. The same logic applies to cooking, cleaning, and almost every meaningful home task.
Humanoid form factor is not an aesthetic choice. It’s a forced consequence of the environment robots have to operate in.
The 1985-coding analogy
In 1985, learning to code was an obscure technical skill. By 2015 it was the dominant economic skill of a generation. People who picked up programming early — even before knowing what specifically they would build — were positioned for a wave that hadn’t fully arrived yet.
The same bet on spatial intelligence today looks like:
- Robotics research and engineering
- Sim-to-real transfer
- Manipulation learning
- 3D perception
- Tactile sensing
- Reinforcement learning for control
- Foundation models for robotics (RT-2 and successors)
These fields are where current researchers and startups are still small in number. They will likely look obvious in 10 years the same way “should have learned to code in 1985” looks obvious now.
The honest counterarguments
Worth keeping in mind:
- Timeline risk. “Robotic hands are 5 years away” has been true for 20 years. It might genuinely be a 10–15 year problem, which is hard for individuals to wait out.
- General-model risk. Knowledge AI did not develop as a modular ecosystem — we ended up with one model handling almost everything (GPT, Claude). Spatial AI might similarly skip the “module marketplace” phase entirely, in which case the value accrues to whoever trains the one general manipulation model, not to specialist developers.
- Data moat risk. Whoever has millions of robots in the field collecting manipulation data (likely Tesla, Figure, or similar) has a compounding advantage independent module developers can’t match.
- Form factor risk. Some tasks might still be served better by specialised non-humanoid robots than by general humanoids — even at home. A roomba is the existence proof.
None of these kill the broader thesis (“spatial intelligence is the next frontier”) but they affect how to bet on it. The safer bet is to build expertise in the field rather than to predict which specific company or product wins.