I wrote recently about what happens when shipping software is no longer the hard part. The short version: the bottleneck moves from execution to judgment. From "can you build it" to "should you build it."
But that essay only covered half the story. Because software is not the endgame. The endgame is physical.
The screen was a constraint
Everything AI has done so far has been trapped behind glass. It writes code, generates images, drafts emails, analyzes data - all inside the digital world. Impressive, but fundamentally limited. The digital world is a small fraction of the economy. The vast majority of value creation happens in physical space: manufacturing, logistics, construction, agriculture, energy, healthcare.
The screen was a constraint, not a destination. And that constraint is breaking.
What physical AI actually means
Physical AI is not just "robots." It is AI that can perceive, reason about, and act in the physical world. That includes:
- Humanoid robots that can navigate unstructured environments - warehouses, construction sites, homes
- Autonomous manufacturing where factories reconfigure themselves based on demand
- AI-driven logistics that can load, route, and deliver without human coordination
- Embodied agents that can inspect, maintain, and repair physical infrastructure
- Autonomous labs that can run experiments, synthesize materials, and iterate on designs
Each of these exists in some form today. Boston Dynamics robots walk. Autonomous vehicles drive. Robotic arms assemble cars. But they are narrow, brittle, expensive, and require enormous engineering effort to deploy.
That is changing at the same speed that software changed. And for the same reason: the AI models are getting good enough to handle the complexity that kept these systems bottlenecked.
The key unlock is foundation models for the physical world. The same transformer architecture that learned to predict the next word in a sentence is now learning to predict the next action in a physical environment. Vision-language-action models can see a scene, understand a natural language instruction, and execute a physical task. This is not science fiction. Google DeepMind, Tesla, Figure, 1X, and a dozen others are shipping systems that do this today.
The double collapse
Here is what gets interesting. My previous essay argued that when software shipping becomes free, four things become valuable: problem discovery, taste, distribution, and judgment.
Physical AI triggers the same collapse - but for atoms.
In the old world, manufacturing something required factories, supply chains, tooling, logistics infrastructure, and enormous capital. The gap between "I know what to make" and "it exists and people have it" was even larger in physical space than in digital space.
Physical AI collapses that gap. Not all at once. Not overnight. But the trajectory is clear.
When a humanoid robot can assemble a product from a CAD file, the factory is software. When an autonomous truck can deliver it, the supply chain is software. When an AI agent can design the product by understanding customer needs, the R&D process is software.
At that point, the cost of creating and distributing a physical product approaches the cost of creating and distributing a digital one. And the same bottleneck shift happens: the hard part is no longer making the thing. The hard part is knowing what to make.
Why this matters more than software
Software eating the world was a $5 trillion story. Physical AI eating the world is a $50 trillion story. Because the physical economy is an order of magnitude larger than the digital one.
Global manufacturing: $16 trillion. Logistics: $9 trillion. Construction: $13 trillion. Agriculture: $4 trillion. Healthcare delivery: $9 trillion. Energy operations: $6 trillion.
These industries have been largely untouched by the AI revolution so far. Not because AI is not useful in them - it is enormously useful. But because the interface between AI and the physical world was too primitive. You could use AI to optimize a schedule or predict a failure, but you could not use it to actually do the physical work.
That interface is being built right now. And when it is ready, every argument about software applies to the physical world, amplified.
The new infrastructure layer
Physical AI does not run on the same stack as digital AI. It requires a different infrastructure:
Compute at the edge. You cannot send video from a robot to a cloud data center and wait for a response. Physical actions require millisecond latency. This means running capable models on local hardware. Neuromorphic chips - processors designed to run neural networks at 1,000x less power than GPUs - become critical. The robot needs a brain in its body, not in a server farm in Virginia.
Energy density. A humanoid robot doing physical work for eight hours needs energy. A lot of it. Current battery technology is adequate for some applications but not for heavy industrial work. This is where energy infrastructure - grid capacity, generation, storage - becomes directly relevant to AI deployment.
Simulation environments. You cannot train a physical AI agent by letting it break things in the real world for a million iterations. You need simulation environments that are physically accurate enough for trained behaviors to transfer to reality. This is an unsolved problem at scale. The sim-to-real gap is the single biggest bottleneck in physical AI today.
Standards and interoperability. A robot from Company A needs to work alongside a robot from Company B in a warehouse designed by Company C. The physical world does not have APIs. It needs them. MCP (Model Context Protocol), A2A (Agent-to-Agent), and emerging standards for embodied AI are the early plumbing of this new layer.
Each of these is a massive market opportunity. And they are all prerequisites - physical AI does not scale without them.
What humans do when physical work is automated
This is the question everyone asks, and most people answer badly. The pessimistic answer: mass unemployment, social collapse, dystopia. The optimistic answer: everyone becomes a creative, we all do meaningful work, utopia.
Both are wrong because both assume a binary transition. The reality is a gradient. Physical AI will automate specific tasks within jobs long before it automates entire jobs. A construction worker with a robot assistant builds faster and safer. A surgeon with a robotic system operates with more precision. A farmer with autonomous equipment manages more land.
The pattern from software repeats: the people who thrive are the ones who move up the abstraction stack. From doing the physical work to directing it. From operating the machine to deciding what the machine should do. From execution to judgment.
And the same bottlenecks apply:
Problem discovery. What physical problems are worth solving with automation? Not everything benefits from a robot. Understanding where physical AI creates real value requires deep domain knowledge.
Taste. A robot can assemble a chair. It cannot decide whether the chair should exist, what it should look like, or how it should feel to sit in. Product sense, design intuition, and understanding of human needs become more important, not less.
Distribution. Building a physical product with AI-assisted manufacturing is one thing. Getting it to the right people, in the right context, at the right time, is the same distribution challenge it has always been.
System integration. The physical world is messy. Factories have legacy equipment. Warehouses have constraints. Supply chains have dependencies. The people who understand these systems - and can integrate new AI capabilities into them without breaking everything - are the most valuable people in the room.
Where the two collapses meet
The interesting thing about the simultaneous collapse of software and physical barriers is what happens at the intersection.
A person with deep domain knowledge in, say, industrial heat processes can now:
- Build a software tool to optimize those processes (software is free)
- Design a physical system to implement the optimization (physical AI is emerging)
- Deploy both to a specific customer (distribution still requires relationships)
Ten years ago, that person would have needed a software team, a hardware team, a manufacturing partner, and $10M. Now they need a laptop, an AI subscription, and access to a robotics platform.
The founder who understands a physical domain deeply, and can leverage both software AI and physical AI to solve problems in it, has an advantage that is nearly impossible to compete with. Because the domain knowledge is the moat. Everything else is becoming infrastructure.
This is why I work across energy, AI, and neuromorphic computing simultaneously. They are not separate bets. They are layers of the same stack. Energy powers the physical world. Neuromorphic computing gives AI a body. Software AI gives it a mind. The founder who holds all three layers has leverage that compounds.
The timeline
Software AI made shipping code free in roughly 2024-2026. Two years from "interesting tool" to "structural transformation."
Physical AI is earlier on the same curve. Foundation models for robotics are where language models were in 2022 - impressive demos, limited real-world deployment, rapid improvement. If the curve follows a similar trajectory:
- 2026-2027: Physical AI capable enough for structured environments (warehouses, factories with controlled conditions)
- 2028-2029: Capable enough for semi-structured environments (construction sites, agriculture, healthcare facilities)
- 2030+: Capable enough for unstructured environments (homes, cities, wilderness)
This is speculative. Physical AI is harder than software AI because the real world has physics, and physics does not forgive errors the way a text editor does. But the rate of improvement is steep, and the investment is enormous.
The only question
When building software is free and building physical things is approaching free, the only question that matters is: what should exist that does not yet?
Not what can you build. Not what would be cool to build. What should exist? What gap in the world, once filled, makes everything downstream work better?
That question requires understanding people, systems, physics, economics, culture, and politics. It requires judgment. It requires taste. It requires the kind of thinking that no AI system can replicate, because it depends on caring about the outcome in a way that only humans do.
The tools have never been more powerful. The barriers have never been lower. The only bottleneck left is knowing what matters.
That has always been the real job. We are just finally running out of excuses to avoid it.