
Physical AI will be won by whoever learns fastest in the physical world. Most incumbents already own the scarcest input, and they are letting it evaporate.
In late 2024, a Chinese battery executive offered a one-line autopsy of Northvolt, Europe’s best-funded bet on battery independence. “We can raise a factory’s battery yield to 96% in just four months,” he said. “Northvolt took four years to reach only 70%.”
Northvolt did not fail because its chemistry was flawed or because it lacked capital or customers. It failed because it could not learn quickly enough.
By March 2025, the company had filed for bankruptcy in Sweden, with about $8 billion in debt. In its final statement, Northvolt noted that it had improved production yield by 50% over the previous six months. It was learning. It was not learning fast enough to outlast its cash.
That is the clearest description of physical AI, and it does not involve a single humanoid robot.
The public conversation frames physical AI as a hardware race: who builds the most capable humanoid, who trains the best robot foundation model. That is the wrong lens. Machines have operated in the physical world for a century. What is new is that the cost of learning in the physical world is about to fall.
In atom-based industries, the rate of learning has always been the most important competitive variable. Physical AI is about to make it the variable that determines who sets prices.
The misdiagnosis
Capital has chosen its narrative. Robotics startups raised $18.8 billion in the first half of 2026, more than in all of 2025. Barclays Research projects a $200 billion humanoid market by 2035. Nvidia’s Jensen Huang declared in January that “the ChatGPT moment for robotics is here.”
Now look at what was deployed. In 2025, factories installed more than 600,000 industrial robots, bringing the global operational stock to a record 5 million. Worldwide humanoid shipments that year: 13,317.
The point is not that humanoids are irrelevant. The point is where value will accrue. Both the hardware and intelligence layers are on paths to commoditization. Barclays estimates that humanoid production costs have fallen 30-fold over the past decade.
Physical Intelligence open-sourced the weights for its first generalist robot policy, π0, in February 2025. Nvidia released open-world models and an open vision-language-action model in January 2026.
When the most visible layers of a stack become cheaper and more open at the same time, value migrates to what does not. In physical AI, that is three things a competitor cannot download: experience data, the capacity to verify it, and an operating model that lets the organization act on what it learns. Robots and models are how you collect experience, but they are not the asset.
Wright’s Law, unbundled
In 1936, aeronautical engineer Theodore Wright observed that each doubling of cumulative aircraft production roughly halved unit costs. The pattern has held across industries ever since. Solar module prices have fallen about 20% with each doubling of installed capacity, from $106 per watt to $0.38 over four decades.
Wright’s Law carries a hidden assumption that has shaped industrial strategy for a century: experience came bundled with volume. To learn, you had to make things. The firm that shipped the most units descended the curve fastest, and that is why scale was the moat and why late entrants had to buy volume to buy learning.
Many ran out of money before the learning arrived. Northvolt is the cautionary case, not the exception.
Physical AI unbundles experience from volume. An autonomous lab runs hundreds of characterized experiments, while a conventional team runs a dozen. A simulation calibrated to plant reality explores process windows no one would risk on a live line; a production run, captured with its decisions and outcomes, becomes a labeled trial rather than just output. Each generates experience without requiring a sale.
The arithmetic is unforgiving. A firm that learns 30% per doubling of volume, instead of 20%, reaches the same cost position with roughly one-fifth of the cumulative production. That is not a productivity gain. It changes who can enter a market, how quickly an incumbent can be caught, and how much capital it takes to reach a competitive cost.
Physical AI is a way to buy doublings of experience without buying doublings of volume.
There is no internet for atoms.
Large language models were built on a gift: the internet, a free, pre-assembled record of human language. Physical AI has no equivalent. YouTube adds roughly 300 million hours of video each year; π0 was reportedly trained on about 10,000 hours of robot data.
Video helps, but it only shows what happened. It does not record force, contact, or the action that caused the outcome.
So the training data for physical AI has to be generated where physics happens: in labs, on production lines, in the field. That is why the most aggressive capital in the category is going into facilities, not just models. Periodic Labs raised a $300 million seed round to build automated labs that synthesize and characterize materials. Lila Sciences, with about $550 million raised, is building roughly 22,000 square meters of automated lab space.
Here is the provocation for incumbents. Startups are spending hundreds of millions of dollars to build places where physical experience can be generated. Incumbents already own those places: reactors, coating lines, furnaces, fabs, pipelines, and field fleets that run every day. Yet most of that experience evaporates.
It evaporates because telemetry is not experience. A plant historian records that the temperature fell from 182 to 176 degrees, but rarely records that an operator changed the setpoint, why she did so, what she saw, or whether yield improved three shifts later. Experience is action, context, and outcome, linked and trusted. Most industrial operations were designed to produce output, not learning; that design choice is now a strategic liability.
The decade is in the middle
Ask where AI for science gets its headlines, and the answer is discovery. Ask where the time goes, and the answer is elsewhere. A new material typically takes 10 to 20 years to move from discovery to commercialization, according to the U.S. Department of Energy. Qualification alone can take three to five years.
The expensive part is rarely finding a candidate. It is turning a result that worked once in a lab into a process that works every time, at volume, at a price someone will pay. A program that shortens a qualification cycle by a year is worth more than one that generates ten thousand more candidates.
Generation is cheap. Verification is the moat.
In November 2023, a Nature paper reported that an autonomous lab at Lawrence Berkeley National Laboratory, the A-Lab, had synthesized 41 novel inorganic compounds in 17 days. Solid-state chemists challenged the claim, arguing that the AI had misidentified phases in the X-ray diffraction data. In January 2026, Nature published a correction. The headline result was revised to 36 compounds from 57 targets, and the word “novel” was removed.
Both halves of the lesson matter. The autonomy worked: robots ran for 17 days and produced real compounds. What failed was the judgment layer, the step that determines whether a result means what you think it means.
AI reduces the cost of generating candidate ideas without reducing the cost of choosing among them. Physical systems make that problem sharper. In software, an unverified output is a bad answer. In a plant, it is a rejected batch, a recall, a safety incident, or a qualification failure discovered two years later.
The frontier of robot learning points the same way. According to the company, Physical Intelligence’s π*0.6 more than doubled throughput and roughly halved failures on its hardest tasks. The gains did not come from more pretraining; they came from learning on the job, with human teleoperators correcting the robot when it erred. Human judgment was not removed from the loop; it was what made the loop work.
This is where most incumbents go wrong. They move right to the sensors, cameras, a robot cell, and an AI platform, and land in what I call the sensor landfill: oceans of data and autonomous output that no one trusts enough to change a recipe. Eighteen months later, the conclusion is “AI doesn’t work in our plant.” The better path builds verification capacity first or in lockstep.
The geography of experience
If experience is the asset, follow where it is accumulating. In 2025, China installed 354,000 industrial robots, 59% of the global total. The United States installed about 38,500. China accounted for roughly 85% of humanoid deployments.
The West leads in frontier models. That lead matters less than it appears if physical AI rewards accumulated experience over model quality. Under Wright’s Law, the economy that compounds doublings fastest sets the cost curve, and everyone else prices against it. The executive’s line about Northvolt was not a boast, but a statement about the learning rate.
What should leaders do?
- Manage learning velocity, not automation coverage: Robots deployed, use cases live, and sensors connected do not measure learning. Measure the cost and time to produce one validated learning, meaning a result trusted enough to change a recipe, a setpoint, or a design. Then manage that down.
- Instrument decisions, not just machines: Capture setpoint changes, operator overrides, rework calls, and engineering change requests, and link them to what happened next. It is unglamorous data engineering. It is also the asset.
- Fund the middle: Shift investment from discovery headlines toward scale-up and qualification, where time and capital are spent.
- Build verification before generation: If you cannot characterize and trust new output at the rate you generate it, more generation produces more landfill, faster.
- Own the loop, rent the model: Negotiate data rights in every robotics, platform, and lab-automation contract as if they were intellectual property, because they are. A vendor that trains its next model on your process data is effectively buying your learning curve at a discount.
- Turn experts into trainers: Your best process engineers and operators are the human corrections in the loop. Capture their judgment while you still have access to it.
The machine was never the advantage.
Go back to the factory that reached 96% yield in four months. It did not have a humanoid. It had a faster loop: engineers who had ramped lines before, equipment instrumented to show them what was going wrong, and the discipline to change the process daily based on what the line taught them. That loop used to take a decade and a supply chain to build.
Physical AI makes it buildable, partly purchasable, and, for the first time, partly independent of volume. That is the opportunity for incumbents who own the places where physics happens and the threat from anyone who learns faster there.
Software ate the world because copying was free. Atoms will not be copied. They will be learned, and the learning compounds for whoever starts first.


