The Physical World Gap

We talk a lot about how quickly artificial intelligence is getting better. Models gain capability in monthly increments, software is written in minutes, and an improved AI system can in theory be available to millions of people overnight.

Out of that development grows a fascinating idea: once AI can take over almost any cognitive work and robots do the same with physical work, the cost of goods and services falls dramatically. Some visions go further. Work becomes optional, abundance becomes the normal state, money eventually becomes meaningless.

Technologically, that idea is remarkably coherent. Socially, it is far more complicated. Because between what technology can do in theory and what actually happens in the real world lies an enormous gap.

I call it: The Physical World Gap.

The Physical World Gap is the widening distance between the pace at which technological capability emerges and the pace at which the physical and institutional world can absorb, implement and scale it.

And from that follows the thesis this piece sets out to justify: the more capable AI becomes, the more the bottleneck of progress shifts from invention to implementation. The scarce question is no longer "Can we build it?" but "Can reality absorb it?"

Software scales. Reality has to be built.

The digital world has an extraordinary property: it can be copied. An algorithm is developed today and deployed a million times tomorrow. An additional digital copy costs almost nothing.

The real world works fundamentally differently. A second house needs more material. A second robot needs more motors, chips, sensors and energy. A new data centre needs land, transformers, power lines and cooling. Copper has to be mined, processed, transported and eventually recycled.

AI can improve all of these processes and automate large parts of them. But it cannot compile matter. Intelligence scales digitally. Matter does not.

A few numbers show how large that difference is.

1. Raw materials: a design takes hours. A mine takes 18 years.

Suppose an AI designs a significantly better battery tomorrow. The design is finished within a few hours. Then physical reality begins.

According to S&P Global, mines that entered production between 2020 and 2023 took an average of 17.9 years from the discovery of the deposit to first output. The trend is rising: between 2005 and 2009 it was still 12.7 years. Most of that time is not spent building, but on exploration, studies and permits.

That is the core of the problem. AI compresses the design phase into hours. But the design phase was never the bottleneck. Even if robots eventually take over every step, the mine still has a maximum extraction capacity, the port a maximum throughput, and the ship still needs weeks to cross the ocean.

Robots do not eliminate these systems. They automate them.

2. Energy: more than 2,500 gigawatts waiting for a grid connection

Almost every vision of a highly automated economy comes down to the same resource: energy. AI needs data centres, data centres need electricity, robots need electricity, and so do the factories that build both.

Demand is already accelerating: the International Energy Agency expects data centre electricity consumption to roughly double by 2030, to around 945 terawatt hours. That is approximately the current electricity consumption of Japan.

Supply is not keeping up. Worldwide, according to the IEA, more than 2,500 gigawatts of generation, storage and large-load projects are stuck in grid interconnection queues. In the United States alone, around 2,060 gigawatts were waiting at the end of 2025, and the median wait from interconnection request to operation was more than five years according to Lawrence Berkeley National Laboratory.

An AI model gets markedly better in months. A power plant, a substation or a high-voltage line does not. In Germany the pattern is familiar from projects like SuedLink: well over a decade passes between the first plans and operation.

An abundance economy would therefore not just be an AI project. It would be one of the largest energy and infrastructure projects in human history.

3. Infrastructure: a robot needs a world it can work in

An intelligent robot on its own changes no economy. It needs an environment: roads, grids, ports, warehouses, factories, charging points, maintenance.

It is one thing to let a robot work in a controlled factory hall. It is something else entirely to integrate millions of machines into a chaotic physical world that grew over decades and was never built for autonomous systems.

The decisive question is therefore not only: how good will our machines become? But also: how quickly can we change their environment?

4. Supply chains: automation removes labour, not the chain

Imagine a fully autonomous copper mine. Robots drill, autonomous trains take the ore to the port, autonomous ships carry it, autonomous trucks deliver it to the factory. Human labour in that chain could approach zero.

But the supply chain still exists. The mine can fail, a port can be blocked, a country can restrict exports, a war can cut a transport route. Anyone who has worked with supply chains since 2020 knows every one of these scenarios from practice.

Automation removes people from a supply chain. It does not remove the supply chain.

5. Regulation: capability is not permission

The clearest illustration is autonomous driving. Google started its self-driving project in 2009. The technology demonstrably works today, Waymo operates commercially without safety drivers. And yet in 2026, seventeen years later, this is happening in only a handful of metropolitan regions.

Why? Because the decisive questions stopped being technical long ago. Who is liable in an accident? What error rate do we accept? Who certifies the system? How do insurers price the risk? Which authority has jurisdiction?

Technological capability does not mean social permission. The more AI leaves the digital world and interacts directly with people, the more that difference matters.

6. Trust: people have to hand over control

A technology can work objectively and still not be accepted. An autonomous vehicle can drive more safely than a human on a statistical basis, and parents still find it hard to put their child in alone. An algorithm can be superior at a diagnosis, and the patient still wants to speak to a doctor.

People do not optimise for statistical efficiency. We respond to trust, control, responsibility and habit. A technology can be technically ready long before a society is ready to trust it. More compute does not solve that problem.

7. Habits: people do not install updates

Occupations become part of our identity. Companies develop processes, public authorities develop procedures, societies develop norms. And as soon as those structures exist, interests in preserving them emerge.

A company could in theory become considerably more efficient through AI and still barely change its organisation for years. Because an org chart does not only show how work is distributed. It also shows how power is distributed. Technological change almost always puts that distribution in question.

8. Incentives: every efficiency gain creates losers

When an AI automates ten processes, ten efficiency gains appear. At the same time, existing structures lose significance: occupations, departments, business models, status.

Whoever loses has a rational incentive to slow change down. That is not backwardness, it is incentive structure. So it is not enough for a technology to be objectively better. A society has to give enough people a reason to support its adoption.

9. Institutions: code scales globally, law stays local

Software can be rolled out worldwide in weeks. Laws have to be debated, jurisdictions clarified, liability questions answered, international standards negotiated. And different states arrive at different answers.

An AI system could in theory run tomorrow in Munich, San Francisco, Shanghai, Lagos and Mumbai at the same time. But it meets completely different legal orders, infrastructures and cultures there. Anyone forecasting a global technological future therefore has to forecast more than technology. They have to forecast the institutional development of nearly every society in the world.

10. Geopolitics: the world is not a closed system

There is no such thing as "the world economy" as a single actor with a shared goal. There are states that control raw materials, protect industries, restrict exports and compete for technology.

The more important compute, energy, semiconductors and certain raw materials become, the greater the strategic value of controlling them. Technological abundance can therefore, paradoxically, create new geopolitical competition rather than end it.

11. Scale: a prototype is not yet a civilisation

Building a machine that can do something is one thing. Integrating a billion of them into the world economy is an entirely different problem.

For context: according to the International Federation of Robotics, global industry currently installs a little over half a million industrial robots per year. The vision of a largely automated physical economy would require not hundreds of thousands of machines, but billions. Which in turn would require factories, raw materials, chips, batteries, energy, maintenance, capital and approvals on a scale that has never existed.

The irony: we first need enormous amounts of the existing economy in order to build the automated economy at all. A prototype proves technical capability. Only scale changes society. In between lies the Physical World Gap.

The real bottleneck is time

Almost all of these problems share one feature: they are solvable. We can produce more energy, automate mining, expand infrastructure, change laws, build trust.

But these processes run at completely different speeds. AI changes in months. A mine takes 18 years. A grid connection takes five. A transmission line more than a decade. A legal system, a public authority, a corporate culture move at their own tempo. And billions of people all the more so.

That is why the decisive variable eventually stops being technological capability. It is synchronisation.

The Physical World Gap Framework

An AI-driven transformation can be thought of in four layers:

1. Capability

What can the technology do? AI produces intelligence, knowledge, decisions, designs and software.

2. Translation

How does that intelligence reach the physical world? Robotics, autonomous machines and intelligent infrastructure translate digital intelligence into physical action.

3. Adoption

Is society allowed and willing to deploy that capability? Laws, liability, trust, habits and political decisions determine actual adoption.

4. Scale

Can the system become large enough to matter socially? Energy, raw materials, factories, supply chains and geopolitics determine scale.

Only when all four layers work does technological capability become social reality. Many forecasts essentially extrapolate the first curve alone. But society changes at the pace of the slowest one.

How do we close the gap?

The answer cannot be to slow technological progress. We have to accelerate the other side. Four levers strike me as central:

Robotics has to become the bridge. The greatest social value of robots may in the long run not lie in replacing individual human tasks. It lies in being able to change the physical world itself faster: robots that build factories, operate mines, maintain grids and recycle materials. The more we automate the production of physical infrastructure itself, the faster the real world can respond to digital progress.

Infrastructure has to become machine-ready. Instead of only building machines that cope with a world built for humans, we should gradually optimise the physical world for machines: autonomous ports, robot-ready factories, standardised interfaces. The big productivity jump probably does not come from the perfect humanoid robot, but from the combination of intelligent machines and an environment built for them.

Regulation has to learn faster. Regulation does not need to disappear, quite the opposite. But it has to become more iterative: regulatory sandboxes, limited approvals, real test areas, data-based thresholds, stepwise expansion of autonomy. Instead of only distinguishing between permitted and forbidden, institutions can themselves become capable of learning.

Incentives have to be aligned. People often resist change not because they fail to understand it. Sometimes they understand it very precisely and recognise that they lose out. If AI creates enormous productivity while the gains are extremely concentrated, resistance becomes rational. A successful transformation therefore needs not only intelligent machines, but intelligent incentive structures.

The bottleneck does not disappear. It moves.

Technological progress does not eliminate scarcity. It relocates it.

When intelligence becomes cheap, energy becomes more important. When energy becomes cheap, materials become more important. When materials are available, production capacity becomes more important. Then infrastructure. Then regulation. Then trust. Then interests and power. And when that barrier falls too, the next one appears.

Progress does not remove bottlenecks. It moves them. Perhaps that is a better definition of progress: not a world without limits, but a civilisation that can move its limits ever faster.

And from that follows the economically most interesting consequence: as intelligence becomes less scarce, the relative value of everything that constrains its implementation rises. Every bottleneck creates a market.

If compute grows exponentially while grid infrastructure does not, the value of generation, grids and storage rises. If AI develops products in hours while factories need years to scale, the value of flexible, highly automated production rises. If autonomous systems work technically while regulation cannot follow, the value of certification, insurance and governance systems rises. And if technology moves faster than social trust, trust itself becomes an economic asset.

Perhaps the biggest trend of all: the automation of automation. Machines that do not only produce goods, but create the production capacity for more machines. Robots build factories, factories produce robots, robots build energy infrastructure. At that point the physical economy itself begins to scale faster. Perhaps that is the moment when the Physical World Gap first genuinely starts to close.

The real question

Today we focus on making AI more intelligent. But we may reach a remarkable point: our machines would be ready for a future our world is not ready for. Not for lack of intelligence, but because of energy, raw materials, infrastructure, laws, interests, trust and time.

Then AI would no longer be the bottleneck. Reality itself would be. That is precisely the Physical World Gap.

The decisive question of the coming decades is therefore not only: what can AI do? But: how fast can reality follow? The big winners of the next technological era may not be those who own the most intelligent AI. They may be those who solve the hardest problem: transferring the speed of the digital world into the physical one.

If you want to read on into how this question continues with ownership and distribution, you will find it in Beyond Money: Who Owns Abundance?.

Sources referenced in the text: S&P Global (mine development timelines), IEA (data centre electricity demand, global grid connection queues), Lawrence Berkeley National Laboratory (US interconnection queues), International Federation of Robotics (robot installations).