Automation, robotics, energy, and what happens when labour becomes a manufactured input. Leo Tervit.
Abstract. Artificial intelligence is usually discussed as software. I think its more important consequence will be physical. If machines become cheaper than people at enough tasks, labour is replaced by capital, and capital requires electricity, semiconductors, motors, transformers, factories and raw materials. Robotics extends that substitution into the physical economy; the strongest version is an industrial system in which machines perform much of the work required to expand machine capacity itself. Human labour then stops being the main limit on production. We are not there today: the ILO still expects transformation to be more common than outright replacement in AI-exposed work [11]. But the physical buildout is already visible. Data centres consumed about 485 TWh of electricity in 2025 and are projected by the IEA to approach 950 TWh by 2030 [1], while 4.664 million industrial robots were already operating worldwide in 2024 [3]. I think these are early pieces of the same process.
The economic purpose of artificial intelligence is automation. There are other uses for it, but this is the one that matters most for the structure of the economy. A business buys software, compute or machinery because it expects the same output with fewer people, more output with the same people, or some combination of the two. When the machine becomes cheaper than the worker, sentiment is not a durable defence of the job.
For now the substitution is easiest where work already lives on a computer. Code, research, drafting, customer support, analysis, design and administration can all be attempted without solving the harder problem of acting in the physical world. The ILO estimates that 25 percent of global employment sits in occupations with some generative-AI exposure, but only 3.3 percent is in its highest exposure category, and it expects transformation to be more common than outright redundancy [11]. The IMF uses a broader measure and puts close to 40 percent of global employment in AI-exposed occupations [5]. Exposure is not unemployment. It is where substitution can begin.
I do not think most jobs disappear next year, or in one clean wave. I do think the usual assumption that displaced labour will always be absorbed into new human work deserves less confidence than it is given. Previous machines were usually narrow. A loom replaced weaving. An excavator replaced part of digging. A spreadsheet replaced arithmetic. A sufficiently capable AI system is different because the same underlying intelligence can be pointed at many kinds of work. Generality makes automation portable.
The same logic becomes more important once intelligence leaves the computer. A model that writes software changes the cost of software. A model that can see, plan and control machinery changes the cost of labour itself. That is the step I care about.
From there the argument is straightforward. AI makes some cognition cheap. Robotics turns cheap cognition into machine action. Machine action replaces human hours with capital. Capital requires electricity, hardware and materials. If machines eventually perform a large share of the work required to expand that capital, the process becomes recursive. Labour is no longer the scarce input it used to be. Energy and matter are.
That is the bet behind this paper.
Economically, a job is not protected because it is familiar, skilled or socially respected. It is protected while the human remains the cheapest acceptable way to produce the required result.
A firm compares two bundles. The human bundle contains wages, benefits, hiring, training, management, absence, turnover and error. The machine bundle contains capital expenditure, financing, depreciation, electricity, compute, maintenance, integration and failure. Automation happens when the second bundle becomes cheaper after quality and risk are taken into account.
This is why a machine does not need to be better than a person at everything. It needs to be good enough at the particular task. A mediocre autonomous system that works twenty-four hours a day can be more valuable than a highly capable human if the task is repetitive and the cost difference is large enough. The threshold will be different in every industry and every country, but the comparison is the same.
There is already evidence that automation can weaken labour even while raising productivity. Acemoglu and Restrepo found that greater industrial-robot exposure in U.S. local labour markets reduced employment and wages over the period they studied [6]. That work predates modern generative AI, so it cannot tell us what present systems will do. It matters because it disposes of the comforting idea that productivity gains must immediately reappear as better jobs for the same people.
Human intelligence is expensive to reproduce. It takes years to educate a person. That person can only work so many hours, may leave, may refuse the task, and must be paid again next month. Machine intelligence has enormous fixed costs, but once a capability exists it can be copied into another process at a much lower marginal cost.
That difference matters more than whether a model is conscious, creative or genuinely intelligent in the philosophical sense. Employers do not need to settle those questions. They need output.
The first labour market for AI is software because software has almost no physical interface problem. A programmer produces code. A lawyer produces research and text. An analyst produces a model or recommendation. A designer produces an image or specification. The machine can already reach the input and output.
As reliability improves, the boundary moves. Some occupations will shrink. Others will keep the same title while one person does what several people did before. Some will expand because lower costs create more demand. The composition will differ by industry. I am less interested in the exact occupational sequence than in the direction: more useful work becomes something that can be purchased as capital rather than hired as labour.
I doubt the final human jobs will simply be the ones AI cannot do. Many may survive because people insist that a person does them.
Politics is the obvious case. A machine may eventually be able to analyse policy better than a minister, but analysis is not the whole job. Political authority has to be embodied in somebody the public can remove, blame, persuade or distrust. The same applies to judges, military command, senior legal functions and parts of regulated finance. Responsibility is itself a function.
There is another class of work where the product is partly the relationship. High-end asset management, advisory work, negotiation, private banking, sales and certain front-office roles may become heavily machine-assisted without becoming fully machine-fronted. A client can prefer a person even when the machine does most of the underlying analysis.
And there will be work that survives because human presence becomes a luxury. A human bartender, barista, concierge, craftsman or adviser can become more valuable precisely because machine service is cheap. We already pay premiums for handmade goods that machines can produce more consistently. Human time can acquire the same status.
This is an important distinction. Human employment can survive without human technical superiority.
AI becomes much more economically important when it can act on the physical world.
Industrial robotics is not new. In 2024, 542,000 industrial robots were installed worldwide and the operational stock reached 4.664 million units [3]. Most of these machines are narrow. They weld, pick, place, paint, cut or move in environments designed around them. Their advantage is repetition, not general intelligence.
The hard part of general physical work is variation. Real workplaces contain damaged parts, unusual objects, incomplete instructions, people walking through the workspace and events that were not anticipated by the programmer. Traditional automation avoids this by controlling the environment. More capable AI attacks the opposite side of the problem: make the machine better at dealing with the environment as it is.
The International Federation of Robotics now describes the shift from rule-based automation toward systems using analytical, generative and agentic AI to operate more autonomously in complex environments [4]. That does not mean a general robot worker exists today. It means the bottleneck is moving.
Humanoid robots attract attention because the world was built for human bodies. Doors, stairs, shelves, tools and workstations all assume our dimensions. A machine that can use the same environment without redesigning it could be useful.
But I do not care whether the winning machine has two arms and a face.
An autonomous haul truck is a robot. A warehouse system is a robot. A mining drill that positions itself is a robot. So is a field machine, an automated crane, a machine-vision sorting line or a construction system. In many industries the best robot will look nothing like a person because there is no reason to inherit the limits of the human body.
The important threshold is economic, not aesthetic: when can machinery perform enough varied physical work that employing a person becomes the expensive option?
Software automation competes with the wage bill attached to digital work. Robotics competes with the wage bill attached to the physical economy.
That includes manufacturing, warehousing, transport, mining, agriculture, construction, cleaning, maintenance, food preparation and a long list of jobs that are currently protected by the difficulty of physical reality rather than by the difficulty of the underlying decision.
If robotics works at scale, the demand created by AI stops being mostly demand for GPUs in data centres. It becomes demand for the machinery that replaces labour everywhere else. Motors. Actuators. Bearings. Sensors. Cameras. Power electronics. Batteries. Machine tools. Factories. Grid connections. More compute. More electricity.
This is where the AI story stops being mainly a software story.
When I describe machines becoming self-generating, I do not mean a humanoid robot walks into a factory with ore and emerges with another humanoid robot.
Human industry does not work that way either. Mines extract ore. Refineries process it. Truck factories build trucks. Semiconductor fabs build chips. Power stations produce electricity. Construction firms build plants. Logistics networks connect the pieces. No single worker or company contains the whole system.
The relevant question is how much human labour remains inside that network.
Imagine autonomous mining equipment extracting ore. Autonomous vehicles move it. Automated plants refine it. Machine-operated factories make motors, bearings, structural parts and electronics. Robotic systems assemble equipment. AI schedules production and maintenance. Automated construction systems extend factories, roads and power infrastructure. Humans still exist in the chain, but their share of the work falls.
At some point the machine system begins to contribute materially to the production of more machine capacity. That is the threshold that matters. I will call it recursive industrial automation.
It does not require perfection. If building and operating the next factory takes one tenth of the human hours required by the last one, the economy can add productive capacity much faster with the same labour force. If the machines produced by that factory then help build the next factory, the process feeds itself.
| Stage | What machines do | What becomes scarce |
|---|---|---|
| 1. Digital automation | Information and cognitive tasks | Compute, chips, data-centre power |
| 2. Physical automation | Increasingly varied physical tasks | Robotics hardware, integration, electricity |
| 3. Autonomous industry | Extraction, logistics, factories, maintenance, construction | Grid capacity, industrial equipment, materials, land |
| 4. Recursive expansion | A large share of the work needed to add more machine capacity | Energy, raw materials, build time, law and politics |
We are clearly not at Stage 4. Most deployed robotics remains narrow, and present AI systems still need significant supervision. The table is not a forecast calendar. It is the path implied by taking automation seriously as an economic process.
If progress stalls at Stage 1, the data-centre boom can still be enormous. If it reaches Stages 2 and 3, the physical economy starts reorganising around machine labour. Stage 4 is where the traditional relationship between labour and production finally breaks.
The energy thesis does not depend on mass-market robots. AI already consumes enough electricity to matter to grid planning.
The IEA estimates that data centres consumed about 485 TWh of electricity in 2025 and projects roughly 950 TWh by 2030, around 3 percent of global electricity demand. Electricity consumption at AI-focused data centres grew by about 50 percent in 2025 and is projected to triple between 2025 and 2030 [1]. In the United States, Lawrence Berkeley National Laboratory estimates in its 2026 update that data centres could account for 11.8 percent of electricity consumption in 2030 in its central estimate, with a scenario range of 9.5 to 15.3 percent [2].
Those numbers matter, but the local problem is more important than the global percentage. Data centres are concentrated loads. A country can have enough electricity over the year and still lack the substation, transformer or transmission capacity to connect a new cluster where it is wanted. The IEA reports that the power density of AI servers increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027 [1].
That turns demand for intelligence into demand for electrical equipment.
Labour is not replaced by nothing. It is replaced by capital that has to be powered.
A robot hour requires electricity at the machine. The control system requires compute. The compute requires servers, networking and cooling. The factory around the machine requires power. The parts inside it had to be mined, processed and manufactured. Moving work from people into machinery changes where the economy consumes energy.
A simple way to express the direction is:
$$\text{human labour} \longrightarrow \text{capital} + \text{compute} + \text{electricity} + \text{materials}.$$
This is not an accounting identity. Human beings also consume energy indirectly through food, transport and housing. The point is that machine labour makes industrial energy a more direct input into production.
If a few machines replace a few workers, the effect is small. If autonomous systems replace billions of hours of work and allow mines, factories and logistics networks to operate at higher utilisation, the effect is structural.
AI hardware and models will become more efficient. That part is almost certain. The IEA estimates that electricity required for individual AI tasks has been falling extremely quickly in recent years [1].
But lower cost changes behaviour. When inference gets cheaper, firms run more inference. When machine vision gets cheaper, more objects are inspected. When robotic labour gets cheaper, tasks that were never worth automating become worth automating. The quantity of machine work is not fixed.
This is already visible in data centres: energy per task is falling while total electricity use is rising because usage is expanding faster [1]. I expect the same logic to matter in robotics if deployment becomes cheap enough.
The energy argument therefore does not require machines to be inefficient. It requires demand for machine work to expand faster than efficiency reduces the cost of each unit.
I do not think this thesis requires pretending to know the exact future generation mix.
Nuclear may become more valuable because AI and industry want reliable power. Natural gas can provide dispatchable generation where pipelines already exist. Solar and wind can add capacity quickly where land, transmission and storage cooperate. Geothermal may matter in suitable locations. Fusion would change the picture if it becomes commercially practical, but no serious near-term thesis should depend on it.
The common requirement is usable electricity.
That is why the grid may matter as much as generation. The IEA estimates that meeting announced national energy and climate goals already requires adding or refurbishing more than 80 million kilometres of grid by 2040, roughly the length of the existing global grid [7]. AI arrives on top of that. The U.S. Department of Energy reports that distribution-transformer lead times that were three to six months in 2019 had extended to roughly one to two years or longer by 2024, with some large transformers taking three to four years [8].
Code can be copied instantly. A transformer cannot.
The phrase artificial intelligence hides an industrial supply chain.
Models require accelerators, memory, networking, storage and advanced packaging. Those require fabs, fabrication equipment, chemicals, gases and substrates. Data centres require racks, cooling, switchgear, power conversion, backup systems and buildings. Buildings require concrete, steel, land and grid connections.
Robotics adds another bill of materials: motors, actuators, gears, bearings, encoders, cameras, force sensors, power electronics, wiring, compute, batteries and structural materials. Purpose-built industrial machines add their own specialised components.
The more successful the software becomes, the more of this physical layer has to exist underneath it.
Software can be duplicated almost for free. Matter cannot.
An AI system can generate another factory design in seconds. It cannot generate another tonne of copper in seconds. It can optimise a grid. It cannot copy a transformer. It can design a better motor. The motor still has to be made.
This difference becomes more important as intelligence gets cheaper. Once design, planning and coordination are abundant, implementation becomes the visible bottleneck. The machine can propose a million factories. The economy still has to build them.
Copper is the obvious example because grids, motors and data centres all use it, but the thesis is not about one metal. The IEA's 2026 Critical Minerals Outlook expects strong growth in demand for key energy minerals through 2040 and projects copper to have the largest absolute increase, about seven million tonnes. Even after announced projects, it still shows a material gap between expected mined supply and 2035 requirements [9].
Modern digital and automated systems also depend on aluminium, steel, silicon, silver, tin, tungsten, gallium, germanium, tantalum, platinum-group metals and rare earths in specialised applications [9]. Engineering will substitute away from expensive inputs where it can. That is how markets respond. It does not abolish the requirement for matter.
A resource argument becomes nonsense if it assumes supply is fixed. Higher prices produce new mines, recycling, substitution, efficiency and new technologies. Automation itself may lower the cost of extraction.
The problem is time.
Software can gain millions of users in months. Mines, transmission lines, power stations, processing plants and heavy factories can take years. Geology does not move faster because the model got better. Permitting does not automatically move faster either.
This mismatch is one of the main reasons I am interested in the physical side of AI. Intelligence can create demand faster than the physical economy can answer it.
If recursive automation ever becomes powerful enough, continued growth means directing more energy through more matter. Land, mineral deposits, environmental tolerance and energy flows become the limits at the margin.
This is also where automation connects to space. Off-world industry is extremely expensive when every productive site needs people, life support and constant human supervision. It becomes less absurd when machines can extract, build and maintain far from human settlements. Automation does not make asteroid mining or lunar industry inevitable. It removes one of the largest costs.
If industrial activity ever expands materially beyond Earth, I expect autonomous machinery to arrive before large human workforces do.
The strongest limit on automation may eventually be people refusing to permit it.
A machine can become capable of doing a job before the law allows it to hold the responsibility attached to that job. The EU AI Act already requires high-risk AI systems to be designed for effective human oversight [10]. The rules will change, but the principle is unlikely to disappear in every important domain. Societies care about who can be blamed, sued, removed, imprisoned, voted out or held to a professional duty.
That is why law, medicine, finance, critical infrastructure, defence and government may retain humans at the point of final authority long after machines dominate the underlying analysis.
The last human worker in a process may be there to carry liability.
There is also a simpler constraint: humans like other humans.
If machine service becomes cheap and competent, human service can become a premium good. Restaurants may employ people because customers prefer them. Hotels may retain human hosts. Wealthy clients may still want human advisers. Education and care may preserve human contact even when machines can perform much of the technical work.
A barista is an inefficient coffee machine. That does not mean the barista disappears if the inefficiency is part of what the customer values.
This is why I expect some front-facing and luxury work to survive longer than its technical necessity would suggest.
Automation creates concentrated losers. The owner of the machine receives a stream of savings. The displaced worker can lose an income immediately. A technology can raise national output and still be opposed rationally by the people who bear the cost.
Governments can slow deployment through labour law, licensing, safety rules, liability, tax treatment and outright restrictions. They can also accelerate it through infrastructure, procurement and legal clarity. The deployment curve therefore does not have to resemble the capability curve.
Software improves at software speed. Societies do not.
A world in which people still work for wages is easy to understand. A world in which machines produce almost everything can at least be imagined as abundant. The unstable period is between them.
In that period, automation may remove bargaining power from labour before it makes necessities cheap enough to compensate. Productivity can rise while household security falls. Asset owners can become richer while wage growth weakens. The economy can become better at producing things while becoming worse at distributing claims on those things.
This happens because wages do two jobs. They pay for labour and they distribute purchasing power. A company can benefit from removing the wage cost, but the economy still needs somebody to buy the output.
A robot does not buy a house because it received a salary.
There are many ways around this problem: lower prices, wider capital ownership, public provision, transfers, taxation, sovereign investment funds, new forms of work. My point is only that some mechanism has to replace the distributive role of wages if wages cease to do it.
The darker equilibrium is not necessarily communism. It can be a highly productive private economy with a population increasingly dependent on transfers.
Imagine that the compute, robots, energy infrastructure, land and intellectual property remain concentrated. Labour income becomes less important. Most people still need housing, food, transport, healthcare and entertainment, so the state transfers purchasing power back to them through income payments, subsidised essentials, public housing or other provision.
Such a society can be materially rich and politically regressive at the same time. A person may have access to cheap food, transport and entertainment while owning almost nothing that produces income. Dependence can coexist with abundance.
The important division is no longer worker and employer. It is owner and non-owner.
The optimistic version is also real.
If autonomous production pushes the cost of goods and services low enough, and if ownership or access to productive capital is sufficiently broad, people may simply need to sell much less of their time to live well.
That would be a profound break with normal life. A conventional full-time career can absorb roughly ninety thousand hours over several decades before commuting is counted. Scarcity made that exchange necessary. It is not obvious that it remains so once machines can perform most economically useful work.
Travel, education, transport, manufacturing, construction and many services could become much cheaper when labour is no longer the dominant cost. People would still compete for scarce land, status, attention and unique experiences. They would still work voluntarily. But work would no longer have to be the admission price for an ordinary life.
I do not know which version wins. Technology does not decide that question by itself. Ownership does.
In a labour economy, a person with little capital can still obtain a claim on output by selling time. In a heavily automated economy that route weakens.
That makes ownership more important, not less.
If productive capital is broadly owned, automation can turn into leisure. If it is narrowly owned, automation can turn into dependency. The engineering can be identical in both cases.
This is why I do not think the final political argument about AI will be mainly about whether the models are intelligent. It will be about who owns the productive system built around them.
The cleanest way for this thesis to fail is for AI capability to plateau before it becomes a reliable substitute for much work. Present systems are still brittle. Long-horizon agents fail. Models hallucinate. Integration is expensive. If those problems remain fundamental, AI stays a useful tool rather than a replacement for labour.
Robotics could also remain the wall. The physical world punishes errors more severely than text. Hardware breaks. Edge cases are endless. Dexterity is hard. It is entirely possible that AI automates a great deal of office work while general physical labour remains stubbornly human for decades.
Relative prices could work against automation. In countries where wages are low, capital is expensive and electricity is unreliable, a person may remain cheaper than a machine even when the machine is technically capable. Automation will not spread evenly.
The energy thesis could be weaker than I expect if efficiency improves faster than usage expands. Small models may handle most workloads. Robotics may require less energy than expected. Compute may become radically more efficient. The current data-centre evidence points toward rising total consumption despite rapid efficiency gains [1], but that relationship is not a law of nature.
Politics may also stop the process. Governments can preserve human employment by law, restrict autonomous systems, tax machine substitution or require humans to remain responsible. That may be inefficient in a narrow economic sense and still be politically stable.
Even if the technological thesis is right, the investment can be wrong at a given price. Shortages attract capital; power equipment can be overbuilt, semiconductor capacity can become a glut, and commodity prices can fall after new supply arrives. Railways changed the world and still ruined investors. A correct view of the future is not the same as a good entry price.
These are not footnotes to the argument. They are the reasons the future described here may arrive slowly, incompletely or not at all.
Artificial intelligence matters when it changes the cost of work.
The first stage is cognitive. The second is physical. If machine intelligence becomes reliable enough to control machinery across varied environments, then labour begins to move from something firms hire to something they manufacture. If automated industry then provides a growing share of the work required to expand automated industry, the traditional labour constraint weakens further.
The economy does not escape scarcity when this happens. Scarcity moves.
It moves into electricity, compute, grid connections, motors, factories, minerals, land and the time required to build physical things. It also moves into law and politics, because humans can refuse to delegate authority even when delegation is technically possible.
The near-term evidence is already physical. Data-centre electricity demand is rising quickly [1, 2]. Millions of industrial robots are already in operation [3]. Grid equipment has long lead times [8]. Critical-mineral supply cannot be expanded at software speed [9]. None of this proves that human labour becomes obsolete. It shows what the economy has to build if the substitution continues.
My investment view follows from that. I am less interested in predicting which interface, model or robot brand dominates than in the physical requirements common to many successful versions of the future. More machine work means more capital. More capital means more hardware. More hardware means more energy and materials.
The social question is harder. If labour stops being necessary before ownership becomes broad, the result can be dependency on an extraordinarily productive system controlled by relatively few owners. If the cost of production falls far enough and access to capital broadens, the same system can remove a large amount of compulsory work from human life.
Both outcomes begin with the same technological event: machines become cheaper than people at more and more of the things the economy pays people to do.
Artificial intelligence may begin as software. Its future is industrial.