The Abundance Paradox

A developer argues that the transition to an AI-driven post-scarcity economy will be far messier and more unequal than either utopian or dystopian narratives suggest, drawing on Iain M. Banks' Culture novels for the destination and Asimov's robot stories for the transition. The essay contends that as intelligence becomes cheap, demand and scarcity rents shift to the physical layer of AI — compute, electricity, datacenters, transformers, and robotics supply chains — making those inputs the key investment focus.

The abundance paradox I’ve been reading Iain M. Banks recently and I keep coming back to the economics of the Culture rather than the spaceships. The Culture is basically post-scarcity. There is no money, nobody needs a job, machines can make almost anything people need and the artificial intelligences running the civilization are vastly more capable than humans. Banks once described work there as something closer to play because nobody has to do it in order to survive. This sounds like science fiction because, obviously, it is. But the basic economic question no longer feels completely detached from reality. If AI keeps improving and robotics eventually makes machine intelligence useful in the physical world, we may end up in a situation where a very large part of human labor is no longer economically necessary. People usually jump from that idea to one of two conclusions. Either it becomes a utopia where everything is cheap and nobody needs to work, or it becomes a dystopia where a small number of companies own all the machines and everyone else is unemployed. I suspect both views are too clean. The long-term destination could be surprisingly abundant while the transition is extremely messy, and possibly much more unequal than today. This is also where I think the investment implications become interesting. Banks has the destination. Asimov may have the transition. Banks imagined a society that had already solved the difficult part. Technology is so advanced that ownership of productive assets no longer matters very much because almost anything reproducible is effectively free. There is little reason to accumulate money if you can already have almost anything you want. Asimov’s robot stories often feel closer to the world we are likely to experience first. Robots exist, they are useful, productivity rises, but people still care about jobs, income, social status and ownership. Some societies benefit much more from automation than others and the technology creates political and social tension before it creates abundance. That seems more plausible to me as a transition path. We already live in a strange version of this. Information has become almost absurdly cheap. A person with a phone can access more books, maps, music, video, software and scientific information than almost anyone in history. I can talk to an AI that can explain quantum mechanics, help me write code, translate Russian or discuss a company filing for effectively nothing. At the same time, an apartment in central London is still very expensive. Technology made one category abundant and increased demand for other scarce things. This is an old pattern, but AI may push it much further. If intelligence becomes dramatically cheaper, the value of the things required to turn that intelligence into useful economic output can rise. We are already seeing this with compute and electricity. Better models create more demand for GPUs, networking, datacenters and power. Those datacenters then require transformers, cables, turbines, substations, cooling equipment and permitted land. The same thing can happen with robotics. A humanoid robot may eventually reduce the cost of labor, but producing millions of robots would increase demand for motors, bearings, magnets, rare earths, batteries, semiconductors and factories. So an economy moving toward abundance can still create enormous scarcity rents during the transition. In fact, the abundance may be the reason those rents appear. This is one of the reasons I am much more interested in the physical layer of AI than I used to be. It is easy to imagine software getting cheaper. It is much harder to quickly manufacture another gigawatt of reliable power or another million transformers. What happens when labor becomes cheap? Imagine that, at some point, a useful general-purpose humanoid robot costs $20,000. I don’t mean a robot that can walk around a stage and wave. I mean a machine that can actually perform a broad range of economically useful physical work with reasonable reliability. The economics would change quickly. A human worker may cost an employer $50,000 or $80,000 a year once wages, benefits, taxes and other costs are included. A robot is capital equipment. It can potentially work several shifts, it does not need sleep and its software can improve after it has already been purchased. If that happens, the cost of physical labor should fall substantially. The more important effect comes later, when automation starts appearing throughout the supply chain itself. Robots can help manufacture other robots, autonomous equipment can operate mines, logistics can become increasingly automated and engineers can use AI to design better factories and machinery. Cost reductions then begin reinforcing one another rather than occurring in isolation. This is the route by which physical goods can become much cheaper. Sam Altman has talked about this kind of future for years. The broad idea is that if intelligence is cheap and machines can increasingly perform physical work, then the human labor component embedded in many goods falls toward zero. I agree with the direction, although the timing matters enormously because physical bottlenecks may persist for years before those cost declines reach consumers. Building a robot is harder than generating a paragraph. Building a factory is harder than deploying an AI agent. Expanding the electrical grid can take years. Mining and processing raw materials takes time. Regulation, land and local politics can slow everything down further. There could therefore be a substantial period where digital labor has already become cheap but physical life remains expensive. That period looks economically very different from the Culture. You can make the house cheaper. You cannot make another Malibu. There is another reason I don’t think post-scarcity means all prices fall. Some things are genuinely scarce. Suppose robotics and automation reduce the cost of constructing a high-quality house by 80%. That would be an extraordinary achievement, but it would not reduce the supply of Malibu coastline, historic neighborhoods in Paris or land next to Central Park. The structure may become cheaper while the land becomes more expensive. This can happen precisely because society becomes richer. If more people can afford beautiful houses but only a fixed number can live in the most desirable places, the value of those places rises. The same logic applies outside property. AI can create unlimited entertainment, but it cannot create another original Picasso. It can produce millions of songs, but it cannot create more seats for a Champions League final. It can design a better city, but it cannot create another Venice. I think something similar will happen inside the AI economy itself. If machine intelligence becomes abundant, scarce access to power, grid interconnections, semiconductor capacity and certain types of land may become much more valuable. The cheaper the intelligence becomes, the more economic activity we may try to build around it. That is why I don’t see AI abundance as an argument against investing in scarcity. In the medium term it may be exactly the opposite. This is also why I am skeptical when people say that AI will eventually make everything so cheap that current infrastructure spending must be a bubble. The eventual destination tells us almost nothing about how much infrastructure needs to be built to get there. If anything, a world of extremely cheap machine intelligence probably requires an enormous amount of expensive physical infrastructure first. The uncomfortable part is ownership There is a more important problem than electricity or copper. Suppose a factory employs 5,000 people today and produces €1 billion of goods every year. Twenty years from now it employs 200 people, uses thousands of robots and produces €10 billion of goods. Society has clearly become more productive. More goods exist and less human labor is required to make them. But who owns the factory? If the factory, robots, AI systems and energy infrastructure are owned by a small group of people, there is no automatic mechanism that gives everyone else an equal share of the productivity gains. This is where the simple abundance story becomes difficult. A person can become economically less valuable before the things they need become dramatically cheaper. An accountant could lose much of the value of his work while rent, healthcare and food are still priced according to the old economy. The eventual productivity gains may be enormous, but that doesn’t help very much during the five or ten years in between. This is why I think the transition could be extremely K-shaped. The owner of a successful AI company may become vastly richer at the same time that many white-collar workers see their bargaining power decline. The owner of a datacenter or power plant may benefit from the buildout while ordinary consumers initially experience higher electricity costs. Countries that control compute, capital and energy could gain relative to countries whose competitive advantage is cheap human labor. All of these outcomes are compatible with AI being a huge success. In fact, some of them may happen because AI is successful. The usual argument about whether AI is “good for the economy” therefore misses a lot. Aggregate GDP can rise while the distribution of that GDP becomes much more uneven. Productivity can improve while some workers become poorer. Consumer goods can become cheaper while land and financial assets become more expensive. The direction of aggregate wealth and the distribution of that wealth are separate questions. Banks’s Culture solves this by making ownership almost irrelevant. Once machines can provide almost anything, there is little economic power in owning the means of production because production itself is no longer scarce. A form of technological socialism or as popularly said Fully Automated Luxury Communism. We are nowhere near that condition. Before ownership becomes irrelevant, ownership may become extremely important. The road to abundance may be very capital intensive There is something slightly funny about the current AI debate. People who believe in abundance often imagine a future where goods are cheap, labor is cheap and intelligence is cheap. People who are skeptical look at today’s enormous capital expenditure and conclude that the spending cannot possibly make sense. Both can be true at different points in time. To get extremely cheap intelligence, we may first need trillions of dollars of datacenters, chips, power plants, transmission infrastructure and factories. To get cheap machine labor, we may need to manufacture hundreds of millions of robots and completely rebuild large parts of the industrial supply chain. There is no contradiction between very high capital expenditure today and very cheap output later. Railways were expensive to build and cheapened transport. Semiconductor fabs are extraordinarily expensive and have made computation vastly cheaper. The infrastructure required to create abundance can itself be scarce and expensive while it is being built. This is where I think a lot of the investment opportunity sits. The most obvious AI trade is to own the companies making the intelligence. That may work very well. But if intelligence becomes progressively cheaper, some of the largest profits may migrate toward whatever remains constrained. Today that might be advanced GPUs and HBM. Tomorrow it could be networking or power. Later it could move toward transformers, gas turbines, nuclear generation, copper, grid connections, industrial robots or even land around large energy sources. The bottleneck can move. I don’t think there will be one permanent winner collecting all the rents. The economic value is likely to move through the stack as each constraint is relieved and the next one becomes important. This is basically the scarcity-stack thesis we have been building in the portfolio. The reason I like it is that it does not require me to know exactly which AI lab eventually dominates. If AI demand continues growing, somebody has to build the physical system underneath it. Where I think this leads I don’t know whether we ever reach anything resembling the Culture. Energy constraints could persist for decades. Governments could slow deployment. The cost of physical systems may decline much more slowly than the cost of software. It is also possible that AI works extremely well but the ownership structure prevents broad abundance from arriving for a very long time. Still, the direction is difficult for me to ignore. We are already seeing intelligence become cheaper. AI systems are moving from answering questions toward performing actual work. Companies are spending extraordinary amounts of money because they want more compute than currently exists. Robotics is improving rapidly enough that serious companies are now building factories around the assumption that general-purpose robots will eventually be useful. If these trends continue, the economic importance of human labor should decline. At that point, the old question of “what will everyone do?” becomes less interesting than “who owns the productive system while this is happening?” I am actually not very worried about humans finding things to do. People already spend enormous amounts of time on activities with little economic value. We play games, dance, paint, travel, compete in sports, collect things, gossip, read novels, argue online and learn languages nobody is paying us to learn. This is discussed in great length in the book Deep Utopia but Nick Bosstrom Humans are perfectly capable of inventing reasons to get out of bed. The difficult part is getting to a world where you don’t need an employer’s permission to afford the bed. Banks may ultimately be right that sufficiently advanced technology makes the distinction between work and leisure much less important. I can imagine a future where machine intelligence and machine labor produce enough abundance that economic survival stops being the organizing principle of human life. But before we get anywhere near that, we may go through a period where intelligence is cheap, labor is becoming cheap, physical bottlenecks are still expensive and the ownership of productive assets matters more than ever. That period could last a long time. For an investor, it may also be the most important part. K-Shaped is AI-native: the ideas, investment judgments and portfolio decisions are mine; I use AI extensively for research, stress-testing, organization and drafting. I review and take responsibility for everything published.