The question of whether artificial intelligence will destroy jobs has become almost quaint. Economists studying the arrival of artificial general intelligence are no longer asking about displacement. They are asking a far more revealing question: in a world where AI can produce nearly everything, what remains scarce enough to command value? A conversation between Alex Imas, director of AGI economics at Google DeepMind, and Phil Trammell of Epoch AI on the Dwarkesh Podcast mapped the forks in the road ahead, and the answers are anything but comfortable.
The Pattern Ricardo Missed Two Centuries Ago
Imas began the discussion with a deliberate caution about economic forecasting. In the early 1800s, David Ricardo warned that machinery would destroy employment and trigger social collapse. Every occupation he specifically named was eventually automated. Yet if Ricardo woke up today and saw prime-age employment in the United States near its highest level since the 2000 peak, he would be astonished. What he missed was structural adaptation: as automated goods became cheaper, the income they freed up flowed into new kinds of demand, and new kinds of employment emerged to meet it.
But Imas was careful not to assume the same loop would repeat. The point was not that history guarantees a happy ending. The point was that prediction is treacherous, and the real task is figuring out what data to collect and what scenarios to prepare for.
Relational Capital as the Last Refuge of Value
Imas anchored the discussion around what he called the relational sector. This is the part of the economy where the fact that a human is involved in the process is itself part of the value of the product or service. A ballerina’s performance, a barista pouring coffee, a physician delivering a diagnosis face to face. Even if AI can produce an identical output, consumers pay a premium for the knowledge that a person did it.
The experiments his team ran were revealing. Show a single art print and tell a subject it was made by a person or by AI, and the human-made piece commands a higher willingness to pay. But tell the same subject that five hundred identical prints were made, and the human-made piece collapses in value. The AI piece barely moves. The premium is not about the object. It is about the perceived connection to a specific artist. That connection evaporates the moment the work becomes scalable.
This is the core of the problem. For relational capital to sustain a large share of the economy, a broad range of occupations must elicit enough willingness to pay for the human element. The necessary data on how strong that willingness actually is does not yet exist.
Robots Multiply, Ballerinas Do Not
Trammell approached relational capital from a wider angle. Imagine AI handles every upstream supply chain for physical goods. Robots can be multiplied next year by any factor you want. Ballerinas cannot. The supply of the physical objects becomes nearly infinite, but the people performing the live, embodied services remain constrained.
If the marginal utility of physical goods declines fast enough — if people get full quickly — then surplus income flows to the relational sector, and labor’s share of the economy holds up. But if the variety of goods that robots can produce explodes so rapidly that consumers never reach satiation, the relational sector shrinks as a share of the whole economy.
Trammell reached for a historical analogy from the Mongol era. Fix the set of goods available at the time: horses for transport, yurts for housing, fermented mare’s milk for sustenance. Then ask what would happen if automation advanced within that fixed set. Horses and yurts would saturate, and all income would flow to the entertainer — the singer commanding the entire surplus. But that is not what happened. Horses, yurts, and ger diversified into forms no one predicted. The singer’s share remained trivial. Trammell argued that the same structural logic is likely to repeat.
The Less Comfortable Reading of Moore’s Law
Trammell offered one of the sharpest observations of the conversation. There is another way to read Moore’s Law. It is not that the density of transistors doubles every eighteen months. It is that the value of a unit of computation halves every eighteen months. Humanity has been finding uses for computation so aggressively that the rate of demand growth has kept pace with the rate of cost decline, maintaining the illusion that computing power is getting cheaper. What is actually happening is that we are burning through potential applications at a staggering speed.
AI may have broken this pattern for the first time. Trammell noted that NVIDIA’s H100 costs more per unit today than its predecessor did three years ago. Supply of compute has grown dramatically, but the models have become so much smarter that the opportunity cost of using them has risen. Prices did not fall during that period. This suggests that demand for compute is extremely elastic — the cheaper it gets, the more uses emerge. Unlike oil or steel, where demand eventually hits a ceiling, compute may be a good that keeps generating new applications as its price drops.
Imas reinforced the point. AI is creating new kinds of capital uses continually. Saturation does not set in the way it does for a commodity like insulators or gasoline. People want to use more compute for more things, and that structural hunger shows no sign of fading.
Why the AI Depression Scenario Is Hard to Build
In February 2026, Citrini Research published a report that shook markets. It described a scenario in which AI eliminates massive numbers of white-collar jobs, destroys consumer demand, and crashes the S&P 500 by 38 percent by 2028. Imas examined the scenario directly and concluded that the conditions required for wealth to generate negative economic growth are extreme. Capital owners would have to be fully satiated in both consumption and investment, with surplus funds finding no outlet at all. That is almost impossible to imagine in a world with AGI. No one builds a data center and then stops. No one develops a new model and then refuses to fund the next one.
Both economists agreed with Molly Kinder’s framing of the “messy middle” — the transitional period between the arrival of powerful AI and the stabilization of a new equilibrium. That period is the real danger zone. But Trammell saw the window as narrow. If AI is capable enough to replace most software engineers entirely, it should also be capable of replacing accountants and analysts. The cost savings and productivity gains from full replacement would be enormous, making it hard to sustain a prolonged depression scenario.
The Cost of Keeping a Human in the Loop
The conversation took a cold turn when it examined the cost of retaining human labor in production. Even if AI handles ninety percent of a task and humans handle the remaining ten percent, employment can be preserved as long as that ten percent adds enough value. Trammell predicted that once AI becomes sufficiently advanced, the cost of integrating a human into a production line will outweigh the benefit. AI systems will communicate in neuralese — exchanging numerical representations directly inside neural networks, thinking thousands of times faster than any person. A factory or design process built for AI speed and precision will treat a human worker as a bottleneck that degrades overall quality. The O-ring theory, where one weak link destroys the entire product, will begin to operate in the opposite direction: excluding humans will become the way to protect integrity.
The People Who Cannot Stop Accumulating
The most unsettling part of the discussion was about the people who never get full. Historically, the ultra-wealthy eventually consumed their fortunes, created foundations, or passed wealth to children who were less capable and dissipated it. That dissipation acted as a natural redistributive mechanism. But figures like Elon Musk talk about building Mars colonies and launching Starships. Accumulation has become the goal itself.
Trammell’s insight was cold. The only reason such people have not dominated the economy for centuries is that death and incompetence dispersed their assets across generations. If lifespans extend significantly, or if AI manages portfolios indefinitely, the dispersal shock disappears. Wealth concentration accelerates.
Then comes the von Neumann probe — a self-replicating spacecraft that builds copies of itself and spreads across star systems. For an entity like that, the marginal utility of compute never saturates. There is always another star to reach. If even a single agent with unlimited appetite for compute exists, its savings rate will outstrip every other actor, and over a long enough horizon it will own nearly everything. For the von Neumann probe, the concept of labor share ceases to have meaning.
What Developing Countries Can Do
Near the end of the conversation, the discussion turned to countries like India and Nigeria that are not building frontier AI models or fabricating advanced semiconductors. How do they access the wealth AGI generates?
Trammell’s answer was blunt. Buy the index. If AI transforms every industry, the benefits will spread across the entire economy, not just a few companies. Imas pointed to electric utilities as an analogy. Nobody thinks American power companies wield enormous political power, even though they supply an essential resource. The benefits of electricity flowed downstream to users. If AI follows the same pattern, owning the S&P 500 captures AGI’s fruits.
But there is a condition. If AI profits concentrate in a handful of private companies that never go public, index investing will not work. Nigeria does not own shares in SK Hynix or Anthropic. The question is whether AI becomes pervasive like electricity or platform-concentrated like social media. That fork in the road will determine whether the rest of the world can buy their way into the future.
Imas emphasized that the existence of competitive open-weight models is the decisive variable. If open models consistently trail frontier models by only six months, then effectively every country has access to AGI-grade intelligence. AI becomes more like electricity than like a proprietary platform. And the index becomes viable again.
The Real Question Is Not When but Which Fork
What emerges from this conversation is an honest admission from economists who know how often their profession has been wrong. For two hundred years, predictions about automation destroying employment have been wrong. But that does not mean they will always be wrong.
What can be said is what data would settle the question. The elasticity of willingness to pay for relational goods. The elasticity of demand for compute. The satiation speed of wealthy consumers. The performance gap between open and frontier models. These numbers do not exist yet. Collecting them is the urgent task.
What becomes scarce after AGI decides the fate of wealth is not a question anyone can answer today. But whether a country or an individual is prepared to ask it will determine which fork they end up taking.
