The better AI gets, the smaller its share of the economy might get – Alex Imas and Phil Trammell
Is it worth your time?
You get a detailed account of how economic models of capital share, demand elasticity, and wealth accumulation apply to an automated future.
This conversation examines how advanced automation and artificial intelligence will alter the distribution of wealth, labor shares, and economic demand. The speakers analyze historical parallels from the Industrial Revolution to evaluate whether automation leads to widespread unemployment or unexpected economic growth.
Capital share and automation. Imas expects that as AI automates production, the overall implications for the capital share of the economy remain ambiguous because human-intrinsic relational services will coexist with automated goods. He notes that roughly 60% of the economy has gone to humans in wages for hundreds of years, a Kaldor fact that survived the Industrial Revolution because structural change created new jobs.
Demand elasticity for software. Trammell argues that demand collapse is unlikely because cheaper software and compute drive higher total spending when demand is sufficiently elastic. He points out that the Budget Lab at Yale found no evidence of a white-collar bloodbath in exposed sectors like software engineering.
Wealth accumulation by agents. Imas warns that future autonomous entities and wealthy humans with non-saturating preferences for accumulation will drive savings rates up until they hold most of the wealth. He explains that the wealthiest humans today prefer compounding wealth over immediate consumption, much like the high returns currently seen in data centers.
Will capital share increase?
Consider the relational sector, defined as services and goods where a human in the loop is part of the value. Because humans are naturally scarce, if automation makes other things abundant, we will still have scarcity in things humans are involved in. In a world where AI can physically do anything humans can, there is a machine economy building factories, doing research, and generating ideas. Humans probably won't be involved in that physical production in the ultimate limit if robotics is solved.
However, we actually want the ballerina or the barista to be human; that is part of the value of a performance or a cafe. Only humans have that preference, creating a human economy where humans provide services for each other. While some wealth flows between humans, some flows out because they want automated goods from the machine-only economy. Much of the machine-only economy is a closed loop because machines don't care about a human barista making them coffee.
Individual forecasts by economists like us are not necessarily useful. A blog post by Andrey Fradkin, Brian Jabarian, and Andrew Koh looking at labor market forecasts found disagreement in every direction. I agree with their advocacy that we should generate prediction markets to get aggregate forecasts and wisdom-of-the-crowd effects.
David Ricardo, a classical economist, wrote during the Industrial Revolution that while lower prices would be great, the automation of value-creating jobs would lead to unemployment and political unrest. Ricardo's predictions were right; those jobs were automated. Yet, if he woke up and learned this, he would be surprised to find the prime-age employment rate in 2026 is the highest it's ever been other than 2000. We have the highest number of potentially employed people since 2000.
Ricardo missed the economics of structural change: everything automated became cheap, leaving people with more money to spend on services. This is the lump-of-labor fallacy; he didn't consider that new jobs would be created. Labor share has gone down, and Phil will talk about this later today.
I've been saying we need a Manhattan Project for data. We lack data on consumer demand elasticities, and the O*NET database of tasks and jobs is rarely updated and super low quality. If there's full employment, we can talk about the relational sector.
The total economy, the sum of goods and services sold, is paid out either to people in wages or to capital, such as rent on buildings and payments to shareholders. For hundreds of years, roughly 60% of the economy has gone to humans in wages, while 30-40% goes to those who own machines, land, and claims on companies. This is a Kaldor fact. It is surprising that labor share remained over 60% after the Industrial Revolution and all subsequent automation. While some say labor share has fallen in the last 20 to 30 years, accounting changes complicate this; Atkinson has a paper showing that if accounting remains constant, labor share hasn't fallen.
If labor and capital are complements, you need both to do anything. If you look at the network-adjusted factor shares of a good down the supply chain—not just the final step, but what went into the machines that automate that step—you'll find labor adds significant value. In the US, computer and electronic products have a stable network-adjusted capital share of around 50%.
I believe a qualitative shift is coming where some goods' network-adjusted capital share goes to one because the entire supply chain can be automated without any part we intrinsically care having a human do. The implications for the overall capital share are ambiguous. Imagine two sectors: the human-intrinsic sector, like ballerinas, and everything else. If we fully automate everything else and satiate quickly, the quantity of non-ballerina goods goes to infinity, but marginal utility drops faster than quantity rises.
In the task-based model of jobs, if consumers pay more for a service where every task is automated except for a doctor delivering a diagnosis and providing support, that job is part of the relational sector. Using conjoint analysis of willingness to pay, you might find an increase in capital variety that avoids satiation. If this variety increases fast enough and doesn't happen in the human sector, you can get all the relational goods you want, but labor share still goes to zero. This reminds me of your analogy to a Mongolian economist in 1400 thinking about scarcity and the limits of that analysis.
Consider the goods available to a Mongolian of the distant past. There were jobs that were intrinsically human, like being a singer, and things that were not, like the transportation services provided by horses or different kinds of food. If they held the varieties fixed in both categories and asked what would happen with more automation, they might have said they would satiate in horse-like transportation, yogurt, and yurts. Those shares would go to zero, leaving them to spend all their money on singers.
As we accumulated more wealth and advanced machines, we expanded the range of things other than singers to spend our money on, and the share spent on singers stayed negligible. It is hard to imagine a world with trillions of robots and only some billion-odd humans where the cumulative amount spent on robots is less than what we spend to pay Magnus Carlsen.
The number of transistors in the world has trillion-X’d or maybe quadrillion-X’d. Chad Jones has an interesting result showing that the share of the economy going toward paying for computing and transistors has been decreasing. Not only are we producing more transistors more cheaply, but the value of the marginal transistor is decreasing. I like the pessimistic framing of Moore's law: every 18 months, the value of computation halves because we are running out of uses for computation so fast that it sustains Moore's law.
This is relevant to AI where, for the first time, this may no longer be true. An H100 costs more to rent now than it did three years ago, despite superior technology and more compute in the world, because as models get smarter, the opportunity cost of compute gets higher. We have increased the types of things people demand from capital; as long as that stays the case, the share of the economy going toward compute will keep increasing.
Demand is not just because of scarcity, but because of an intrinsic preference for empathy, connection, and interacting with another person. In one of our experiments, we used an incentive-compatible way to ask how much people were willing to pay real money for an art print. In between-subject conditions, we stated there was only one print, either made by AI or a person; the person-produced print was valued much higher. In other conditions, we said there were 500 prints, and the price for the human-made one dropped significantly because it was no longer seen as making a connection with one artist.
A horse was an input into an output that you could replace. This relational story only works—and we need more data on this—if a human is not a horse, providing value from the output such that if you replace the human, the value of the output decreases.
Messy Middle scenario
Molly Kinder has written about a "Messy Middle" scenario, which made me wonder if a much faster AI takeoff might be better for wealth distribution. AI could automate jobs and cause mass layoffs without creating enough immediate wealth to pay off those displaced and achieve a Pareto improvement where everyone benefits.
While resources saved by replacing humans with AI still exist in the economy, allocative inefficiency and political problems arise. The government doesn't know exactly who was laid off by AI. For instance, is it politically sustainable to give a laid-off Meta worker a $200,000 annual check when many other working people make much less?
I guess that if automation creates such a political problem, the economic pie will also grow rapidly—unless AI is only slightly more productive in those professions, making the capital to replace software engineers nearly as expensive as the engineers themselves. In the long run, Jevons paradox suggests we cannot anticipate how we will use more software, but there will surely be more uses.
Andy Hall wrote a blog post observing that if unemployment increases by just 2%, the political winds change completely. Referring to Molly's essay, I think a "drip scenario" is one of the worst outcomes because of the political economy; people might not be unemployed en masse, but instead move into lower-paying sectors.
Between 1920 and 1940, phone operators were completely automated, but it took 20 years even though the technology existed. A QJE paper shows they were reabsorbed into the economy, mostly underemployed and at lower salaries.
An emergency is a quick uptick in unemployment, perhaps 2-3%. The concern is whether the resources saved from white-collar workers are enough for a broad-based redistribution scheme if they don't grow the economy. Unless you can figure out how to get funds to them specifically, you face the problem of whether you can fund a UBI from layoffs that displace people without growing the technological frontier of what the economy can produce.
Historically, the technological frontier has expanded whenever this happened. It is hard to imagine a scenario where intelligence is just enough to replace a software engineer but still costs nearly as much, providing no abundance effect. If the pie doesn't grow, where does redistribution happen?
For that to occur, you would have to be able to automate white-collar jobs only piecemeal—for example, automating software engineers but not accountants or analysts. My model of intelligence suggests that if you can lay off all software engineers, you have enough capability to automate all kinds of white-collar work. Because AI will be cheaper than human labor and create huge potential savings, the messy middle scenario where we lack the wealth to go around seems unlikely. The question then is the best way to tax and redistribute it.
How to tax and redistribute AI wealth
There is differential complexity in implementing these things and they differ in the timeline of being helpful. Universal basic capital won't generate returns in six months, whereas a negative income tax provides an immediate insurance floor where everyone gets a certain amount of money and is taxed more as they earn more.
I worry about the political economy implications of UBI; if people are dependent on a check, it matters who is in power. We are currently endowed with labor that turns into income, but being at the mercy of elected officials for basic needs is a dangerous power-sharing arrangement. With universal basic capital, you have an ownership share and property rights, but this depends on indexing. The problem with universal basic capital is targeting: what if Anthropic goes to zero but some random robotics company takes over?
A negative income tax shares UBI's risk that someone in power could remove the floor when people can't work. Regarding the wealth tax, I concern that there is no politically sustainable equilibrium at 0.5%. It might start low for a war and escalate until the marginal income tax rate in the US is around 40%, and upwards of 50% in certain states. With a capital tax, one might worry it distorts investment, leading people to ask why they should invest in Anthropic or Intel if the government takes larger shares and dilutes them.
It is worth separating how revenue is raised, what is taxed, and how it is distributed. The government could use a broad-based tax on externalities or land to buy and hand out shares of Anthropic. Similarly, a consumption tax like a European value-added tax allows the government to buy stocks and distribute them to everyone. That was the proposal for privatizing Social Security: giving everybody a basket of stocks.
Why demand collapse is unlikely
People talk about whether there's a white-collar apocalypse. The Budget Lab at Yale is doing great analysis on this; their recent report shows that you have to squint to see anything happening. Across the entire economy, even in the most exposed sectors like software engineering, there's not really anything going on. There might be a signal that junior developers are getting jobs less than before, but that's a "less than before" rather than a level shift, as there is actually an increased demand for senior software engineers. Layoffs are likely just normal layoffs that were rebranded as AI layoffs.
There are public coordination devices. If we enter a narrative where firms not laying people off are seen as not adapting to AI, you'll get a cascade effect of firms keeping up with the Joneses. That's worrying, as a firm might be worse off after layoffs but does them to maintain the perception that they aren't behind the times. You've probably heard anecdotal stories of token counters maximizing tokens.
Right now, we have no evidence of a white-collar bloodbath. If you automate a complementary task, the human labor complementing that automation increases in value. A key statistic for this is elasticity of demand. In the O-ring model of jobs, a job is a series of tasks. If AI automates nine out of ten tasks and a person focuses on the remaining one, the job becomes more productive. If this creates a price effect where the product is cheaper and demand responds by increasing usage, it could lead to more hiring. Many on the internet argue that the uptick in software engineering demand suggests jobs are currently elastic enough.
This elasticity of demand argument is incredibly important for labels people use without understanding the underlying causation, such as Jevons paradox. This is the idea that as something gets cheaper, you want so much more of it that total spending increases, as happened with coal in Britain 200 years ago. However, this only happens if demand is highly elastic. For things like oil or insulin, demand is not super elastic; if they get cheap, it's not like magically there will be so many more cars that we use way more oil than before.
Long-run elasticity is higher than short-run elasticity. Even in the long run, agriculture is the famous example where we could produce way more food if we dedicated the same portion of the economy to it as we did 100 years ago.
Software is a particular kind of good where, as it gets cheaper, we will want more and more of it; it is not an inherent property of markets that you just keep wanting more as prices drop.
Citrini's viral scenario planning predicts that automation and powerful AI will cause a recession. If white-collar workers are automated quickly, they may be laid off and not find work, removing their salaries and causing a slump. This involves a reallocation of wealth and income from lower-income laborers toward tech capital owners.
For this to generate negative economic growth, demand must have a hard bound. Capital holders would eventually have to say, "I've had enough," and stop spending or investing in more fabs and data centers to run AGI. Unlike the Depression, the technological frontier is expanding here. It is very hard for abundance to generate negative economic growth.
One plausible mechanism could be the O-ring theory, which refers to how the Challenger shuttle blew up because one malfunctioning component destroyed the whole thing.
Human employees would be hard to integrate into the machine economy
You can't automate an entire job to an AI right now because you need extreme reliability to avoid destroying the finished good, which explains why there is less automation than there otherwise could be. However, once AIs are advanced enough, integrating humans into production flows will become difficult. Beyond humans being more expensive or less capable, whole production flows will be organized for AI labor, thinking thousands of times faster and communicating in neuralese. Even if a comparative advantage exists for hiring a human, transaction costs and reliability worries will make integration hard.
I want to distinguish this from the point that if you automate nine-tenths of a job, people might shift to the last tenth but face ten times more demanded work. Compare that to Gans and Goldfarb's model of O-ring automation: if you can only automate nine-tenths of a job at a lower quality standard than a human, you might not automate any of it. Symmetrically, we might stop using humans for one-tenth of a job because they cannot match the AI's quality or speed.
This model seems extremely plausible for why more lawyers, accountants, or software engineers are not automated. With lawyers particularly, you need an entity to back up the product, and regulatory layers keep humans in the loop regardless of their ability to perform the service. These frictions on political-type decisions we trust only to humans—legislation, being a judge or juror, or professional licensing—strike me as transitional. What we expect from humans and how we organize politics has changed throughout history, from hunter-gatherer bands to empires. Once an AI-run political system is much more efficient than the alternatives, it will probably out-compete them.
What if some humans (or AIs) value wealth accumulation intrinsically?
In the future, we will have different kinds of entities: AIs. Evolution once selected for agents with specific drives and preferences because they survived the most, and those preferences now determine what a hundred-trillion-dollar world economy produces. Even without catastrophic misalignment where everyone is killed, there will be evolution of firms that have AIs as part of them. Selection will favor agents that grow, save more, and have unsatisfiable demand for resources, with compute being an obvious one.
If a fully autonomous AI has its own welfare and makes its own decisions, I have no prior that it would prefer to deal with humans. It is an important question whether human preferences for trusting and empathizing with other humans versus a simulated AI will change. Some people will see an AI therapist as a superior product and won't need human empathy. Using Jonathan Haidt's framework, others have a moral emotion against offloading social interactions to an AI; they are the ones with a preference for other people. As noted on the last podcast, we are buzzing with natural selection, so even current indifference might evolve into a stronger preference for other humans.
Mark Zuckerberg hires MMA instructors and dancers for his wife's birthday, but most of his wealth is stock in Meta. As a controlling shareholder, he could turn that wealth into dividend income for consumption, but he prefers his wealth to compound and Meta to build more data centers. The wealthiest humans, whose wealth is compounding, have an almost Nick Landian preference for accelerating capital.
If two people satiate equally quickly in capital, but one also likes human-intrinsic services, the marginal value of capital for each remains basically the same. However, if one person doesn't satiate in capital because they are engaged by exploring the universe or turning their head into a galaxy brain, while the other does, the non-satiating person will rationally have a higher savings rate. In the long run, they will hold most of the wealth, and the overall capital share will be the capital share of that person's spending, which is one.
Currently, his investments go toward humans and machines, but I don't think he cares whether his future researchers and engineers are humans or AI. Some rich people don't seem to satiate quickly in capital, so they may save the most in the long run. Even if they reproduce more slowly biologically, that might not matter if they can live forever.
Regarding rich people investing rather than consuming, this depends on the returns to capital. Returns to data centers are currently super high, but if people become satiated with capital, returns will be lower and these rich people will consume more because the incentive to invest is smaller. Consider the general equilibrium: we have gotten tremendously richer since 1820. Many more people are investing, but a consumption response keeps people employed and labor share high. Investment may have to be titrated through actual laborers for it to work. In the future, only consumption is human-mediated because investment can be done by robots.
In a scenario with high labor share, returns to capital will be lower, meaning the growth rate must be lower than what we expect through the period of transformative AI. The capital stock could grow quickly, but the price of capital goods relative to consumption goods could fall faster than the capital stock grows. This is the difference between the potential frontier of technology and realized prices.
If the growth rate is high, I could put my money toward earning 30% interest in data centers, or I could buy a cool product resulting from technological breakthroughs; both are compelling options. We might want a ballerina performance with the same marginal utility as now, but the marginal utility of a robot might be much lower. In units of robots, we want it more than now. If every robot now turns into 100 robots next year, the interest rate in units of robots is 10,000%.
If the price of robots falls really fast, prices adjust in a way too many macro models don't allow for. This is investment-specific technical change: the price of capital falls relative to the price of consumption, rather than the standard macro view of a single "output" chimera allocated to capital or consumption. Every unit of capital next year gives up far less consumption than a unit this year. One robot now turns into many robots next year, but the number of ballerinas remains the same. If those extra robots are different varieties and I don't get satiated, it's a different story. On the investment side, a greedy titan of industry wanting more robots alone could increase the marginal value of robots and decrease labor share.
Historically, greedy titans of industry built libraries. Wealth accumulation was often part of a social interaction among peers and community to be admired. The stylized fact is that you accumulate capital and then buy things. As these titans near the end of their lives, they often hand wealth to children who are worse stewards, failing to grow wealth even at the rate of the economy. Consequently, they give it to a trust. If people live longer or align their trusts to the accumulation process, the evolution feels strong. A few agents thinking this way could determine the preferences of the whole economy because this sector grows faster than others.
Regarding satiation and diminishing marginal utility, if a person has an intrinsic preference for accumulation, your story is right. Once you have enough hedonics and social status—which Rousseau and St. Augustine wrote about—accumulation becomes the goal. We are arguing that a high concentration of wealth allows a few exceptions to the rule to be enough. Historically and today, we see these exceptions; they haven't taken over the economy because of dissipation shocks, such as children squandering wealth or foundations spending it. This preference is not hypothetical.
Even without an intrinsic preference, there are instrumental reasons to value accumulation, such as a desire for political, philosophical, or religious influence. People enter an arms race over what society looks like and what people believe. Similarly, there is total utilitarian philanthropy. As a classical utilitarian, one way to have an almost unsatiating utility function is to create new happy beings. This goes back to Bostrom's astronomical waste point: putting Dyson spheres around stars to turn energy into happy simulations. They have high marginal value for the solar system they occupy because that leads to more solar systems. A von Neumann probe can exist, but we only count final consumption and investment goods as GDP. If we recognize a person that owns itself, optimizing between spending on a baby von Neumann probe that colonizes another star system or a ballerina, and it doesn't value the ballerina much, the labor share might be high the way we usually count it.
What should developing countries do?
Unless you are producing AI models, the hardware that goes into them, Korea's HBM, Taiwan's fabs, or the Netherlands' ASML, you are not in the production chain. I believe the economics profession's biggest lack of allocated resources is thinking about middle-income developing countries in the age of AI.
There are scenarios where AI technology dissipates to Nigeria and other developing countries, leveling the playing field and giving them a capabilities level-up. But in another world, because they lack the resources to train models and the necessary hardware, they get completely left behind. Because of automation, we can now produce commodities in developed countries. This seems an extension of the messy middle case.
The messy middle might only be bad in a narrow range of scenarios, not just because redistribution would be easy as the pie grows, but because the interest rate would be way higher and the price of everything except human-intrinsic goods would fall rapidly. Things must go very wrong for capital to be productive enough to automate work, but not productive enough that the interest rate is high and the price of capital-produced goods falls significantly. Even without redistribution, a little savings will save many people, allowing them to consume a lot using those savings since they start from such a lower level indexed to the global economy. I have no strong feelings on whether this should take the form of sovereign wealth funds investing in supply chains or subsidies to citizens.
Regarding why the descendants of the Rockefellers don't control everything, if our argument about greedy optimizers holds, one reason is that it is very hard to index the economy. Had they decided to have their heirs index the economy, they would be trillionaires now, but before index funds, it was simply too difficult. A very small fraction of the economy from 100 years ago accounts for a majority of the value created now; if you missed those things, your wealth stagnated. There may have been a brief golden window from the creation of index funds until five years ago to grow wealth at the rate of the economy, but we are now in a world of very concentrated returns, especially to private companies.
As we noted in our blog post, the average person has disproportionately less access to this capital. In the US, most of their capital is a random house, which is uniquely ill-suited to be complementary to the production or serving of AI and robots, or the goods the rich will bid up. Land is close to other humans, but modulo relational stuff, it will not be the main factor of production in the future. This is why a Georgist tax would not raise enough money for the programs we will be discussing.
If it gets harder to index the economy, normal people and developing countries lose their main way to have purchase on the wealth from AI, modulo some sort of universal basic income. It is not enough for Nigeria to just own the S&P 500; does Nigeria own a lot of SK Hynix and Anthropic?
Consider ConEd, the electricity provider here. It is a monopoly providing a resource everybody uses, but we do not think of electricity as creating a concentration of political or social power because the downstream benefits came to the users rather than the producing entity. Social media was the opposite: everybody used it, but the rents went to the platform. If our economy is run on AGI the way it currently runs on electricity—a broad fundamental transformation—it looks like electricity. Every company in the future S&P 500 will be there because it leveraged AI.
However, given how concentrated the S&P has become with big tech companies, it is hard to reason about how much of the gains private companies will control. Open models are a big point here. If open models stay six or nine months behind the frontier, then when we hit AGI, everybody has access to this resource in six months. Whether there are runaway gains connects to technical forecasting questions about recursive self-improvement, continual learning, or online learning, which impact whether Uganda will have any purchase on the returns of AGI.
In one world, AGI is concentrated and hard to index. In another, it is like electricity: every company has access, and Nigeria just needs to buy the index and use open models. I would prioritize trying to index, given how fast AI could hit the world. In messy middle or long-timeline cases where AGI is not imminent, it would be leaving value on the table if you could have retrained to be better educated on the latest wave of computing. Perhaps the reason to be pessimistic is that a country is often poor because it has a bad education system, so becoming the best at retraining people to use AI may not be a promising strategy for that poor country.
In developing countries, we've seen a leapfrogging effect with mobile banking, which is much more prevalent in Nigeria than in Germany. With a transformative technology like AI, you could similarly skip the middle step and achieve astronomical growth.
As discussed in our essay and by others, it is already not that hard to index. While there has been an increase in the privatization of returns, well under 20% of the total market cap of non-tiny companies in the US is private. Everyone thinks about OpenAI and Anthropic; if that is where all the wealth accrues, then questions about whether open models stay only a little bit behind are important. However, AI might alleviate the frictions, such as disclosure requirements, that keep companies from going public. I guess the long general trend of lowering those frictions and making it easier for people to index will continue, despite the recent bump in the other direction.
AI will be more popular and more likely to lead to broad increases in prosperity if it is as hard to capture its gains as it was to capture the gains of electrification. There are no anti-electricity people. The current negative narrative around AI exists because people aren't putting out a positive one; it is easier for someone on a podcast to say the jobs you like are going away than to spin up a utopia that doesn't exist yet.
One big cost of commoditized frontier AI models is the tech race dynamic. For safety, you might want fewer frontier companies so each has a buffer to slow down. Regarding the widespread access of returns, I think there is less of a trade-off than some imagine. Some believe either frontier AI gets commoditized and we all enjoy the benefits despite a cutthroat market, or things are safer because of a big gap between the leader and the laggard. You could instead have a relatively big gap where the leader is a public company with widely distributed ownership.
I have recently been thinking that the risk of commodification—which diffuses the ability to use AI for harmful ends—is worth the benefit. I worry that concentrated labs not only prevent the surplus from being widely distributed through society but also create a clear political target for the government, as seen with the Defense Production Act threat against Anthropic. If there weren't one or a couple of labs clearly ahead of others, this kind of threat would be much harder to make.