What Gets Scarce When Intelligence Is Cheap? Two Economists Try to Map the Post-AGI Economy

The conversation most economists are not having about AI goes something like this: forget which companies win. What actually becomes scarce — and for whom — when intelligence is abundant?

That question drove a 76-minute episode of the Dwarkesh Podcast 1 recorded on June 4, 2026, with two economists whose day jobs put them closer to the frontier than most: Alex Imas, Director of AGI Economics at Google DeepMind and Professor of Economics at the University of Chicago, and Phil Trammell, Head of Economics at Epoch AI and Research Scholar at Stanford. Host Dwarkesh Patel frames the session as a "scenario-building" exercise — not a prediction, but a map of which assumptions lead where.

Imas opens with a disarming admission: individual economic forecasts about AI and labor are not very useful. His evidence is David Ricardo, circa 1820. Ricardo watched the Industrial Revolution begin, first predicted prosperity, then reversed himself and warned of mass unemployment. Both predictions were directionally correct about the specific jobs that existed in Ricardo's time — they did get automated. What Ricardo missed is that prime-age employment in 2026 sits near its all-time high. Every job that vanished was replaced by something he couldn't imagine. 1

The lesson Imas draws is structural: instead of asking "what will happen," economists should map which premise generates which scenario, then collect data to figure out which premise is closer to true. "We don't have any data," he says flatly. "I've been saying we need a Manhattan Project for data. We don't have data on consumer demand elasticities. We don't know what they are."

That framing shapes everything that follows.

For three centuries, roughly 60% of GDP has been paid to labor in wages. The remaining 30–40% went to capital owners — machines, land, company shares. This stability — the "Kaldor fact" — survived the steam engine, electrification, and computing. 2

Trammell's contribution is to point out that this stability rests on something specific: nothing has ever been fully automated when you trace the supply chain. Even electronic products in the US, which look almost entirely capital-intensive, have a network-adjusted capital share of around 50% — because the machines that automate the final step were built by workers, and those machines were designed by engineers. Labor is embedded deep in every supply chain.

AGI represents the first technology that could, in principle, close that loop entirely. Some goods could reach a network-adjusted capital share of one. The implications, Trammell argues, are genuinely ambiguous — not because it's unclear that this would be a big deal, but because whether labor share rises or falls depends on a question nobody has answered yet: do automated goods satiate demand, or do they expand it?

The standard framing for "what stays human" is the ballerina: a performance whose value comes precisely from the dancer being a biological person. Machines don't go to the ballet; only humans value human performance. So as everything else gets automated and cheap, spending concentrates on human-intrinsic goods, labor share stays high.

Imas thinks this framing is seductive and wrong. His counterexample: transistors. The world has produced roughly a quadrillion times more transistors than in 1960, yet computing's share of the global economy has been falling for most of that time (the economist Chad Jones documented this). More transistors, lower share. The reason: variety didn't expand fast enough to absorb the abundance.

What looks different about AI, for the first time, is that H100 GPUs cost more to rent now than three years ago, despite far greater supply — because new uses for intelligence are appearing faster than supply can clear the market. If that dynamic persists, the machine sector never satiates, capital share climbs, and labor share falls even if everyone is still employed as ballet dancers and baristas for each other. "One robot now turns into many robots next year," Trammell says, "but the number of ballerinas is the same."

Imas pushes back on the ballerina frame for a second reason: real jobs are bundles of tasks, not single roles. A doctor fills out insurance forms, calls pharmaceutical companies, and occasionally delivers a diagnosis. AI could automate everything except that last moment of clinical connection — and if consumers value the human-in-the-loop enough for just that task, the doctor's job survives as what Imas calls a "relational sector" job. The problem: we have almost no data on how much people actually value human involvement for specific tasks. His lab is running conjoint experiments — real money changing hands, real willingness-to-pay measurements — but the field is early.

The art print experiment is instructive. When told there's only one copy and it was made by a human, subjects pay a significant premium. When told there are 500 copies and they're all human-made, the premium collapses — because the connection to a specific person is gone. AI art, by contrast, commands the same low price regardless of scarcity. Humans value the relationship with a specific maker, not just the human-ness of the medium. Whether that logic scales to medicine, therapy, legal advice, or tutoring is an open empirical question.

Dwarkesh Patel raises the scenario Molly Kinder named the "Messy Middle": AI automates enough to displace many workers, but doesn't create enough new wealth to redistribute and make everyone whole in time. A Pareto improvement in theory, a political crisis in practice.

Trammell and Imas think this scenario is possible but narrow. Their core argument: if AI can really automate all software engineers, it likely has enough general capability to automate accountants, analysts, and most white-collar work simultaneously. The savings are enormous. The scenario where you automate a specific class of workers but the productivity gains are "just a hair" more than the cost of the AI — barely enough to justify the layoffs without generating redistributable surplus — requires a very specific and unlikely knife-edge.

The phone operator analogy is telling: the technology to automate telephone switchboards existed in the 1920s, but full automation took 20 years. Workers drifted into other sectors at lower salaries. No political emergency. No moment of reckoning. That slow drip may actually be worse for political economy than a sharp shock, Imas argues, because a sharp uptick in unemployment — even 2-3% — triggers emergency fiscal responses (COVID proved the government can move fast). The drip scenario produces underemployment without mobilizing redistributive politics.

When Patel asks what the optimal redistribution mechanism is, Imas lays out a menu with frank assessments of each:

UBI: Immediate income floor, but creates dangerous political dependency. "Right now, we're endowed with labor that can turn into income. When that is no longer the case and we are at the mercy of the elected official for basic needs, that feels like a power-sharing arrangement that's really dangerous."

Universal basic capital (giving everyone shares in AI companies): More stable than income transfers because you hold property rights, not a promise. The problem is targeting: what do you put in the portfolio? If you bet on Anthropic and it goes to zero while a robotics startup wins, you're worse off than a broad index.

Wealth tax: History suggests it starts low and escalates. Income taxes began at marginal rates that seem quaint today. Imas worries about investment distortion as rates creep up.

Consumption tax + sovereign wealth fund: A European-style VAT that feeds into government equity purchases, then distributes shares broadly. This was actually the original proposal for privatizing Social Security. Trammell suggests this is underrated because it separates "how to raise revenue" from "what to tax" from "how to distribute" — you can optimize each layer independently.

Both economists land on the same bottom line: you'll need a layered approach, and some tools work faster than others. Universal basic capital takes years to compound; a negative income tax can be activated overnight.

One of the most striking passages in the conversation involves what Imas calls "the biggest lack of resources allocated in the economics profession" — thinking about middle-income developing countries in the age of AI. India. Nigeria. Countries not building models, not making chips, not hosting the fabs.

The standard advice — retrain workers, build data centers — Imas views skeptically. You can't leapfrog to being the world's best at AI-assisted work if your education system is weak. What you can do is index.

The key insight from Trammell: the window when ordinary indexing worked well (buying a diversified equity portfolio and riding the growth of the global economy) has narrowed. Returns are increasingly concentrated in private companies — OpenAI, Anthropic, pre-IPO AI infrastructure firms. Under 20% of non-tiny US company market cap is private, but the companies capturing the most AI upside are disproportionately in that 20%.

Developing countries with sovereign wealth funds should be acquiring exposure to AI supply chains now, before that window closes further. The analogy Imas reaches for is mobile banking in Nigeria — a country that leapfrogged legacy financial infrastructure by going straight to phone-based payments. AI might offer a similar leapfrog opportunity, but only if countries act before the returns are fully priced in. 3

The episode's most practically important framing surfaces late: will AI distribute gains like electricity did, or concentrate them like social media?

With electricity, the downstream benefits went to users of electricity — every company that plugged in. ConEd doesn't run the economy; it just powers it. With social media, the rents went to the platforms themselves, not the users who created the content.

Imas's tentative view: if open models stay 6–9 months behind the frontier and AGI is eventually commoditized, AI looks more like electricity — every S&P 500 company that survives will have leveraged AI, and owning a broad equity index means owning a broad slice of that value creation. If the frontier labs stay private indefinitely and closed-model advantages compound, it looks more like social media.

The open-source trajectory matters more for inequality than almost any policy lever.

One data point grounds all the scenario-building: as of June 2026, the Yale Budget Lab's analysis shows almost no macroeconomic signal from AI-driven unemployment. "You really have to squint to see anything happening," Imas says. 1 Junior developer hiring may be slightly below trend; senior software engineer demand is if anything up. The white-collar disruption is not yet visible at the aggregate level.

Imas offers two competing explanations. One: the O-ring model — if AI automates 9 of 10 tasks in a job but quality requires all 10 to be excellent, you can't automate the job until you can automate the 10th. Reliability thresholds are keeping humans in the loop. Two: social coordination — if firms perceive that not laying off workers signals AI underadoption, you could get cascading layoffs driven by narrative rather than productivity. Token-counting managers, "we're using AI" press releases. Neither story is settled.

The honest summary from a 76-minute conversation between two economists who study AGI for a living: the number of open questions exceeds the closed ones by a factor of several. The useful output is a map of which questions to watch — labor demand elasticities, capital share trends, the gap between frontier and open models, whether human-intrinsic preferences hold under sustained AI exposure — not a prediction. The Mongolian economist of 1400 who tried to forecast today's economy would have been wrong about almost everything. Imas and Trammell are at least honest about standing in the same position.