Google DeepMind has a Director of AGI Economics. That job title did not exist three years ago, and its existence tells you something the lab’s model release calendar does not: frontier labs are hiring economists to design the distribution architecture for a world where cognitive labour is no longer scarce.
Alex Imas, who holds that role at DeepMind, sat for a multi-hour interview with Dwarkesh Patel on June 7 alongside Philip Trammell, an economics postdoc at Stanford’s Digital Economy Lab. The conversation is the clearest signal yet that AGI distribution economics has moved from think-tank speculation into active policy design at the labs themselves.
The central argument is structural. In an AGI scenario, labour scarcity disappears. What replaces it as the binding constraint is a set of factors that cannot be automated: natural resources (land, energy, specific minerals), institutional trust (regulatory permission to deploy), and human attention as a positional good. Who owns those factors determines who captures the gains. The question Imas and Trammell are working on, in plain terms, is whether that ownership distribution can be shaped before the transition, not after.
Imas’s tax argument is specific and worth stating precisely. He holds that capital-gains frameworks systematically under-tax AI productivity gains, and that the correct structure looks closer to a Henry George land-value tax applied to compute capacity, combined with a citizens’ dividend funded from compute royalties. The logic is Georgist: tax the rents on the input that cannot be reproduced, not the returns on labour or investment. This framing sits directly adjacent to the Public Wealth Fund proposals circulating in policy circles this week, which propose sovereign AI compute ownership as a redistribution vehicle.
Trammell offers the more contrarian position on global inequality. The popular concern is that countries outside the AI supply chain lose. His counterargument: if the price of cognitive labour falls by two orders of magnitude globally, every economy that employs educated workers gains in absolute terms relative to the current baseline, regardless of whether it owns a frontier lab. The gains are asymmetric, labs capture disproportionately more, but the distribution is positive-sum at the global level. The historical analogy is not perfect but it is useful: electrification and the internet both concentrated early returns to capital before labour markets adjusted and redistributed the surplus.
The question neither Imas nor Trammell can answer is whether the AI transition follows that pattern. Both are explicit that they do not know when AGI arrives and that their arguments are conditional on it arriving. What is unconditional is the timeline argument: Imas holds that the political coalition for redistribution is materially easier to build before wealth has concentrated in compute owners than after. Once the transition is complete, the actors with the most to lose from redistribution are also the actors with the most capacity to block it.
This is also the counterpoint to the subsidy economics conversation running in parallel. The cost-side debate, about who pays for training runs and whether government subsidies distort the frontier, treats the financial structure of the AI industry as a near-term operating question. Imas and Trammell are working the same problem from the other end: the financial structure of the AI industry is also a decades-long distribution question, and the two angles are not separable.
The piece worth reading against this one is the Public Wealth Fund proposal. Both arguments land on a similar policy instrument, compute ownership as the lever, from different directions. One arrives via fiscal theory, the other via sovereign wealth design. The convergence on compute as the taxable or ownable unit is notable.
Builders and operators focused on product cycles have limited reason to act on this analysis in the next quarter. The policy window Imas is describing is measured in years, not sprints. But the signal that DeepMind created the role at all should recalibrate how seriously you take the distribution question. The labs are not assuming governments will sort this out independently.
Dwarkesh Patel (dwarkesh.com) interview with Alex Imas (DeepMind) and Phil Trammell (Stanford), 2026-06-07.
