Yale economist Restrepo’s paper analyzes wages and growth mechanisms in an AGI world
Yale University associate professor of economics Pascual Restrepo’s paper, We Won't be Missed: Work and Growth in the AGI World (NBER Working Paper 34423, 2025), analyzes long-term economic growth and the labor market under the assumption that artificial general intelligence (AGI) is feasible and computing power continues to grow. The paper divides work into bottleneck work (tasks essential to growth and that must expand) and complementary work (tasks not essential to growth).
The paper’s main theoretical result is that, on the premise of continued computing expansion, the economy will automate all bottleneck work, output will shift from a multiplicative relationship between computing power and human labor to an additive relationship, and the long-term growth rate will equal the growth rate of computing power. Wages converge to the opportunity cost of the computing power needed to replicate human work (computing-equivalent cost), and the labor income share in GDP approaches zero. Some complementary work may still be done by humans. The paper notes that if computing power and human skills are the only scarce resources, then the sum of wages after AGI will not be lower than before AGI, and average wages will be higher.
The paper’s abstract also says that when algorithmic progress becomes the bottleneck, the transition may take a sawtooth shape, and wages may drop sharply to the computing-equivalent level after the last task for a specific skill is automated. The author disclosed that he has served as a member of Anthropic’s Economic Advisory Board since May 9, 2025. The paper is a theoretical derivation and provides no empirical data or specific timeline.
Original sources:
Title: We Won't be Missed: Work and Growth in the AGI World Publisher: Pascual Restrepo (NBER Working Paper 34423) Publication date: October 2025 (working paper); included as a chapter in The Economics of Transformative AI (2026) URL: https://www.nber.org/papers/w34423 Source type: academic working paper Verification status: actually read the abstract and related descriptions in this review, verification date 2026-10-26 Supporting content: distinguishes bottleneck and complementary work; automates all bottleneck work; output becomes additive in computing power and labor; growth is driven by computing expansion; wages converge to the opportunity cost of computing power; labor share approaches zero
Title: Pascual Restrepo on AI, automation, and the future of work Publisher: Yale Department of Economics Publication date: 2026-03-17 URL: https://economics.yale.edu/news/260317/pascual-restrepo-ai-automation-and-future-work Source type: university official introduction Verification status: actually read in this review, verification date 2026-10-26 Supporting content: the paper discusses AGI making computing power a growth constraint; growth is driven by the expansion of computing resources; income flows more to owners of computing power
Title: A Yale economist says AGI won’t automate most jobs—because they’re not worth the trouble Publisher: Fortune (republished on Yahoo Finance) Publication date: 2026-04 URL: https://finance.yahoo.com/sectors/technology/articles/yale-economist-says-agi-won-080200425.html Source type: media report Verification status: actually read in this review, verification date 2026-10-26 Supporting content: the paper’s argument is that most work may not be automated because it is not worth spending computing power on; skills are repriced according to the opportunity cost of the computing power needed to replicate them; labor share approaches zero
ABAB AI Insight
Restrepo’s model is built on two premises: AGI makes it possible, in principle, to use computing power to perform all economically valuable work, and computing power is limited at any point in time but can grow without bound. Under these conditions, growth requires bottleneck tasks to expand in step with it. Because the number of human workers does not increase in proportion to computing power, the model derives the result that bottleneck work is ultimately automated; otherwise, human labor would constrain overall expansion.
After automation, the production function shifts from a multiplicative relationship between computing power and human labor to an additive one. At that point, the pace of economic expansion depends directly on the growth rate of the stock of computing power, without requiring additional improvements in technological productivity. Wages are no longer determined by marginal contribution to growth, but by the opportunity cost of the resources needed to replicate that skill with computing power. Human work can still receive positive wages, but those wages equal the value of the computing power that substitutes for it, rather than its share in total output.
The decline of the labor income share to zero is a share-level result, not a decline of wages to zero in absolute terms. The paper also notes that if one can construct a subeconomy without AGI and maintain the original production, total wages will not be lower than before AGI. The transition path depends on whether the binding constraint is computing power or algorithmic progress: in the former case, wages change continuously; in the latter, a wage drop may occur when the last task for a specific skill is automated. The paper does not make empirical predictions about specific occupations, timing, or policy conclusions.
The observable indicators are whether subsequent research refines the criteria for identifying bottleneck tasks, whether the actual trajectory of computing growth tests the model’s parameters, and whether empirical observation of the long-term change in the labor share is consistent with the theoretical direction.
ABAB News · Cognitive Laws
When all bottleneck work can be automated by computing power, the long-term growth rate equals the growth rate of computing power, and the labor share approaches zero.
Wages are determined by the opportunity cost of the computing power needed to replicate that skill, rather than by the skill’s contribution to total output.
Because complementary work is not essential to growth, it may still be performed by humans if the opportunity cost of computing power is too high.