1 We Wont be Missed:2 Work and Growth in the Era of AGI34 Pascual Restrepo5 pascual.restrepo@yale.edu6 Yale University78910 July 4, 2025111213 Abstract1415 This chapter explores theoretically the long-run implications of Artificial General Intelligence16(AGI) for economic growth and labor markets. AGI makes it feasible to perform all economically17valuable work using compute. I distinguish between bottleneck and accessory work—tasks18essential vs. non-essential for unhindered growth. As computational resources expand: (i) the19economy automates all bottleneck work, (ii) some accessory work may be left untouched by AI20and assigned exclusively to humans, (iii) output becomes linear in compute and labor and its21growth is driven by the expansion of compute, (iv) wages converge to the opportunity cost of22computational resources required to reproduce human work, and (v) the share of labor income23in GDP converges to zero.24 This chapter studies the long-run behavior of wages and growth in an economy where Artificial25General Intelligence (AGI) is developed and computational resources increase over time. AGI allows26the economy to complete all relevant work using computing systems. These systems consume27computational resources but do not require human input, guidance, or effort to accomplish work.28 The key economic problem is how to allocate finite (but growing) computational resources and29human labor to accomplishing the work needed to produce output. The chapter introduces a key30distinction between bottleneck and accessory work.3132 • Bottleneck work comprises tasks essential for economic growth. Output cannot expand indefi-33 nitely unless inputs in bottleneck tasks also expand or become infinitely valuable.3435 • Accessory work is non-essential to growth. Output can expand indefinitely even if these tasks36 are discarded or limited in input.3738 My main theoretical result shows that all bottleneck work is eventually automated while some39accessory work may be left untouched by AI. Once this occurs, output shifts from being multiplica-40tive in compute and human effort to being additive, and the long-run growth rate of the economy41is pinned by the growth rate of compute.42 Despite AI completing all bottleneck tasks without human input, people still hold jobs. They43can contribute by fulfilling bottleneck work. Human labor remains valuable because it saves44scarce computational resources. Alternatively, workers may perform accessory work, where it is45impractical to use compute since we already have too many workers. In the first case, wages are46pinned by the value of computational resources saved. In the later, wages are bounded above by47the value of computational resources it would cost to automate accessory work.48 In sum, the advent of AGI changes the way labor is valued. Before AGI, wages reflected the49importance of bottleneck work and the scarcity of labor with the requisite skill for this work.50With AGI, wages reflect the computational cost of replicating the work produced by all human51labor. Despite the fact that human labor retains some value, its contribution to GDP and growth52becomes vanishingly small, with the share of labor in GDP converging to zero, and all income being53eventually accruing to compute.54 I then expand the analysis to an economy where AGI can be used to complete scientific work,55accelerating the pace of technological progress. Without AGI for science, technological progress56is constrained by population growth (as in the semi-endogenous growth models of Jones, 1995;5758 159 Kortum, 1997; Segerstrom, 1998). With AGI, all scientific bottleneck work is automated and the60rate of technological progress is determined by the growth rate of compute. This may generate61sustained exponential growth despite shrinking population, but does not create a singularity or62infinite growth explosion.63 The analysis here complements existing work on the economics of AI, including several contri-64butions in a previous NBER volume (see Aghion, Jones and Jones, 2019; Acemoglu and Restrepo,652019). Methodologically, I build on ideas from task models, which center on the problem of how to66accomplish multiple work tasks using different production techniques (Autor, Levy and Murnane,672003; Acemoglu and Autor, 2011; Acemoglu and Restrepo, 2018, see). As in these models, the core68economic problem I study is that of allocating work to human labor or AGI systems, and deriving69the implications of this process for wages and output. The focus on the transition to AGI is shared70with complementary work by Anton Korinek and collaborators that influenced some of my thinking71(see Trammell and Korinek, 2023; Korinek and Suh, 2024). Outside of traditional economics, there72is a large and interesting literature on world building exercises imagining how the economy and73world are transformed by AGI (see Kokotajlo et al., 2025; Drago and Laine, 2025).74 The chapter is organized as follows. Section 1 presents the analysis in a production economy75where AGI can be used to produce all production work with compute. Section 2 then extends76the analysis to a semi-endogenous growth economy where AGI can be used for scientific work.77Section 3 concludes with a discussion of policy and issues related to the distribution of income and78meaningful work.798081821 Production and Work with Artificial General Intelligence8384The economy produces output 𝑌𝑡 by completing work—the use of human or computational resources85to accomplish tasks needed to generate valuable output. The set of all valuable work is Ω, with each86specific type of work indexed by 𝜔. I assume Ω is a finite (but large set).187 The quantity of work 𝜔 completed at time 𝑡 is given by8889 190 𝑋𝑡 (𝜔) = 𝐿𝑡 (𝜔) + 𝑄 𝑡 (𝜔). (1)91 𝛼 𝑡 (𝜔)92 1 Infinitesimal similar results but technical complications9394959697 298 Work 𝜔 can be accomplished by human labor 𝐿𝑡 (𝜔) or using computational resources 𝑄 𝑡 (𝜔). The99quantity of work produced by a human per unit of time is normalized to 1. Emulating or replicating100this work using computers costs 𝛼 𝑡 (𝜔) units of compute. We could have 𝛼 𝑡 (𝜔) = ∞ so that it is101impossible to replicate human effort in 𝜔 with computers.102 The quantity of output produced is then103104105 𝑌𝑡 = 𝐹({𝑋𝑡 (𝜔)}𝜔∈Ω ),106107108where 𝐹 is increasing in all entries, differentiable, concave, and exhibits constant returns to scale. I109also assume marginal products converge in [0, ∞] for any converging sequence of inputs in [0, ∞].110 There are different human skills, indexed by 𝑠 ∈ 𝒮, and in quantity 𝐻(𝑠). People of skill 𝑠 can111accomplish work in Ω(𝑠) ⊆ Ω, where the Ω(𝑠) partition Ω. This implies112113 Õ114 𝐿𝑡 (𝜔) ≤ 𝐻(𝑠).115 𝜔∈Ω(𝑠)116117118The partition Ω(𝑠) captures the degree to which skills can only be used for a given type of work119(Ω(𝑠) is a one-to-one mapping from 𝒮 to Ω) or they are general (Ω(𝑠) = Ω and there is a single120universal skill). The assumption that there is no overlap between Ω(𝑠)’s simplifies exposition.121 The total quantity of computational resources at our disposal is 𝑄 𝑡 . Think of this as the total122number of computations the economy can perform per unit of time, given its data centers and chips.123The computational resource constraint of the economy is then124125 Õ126 𝑄 𝑡 (𝜔) ≤ 𝑄 𝑡 .127 𝜔∈Ω128129130The Premise: The objective here is to characterize the behavior of an economy facing two forms of131technological progress: the development of Artificial General Intelligence and the increased expansion132of its computing capabilities. What happens in an economy that reaches AGI and acquires enough133compute to run it?134 The first form of technological progress considered is the development of AGI. This premise can135be expressed in terms of the work equation (1) as:136137Premise 1 (The economy develops AGI.). For all work 𝜔 ∈ Ω, 𝛼 𝑡 (𝜔) reaches a finite value by date 𝑇(𝜔)138139 3140 and converges to a terminal value 𝛼(𝜔) ∈ (0, ∞) from there on.141142 AGI is the knowledge or technology for transforming raw compute into all types of useful143work. Acquiring AGI means we figured out how to train or create computer systems capable of144performing all work currently accomplished by humans. This definition distinguishes between AGI145and computational resources. AGI is the recipe of how to use compute—the input—to accomplish146any type of work—the output. Doing so consumes some of our finite computational resources, 𝑄 𝑡 .2147 While I take this premise as a point of departure in my analysis, it is helpful to clarify three148points about what is being assumed here.149150 • Feasible vs Realized: The definition of AGI states that, in principle, we can replicate what151 people do with enough compute. Depending on the computing costs 𝛼(𝜔) of replicating what152 a person does we may decide not to use computational resources to accomplish this work. The153 key point is that AGI makes it feasible to do so; just because we develop AGI does not mean it154 will be practical to carry out all work with computers, since our computational resources are155 finite and can be applied to more profitable uses.156157 • Physical work: Some work requires interacting physically with the world. The premise here158 is that, when needed, computer systems can control machine and hardware to accomplish this159 work. When appropriate, I re-interpret 𝑄 𝑡 as a bundle of energy and computational resources160 that computers use to carry out both cognitive and physical work.161162 • Social work: one may argue some work requires social interaction and must be carried out163 by humans. The “human touch” and “empathy” of a therapist or healthcare provider may be164 impossible to replicate, creating a premium for work completed by people. The assumption165 behind Premise 1 is that quantity has a quality all its own: any deficit of computer vs humans may166 be overcome by the sheer application of compute to the task at hand. Imagine, for example,167 an AI system that perfectly emulates the best therapists in the world, is at your disposal at168 any time, knows you perfectly, never gets tired, and is open to use 100% of its intellect and169 capabilities to help you overcome your anxieties. The system may be too costly to be practical,170 2 The definition used here differs from other concepts, such as Transformative AI (defined as “AI developments that171172significantly transform the economy”) or Super-intelligence (AGI that keeps recursively self-improves itself, reaching173levels of intelligence that far exceed human ones). The definition of AGI used here does not need AI to be super-174intelligent.175176177178 4179 but would you insist on visiting a human therapist when you could instead be treated by this180 superhuman team of world experts? Imagine an AI system that uses vast amounts of compute181 to diagnose and treat medical conditions. Would you keep a sick children from accessing182 this medical attention because you insist AI lacks “the human touch”? If valid, this argument183 implies that we can still represent such situations by high values of 𝛼(𝜔), which capture184 the extra computing resources needed to compensate for the “human touch”. Whether it is185 practical or not to perform such work with compute is a different question addressed below.186187 The second form of technological progress involves increased computational resources, 𝑄 𝑡 . The188premise here is that computational resources are finite at each point in time but grow over time and189becomes arbitrarily abundant.190191Premise 2 (Abundant Compute.). Computational resources 𝑄 𝑡 are finite but grow without bound in time.192193 This premise is motivated by historical trends and regularities, such as Moore’s law. The premise194is that we will continue to expand computational resources in the future, in the same way we have195expanded them historically, with no obvious ceiling in the near or mid term and a limit that far196exceeds current computational resources.197 Historically, we have greatly expanded computing capabilities since the creation of modern198computers and transistors in 1970. Computational resources can be measured in flops—the total199number of floating-point operations per second that all computers in the economy could collectively200perform. From 1980 to 2007, we increased compute by three orders of magnitude, from 1015 to 1018201flops. Currently, our economy has a peak capacity of 1021 flops. It is estimated that this could202increase to 1054 flops in the long run, suggesting that there is plenty of room for compute to grow.203For reference, a human brain is estimated to perform 1016 –1018 flops, so computational resources are204already en route to exceed human brainpower by orders of magnitude.3205 Progress in computing capabilities is assumed exogenous and is subsumed in the path for 𝑄 𝑡 .206As a benchmark, one could imagine computational resources 𝑄 𝑡 growing exponentially in time—as207in some variants of Moore’s law—, though my results only require 𝑄 𝑡 to become sufficiently large.208 3 See Hilbert and López (2011) for trends in compute over 1980–2007. See AI Impacts (2023) for estimates of current209210compute. See Bostrom (2003) and Bostrom (2014) for estimates of future computational resources. See Sandberg and211Bostrom (2008) and Open Philantropy (2020) for estimates of the computing power of the human brain.212213214215216 5217 Bottlenecks: In what follows, I let 𝐹({𝑋(𝜔)}) denote the output obtained when the quantity of218work is 𝑋(𝜔) and 𝐹𝜔 ({𝑋(𝜔)}) > 0 denote the marginal gain of work 𝜔.219 A core component of the analysis is defining bottleneck and accessory work.220221Definition 1. Work 𝜔 is bottleneck if, for any {𝑋𝑡 (𝜔)} such that 𝐹({𝑋𝑡 (𝜔)}) is unbounded, either i. 𝑋𝑡 (𝜔)222is unbounded or ii. 𝐹𝜔 ({𝑋𝑡 (𝜔)}) is unbounded.223224 The definition captures an intuitive notion of a bottleneck: a type of work that is necessary for225sustained growth, in the sense that it must expand or its price would inflate to exorbitant levels,226capturing an expanding share of output. Potential examples of bottleneck work includes feeding227and sheltering people, producing energy, maintaining the productive infrastructure of the economy,228advancing science, decision making, logistics and delivery, and maintaining national security as229well as policing nations to preserve order and stability. These types of work are mission critical in230the sense that is hard to imagine an economy that keeps growing in a sustained way but lacks any231of these components.232233Definition 2. Work 𝜔 is accessory if, there is 𝑋𝑡 (𝜔) such that 𝐹({𝑋𝑡 (𝜔)}) is unbounded while both 𝑋𝑡 (𝜔)234and 𝐹𝜔 ({𝑋𝑡 (𝜔)}) are bounded.235236 The definition shows that accessory work is the opposite of a bottleneck: we can grow the237economy while keeping this type of work fixed and without commanding too high of a price.238Potential examples include work associated with arts and crafts, performing for others, literature,239and hospitality and fine dining. We may also have design and customer support work, which may240not scale with the economy either. Finally, we may have judicial work, as well as work associated241with religious and civic organizations, as well as nature preservation groups. Even the work242performed by academic economists may prove to be accessory, as it is unlikely to become ever more243valuable in the future. These are just possible examples, as the exact partition of what turns out to244be a bottleneck or accessory work depends on future preferences, the structure of production, and245the key existential problems and challenges faced by future people.2462472481.1 Limit Behavior249250This subsection characterizes the limit behavior of the economy as 𝑡 → ∞ under Premises 1 and2512. I characterize the properties of a competitive equilibrium: an allocation of compute and human252253 6254 labor that maximizes output and where factors of production are paid their marginal products.255This choice highlights the key economic forces at play.256257Proposition 1. All bottlenecks are eventually automated while some accessory work may be left to labor.258259 The proposition clarifies how work is carried out in the AGI economy. All bottleneck work is260eventually automated and produced with compute, while some accessory work is left exclusively to261humans. Human labor may still produce some bottleneck jointly with AGI, adding to the quantity262of work obtained, or may specialize fully in accessory work. In the special case where all work263is a bottleneck, the proposition implies that all work is eventually automated and produced with264compute.265 This last result seems to counter Ricardo’s principle of Comparative Advantage. We can think of266the AGI economy as a world where we engage in trade with “a country of geniuses in a data center,”267exchanging compute for productive work.4 Wouldn’t Ricardo’s principle imply that the AGI and268human countries should fully specialize to maximize the gains from trade? Why is it optimal for269AGI to produce all bottleneck work?270 The reason why this logic breaks is instructive and provides an heuristic proof of the proposi-271tion. Suppose AGI specializes in the production of a subset of work with compute while human272labor specializes in the remaining components. Over time, the economy will become unbalanced,273producing an expanding quantity of the first type of work and a fixed quantity of the later. This274imbalance cannot be optimal: to keep the economy growing one must expand all bottleneck work at275the same rate. The only circumstance in which such unbalanced path can be optimal is one where276the non-automated work is accessory, as claimed in the proposition.5277 Let’s now characterize the behavior of output and wages once we reach the point when all278bottleneck work is automated. Define the compute-equivalent units (CEU) of skill 𝑠 as279280281 CEU(𝑠) = max 𝛼(𝜔).282 𝜔∈Ω(𝑠)283284285This gives the computational resources needed to replicate work carried by people of skill 𝑠 in the286 4 This apt metaphore is from Anthropic’s Amodei (2024).287 5 The idea that Ricardo’s principle of Comparative Advantage calls for full specialization is a common misunder-288289standing. Ricardo’s original 2 × 2 example predicts that at least one of the two countries specializes (see chapter 2 in290Feenstra and Taylor, 2017, for a textbook treatment). When one country is large and endowed with enough resources291(the AGI country in our economy), the equilibrium involves the large country producing both goods and determining292their relative prices. The small country (humans), on the other hand, specializes in a single good, as in Proposition 1.293294295 7296 most computationally complex task they perform. Premise 1 implies CEU(𝑠) is finite for all 𝑠.297298Proposition 2. Output converges to299300 Õ Õ301 𝑌𝑡 = 𝐴 𝑄 𝑡 + CEU(𝑠) 𝐻(𝑠) + 𝐻(𝑠) ,302 303 CEU(𝑠) (2)304 𝑠∈𝒜 𝑠∈𝒩305306307 < CEU(𝑠), where 𝒜 is the set of skills performing only automated work and 𝒩 the set of308for some CEU(𝑠)309skills that perform accessory work not automated.310311 Aggregate production possibilities in the economy become additive in compute and labor, with312labor expressed in units of compute. This holds for any initial production function 𝐹. For example,313if the different forms of work are combined a-la Cobb-Douglas, we would go from an economy314where compute and human labor are combined in a multiplicative way (before AGI is developed)315to one where they are combined in an additive way (once AGI is developed and we acquire enough316compute).317 To understand the result, let’s consider the special case where all work is a bottleneck. Propo-318sition 1 implies all work is automated, which means compute is used for all existing work. At this319point, we can write output (given an allocation of labor) as320321 1 Õ322 𝑌𝑡 = max 𝐹 {𝐿𝑡 (𝜔) + 𝑄 𝑡 (𝜔)} 𝑄 𝑡 (𝜔) ≤ 𝑄 𝑡 .323 324 s.t:325 𝑄 𝑡 (𝜔) 𝛼(𝜔)326 𝜔∈Ω327328329Let 𝑄330 e𝑡 (𝜔) ≡ 𝑄 𝑡 + 𝛼 𝑡 (𝜔) 𝐿𝑡 (𝜔) denote effective resources (in units of compute) allocated to work 𝜔.331We can rewrite the maximization problem as332333 1 e Õ Õ334 𝑌𝑡 = max 𝐹 { 𝑄 𝑡 (𝜔)} 𝑄335 e𝑡 (𝜔) ≤ 𝑄 𝑡 + 𝛼 𝑡 (𝜔) 𝐿𝑡 (𝜔),336 337 s.t:338 𝑄339 e𝑡 (𝜔) 𝛼 𝑡 (𝜔)340 𝜔∈Ω 𝜔∈Ω341342343where the economy maximizes output subject to a total resource constraint pooling compute and344human labor. Constant returns to scale implies a solution of the form345346 Õ347 𝑌𝑡 = 𝐴 𝑄 𝑡 + 𝛼 𝑡 (𝜔) 𝐿𝑡 (𝜔) ,348 349 𝜔∈Ω350351352for some 𝐴 > 0 equal to the rate of transformation of compute into output. To conclude, let’s353354355 8356 turn to labor. The allocation of labor that maximizes total resources assigns all labor of skill 𝑠 to357arg max𝜔∈Ω 𝛼(𝜔), which yields358359 Õ360 𝑌𝑡 = 𝐴 𝑄 𝑡 + CEU(𝑠) 𝐻(𝑠) .361 362 𝑠∈𝒮363364365 These steps clarify the economic logic behind Proposition 2. AGI allows us to produce all366bottleneck work with compute. At this point, compute pins the value of work: having a worker367producing one unit of work 𝜔 is the same as having 𝛼(𝜔) extra units of compute, which is the same368as having 𝐴 𝛼(𝜔) units of output. The best thing labor can do is to specialize in the work that saves369the most compute, which frees computational resources capable of generating an output 𝐴 CEU(𝑠).370 The above reasoning must be modified slightly by the existence of accessory work. Accessory371work that is automated can be handled as above. If some accessory work is not automated, it is372because it is not worth wasting scarce compute on it. Suppose 𝜔 is performed by workers of skill373𝑠 ∈ 𝒩 and not worth automating. Reallocating 𝛼(𝜔) units of compute from producing bottlenecks374(valued at 𝐴 𝛼(𝜔)) to producing 𝜔 (valued at MPL(𝑠)—the marginal product of skill 𝑠 labor) must375lower output, which implies376377378 𝐴 𝛼(𝜔) > MPL(𝑠) ⇒ 𝐴 CEU(𝑠) > MPL(𝑠).379380381This inequality shows that the value of accessory work performed by skill 𝑠 labor is bounded by382𝐴 CEU(𝑠) and must then eventually converge to an additive constant (this last claim is a nontrivial383implication of constant-returns to scale in 𝐹).6384 Proposition 2 shows how the automation of bottlenecks allows output to scale with compute,385sustaining economic growth.386387Proposition 3. Output grows at the same rate as computing resources 𝑄 𝑡 .388389 This provides cause for optimism: In a world of AGI, the economy can grow simply by expanding390computational capabilities. No other form of technological progress is needed after AGI makes it391feasible to turn compute into all kinds of work. This is true even if there is accessory work left392unautomated, as this work does not hinder growth.393 6 A subtle detail behind the statement of Proposition 2 is that human workers fall in two groups. Workers of skill 𝑠394395for whom one element of 𝜔∗ (𝑠) is automated perform work only in 𝜔∗ (𝑠) and all such work is automated. This is group396𝒜. The remaining workers are group 𝒩 .397398399 9400 How is the growing income distributed? In a competitive economy, workers and compute are401paid their marginal contribution to output. These can be read right away from the expression for402output in Proposition 2.403404Proposition 4. The real price of computing resources converges to 𝐴 and real wages to405406407 𝐴 CEU(𝑠) if 𝑠 ∈ 𝒜408 409 410 411 412 𝑊(𝑠) =413 𝐴 CEU(𝑠) if 𝑠 ∈ 𝒩414 415 416 417418 In an AGI economy, workers are paid the value of the compute needed to replicate their work.419Human skill remains valuable—even in a world where all work is automated—because it can be420used to accomplish useful work, saving scarce computational resources. This hints at an important421distinction: AGI does not render labor redundant; it makes it replicable through computation.422Since compute is scarce, human labor retains value, given by the opportunity cost of deploying423computational resources.424 Skills that perform accessory work command a discounted wage, in the sense that CEU(𝑠)425 <426CEU(𝑠). Accessory work left to humans is not a growing source of riches. The reason why this work427is left untouched is that we already have too many workers to do it. This lowers its value (hence the428discount) and makes it impractical to automate it.429 The significant result is that wages become fully decoupled from growth, which in the AGI430economy is entirely driven by expanding computing resources. Despite the fact that the economy431keeps expanding over time, the value of human labor stops growing and remains bounded above432by the value of compute needed to replicate it. Because wages remain bounded while the economy433continues to expand, all income eventually accrues to compute.434435Proposition 5. The share of compute in GDP converges to 1 and the share of labor in GDP to zero.436437 The share of labor in GDP can be bounded by the share of human compute in our total compu-438tational resources, measured by439440 𝑠∈𝒮 CEU(𝑠) 𝐻(𝑠)441 Í442 Share human compute𝑡 ≡ .443 𝑄𝑡 + 𝑠∈𝒮 CEU(𝑠) 𝐻(𝑠)444 Í445446447This converges to zero over time. The numbers on computational resources cited above give a sense448449 10450 𝑠 CEU(𝑠) 𝐻(𝑠), is measured in billions (population)451 Í452of magnitudes. The value of human compute,453times 1015 − 1018 flops. The value of total compute in the economy, on the other hand, could be as454high as 1054 flops, making human compute meager in comparison. These calculations show that455in an AGI economy where labor paid its compute-equivalent value, the labor share will plummet,456as total compute is projected to far exceed human compute. Conversely, most if not essentially all457income will accrue to owners of computing resources.458 The fact that the value of human labor remains capped and shrinks as a share of output does459not imply that AGI made society and workers poor in any sense. Adding up the income generated460by compute and human labor, society becomes richer (and keeps getting richer over time so long461as it expands computing resources!). Even if we leave aside the income accruing to compute, the462transition from the pre-AGI to the AGI economy necessarily makes labor income more valuable as463a whole.464465Proposition 6. The sum of all people wages is higher in the limit with AGI than in the pre-AGI economy466(with 𝑄 𝑡 = 0 and 𝛼 𝑡 (𝜔) = ∞).467468 The proposition shows that workers as a whole benefit from transitioning to the AGI economy.469Why? The arrival of AGI cannot make us worst off collectively, since we could always set up a470no-AGI zone and continue carrying our lives as if nothing had happened. This delivers at least the471same wages we had before.472 The exact argument is as follows: competitive markets imply that the economy arranges pro-473duction efficiently under AGI. There is no rearrangement that can raise output. Suppose we take4741% of human workers and send them to a non-AGI zone. These workers produce (𝑊pre-AGI · 𝐻) × 1%475in the non-AGI zone (an implication of constant-returns to scale), but we give up their wages476(𝑊post-AGI · 𝐻) × 1% in the post-AGI world. Because this rearrangement cannot raise output, the cost477(𝑊post-AGI · 𝐻) × 1% exceeds the gains (𝑊pre-AGI · 𝐻) × 1% and wages in the post-AGI world are higher478than those in the pre-AGI one.479 There are limits to the above argument. If the AGI economy uses some finite resource that is480also needed by human labor or there are diminishing returns to scale, the above argument breaks481down. Still, the above logic shows that the scope for AGI to make human workers collectively less482valuable is limited. The “problem” with AGI is that human labor seizes to be a source of growing483riches, not that it looses value.484485 11486 1.2 Transition to AGI [Preliminary]487488In general, the transition from the current economy to the AGI limit is hard to trace and depends489on how fast compute advances (𝑄 𝑡 ) relative to progress in AGI (the 𝛼’s).490 To simplify the exposition, assume all work is a bottleneck and is eventually automated. Assume491also that 𝛼 𝑡 (𝜔) jumps from infinity to its limit value 𝛼(𝜔) at some finite time 𝑇(𝜔).492 Two polar cases stand out. On one extreme we have a transition where technology binds and493work 𝜔 is automated at 𝑇(𝜔), the exact moment it becomes feasible to do so. This means compute494is abundant and ready to be used in any bottleneck tasks that can be automated. On the other495extreme, we have a transition where compute binds and work 𝜔 is only automated at a future date496˜497𝑇(𝜔), once the economy acquires enough compute to justify this application.498499Proposition 7. Suppose compute binds along the transition. The marginal value of compute 𝐴𝑡 decreases500and converges monotonically to 𝐴 > 0. Moreover, for every skill 𝑠:501502 1. The work in Ω(𝑠) is automated sequentially, from least to computational to highest computational503 requirements.504505 2. Wages 𝑊𝑡 (𝑠) decreases gradually around ˜506 𝑇(𝜔) at the same rate as 𝐴𝑡 for an interval of time and then507 continue growing until the next date is reached.508509 3. At the time work in arg max𝜔∈Ω 𝛼(𝜔) is finally automated, wages 𝑊𝑡 (𝑠) must equal 𝐴𝑡 CEU(𝑠) and510 decrease gradually from there on at the same rate as 𝐴𝑡 .511512As a whole, total wages weakly increase during the transition.513514 When compute binds, the transition to the AGI limit is smooth and gradual. For every skill515𝑠, wages are in general growing over time. This pattern is punctuated by short-lived episodes516around the time some of the work performed is automated. During these events, workers gradually517reallocate away from the automated work, which is eventually completely taken over by AGI. This518process continues until only work in arg max𝜔∈Ω 𝛼(𝜔) remains. Once this work is finally automated,519workers are stuck here and their wages are pinned down by the value of compute they save. From520there on, their wages decrease at the same rate as 𝐴𝑡 until they converge to the values in Proposition5214. On net, this process generates a growing path for total wages, and no sharp declines or “jumps”522in the wage of workers.523524 12525 Proposition 8. Suppose technology binds along the transition. The marginal value of compute 𝐴𝑡 may be526non-monotonic and jump up at points during the transition before converging to 𝐴. Moreover, for every skill527𝑠:528529 1. The work in Ω(𝑠) is automated at 𝑇(𝜔).530531 2. Wages 𝑊𝑡 (𝑠) may jump down at 𝑇(𝜔) for some of the 𝜔 ∈ Ω(𝑠) if displacement effects dominate.532533 3. Wages 𝑊𝑡 (𝑠) necessarily jump down when the last work in Ω(𝑠) is automated534535 4. Wages may exceed 𝐴𝑡 · CEU(𝑠) at various points in the transition.536537As a whole, total wages may jump down at points during the transition and will otherwise increase.538539 When technology binds, the transition to the AGI limit is jagged, uncertain, and haphazard.540The proposition describes a world where compute is abundant and ready to be used the second an541application arrives. Technology firms are experimenting in the back with applications of AI. The542moment an application pans out, the labor market is disrupted, causing the immediate displacement543of some of the people working at the automated tasks. Those who happen to be employed in544applications that prove harder to automate, on the other hand, will see rising wages, far exceeding545what they will earn in the AGI limit. The flip-side is the looming realization that at some future546date, this high wages will evaporate, as this is a premium for the random fact that the work they do547just happened to be the last experiment to pan out.5485495505512 Scientific Work552553Above I considered an economy where computers and people carry out the work needed for554production. Let’s now expand the analysis to account for scientific work. This type of work does555not deliver output, but expands our knowledge on how to produce more efficiently in the future.556 The quantity of output produced is now557558559 𝑌𝑡 = 𝑍𝑡 𝐹({𝑋𝑡 (𝜔)}𝜔∈Ω ),560561562where 𝑍𝑡 is the level of technological sophistication of the economy, improved by completing563564565 13566 scientific work. As in semi-endogenous growth models, this evolves according to567568 𝑍¤ 𝑡 −𝛽569 = 𝑍𝑡 𝐺({𝑋𝑡 (𝜎)}𝜎∈Ξ ),570 𝑍𝑡571572where 𝜎 denotes scientific work and Ξ is the set of all such work. The function 𝐺 is increasing in all573entries, differentiable, concave, and exhibits constant returns to scale. The elasticity 𝛽 > 0 captures574fishing-out effects—ideas get harder to find the further we advance.575 The quantity of scientific work accomplished in this economy is576577 1578 𝑋𝑡 (𝜎) = 𝐿𝑡 (𝜎) + 𝑄 𝑡 (𝜎).579 𝛼 𝑡 (𝜎)580581Scientific work must be accomplished by scientists, whose supply is 𝐻(𝑠) for some skills 𝑠 ∈ 𝒮𝑅 that582can be used in a subset of scientific work Ξ(𝑠). This implies583584 Õ585 𝐿𝑡 (𝜎) ≤ 𝐻(𝑠) for all 𝑠 ∈ 𝒮𝑅 .586 𝜎∈Ξ(𝑠)587588589On the other hand, completing scientific work with AI consumes computational resources. The590computational resource constraint of the economy is now591592 Õ Õ593 𝑄 𝑡 (𝜎) + 𝑄 𝑡 (𝜔) ≤ 𝑄 𝑡 .594 𝜎∈Ξ 𝜔∈Ω595596597 The set of skills used for scientific work are assumed different from those used for production,598denoted by 𝒮𝑃 .599 To simplify the exposition, assume all scientific and production work are bottlenecks.600 The two premises above are now strengthened as follows:601602Premise 1′ (AGI for science). For all work 𝜔 ∈ Ω and scientific work 𝜎 ∈ Ξ, 𝛼 𝑡 (𝜔) and 𝛼 𝑡 (𝜎) converge to603some finite values 𝛼(𝜔), 𝛼(𝜎) ∈ (0, ∞) over time.604605 I also strengthen Premise 2 to:606607Premise 1′ (Exponential compute). 𝑄 𝑡 grows exponentially at rate 𝑔𝑄 > 0.608609 The next proposition characterizes the limit behavior of the economy. To highlight key forces,610611 14612 I assume a constant fraction 𝜆 of compute is allocated to science. Naturally, there is some optimal613value for 𝜆 that depends on how society discounts future consumption flows.614615Proposition 9. Suppose fraction 𝜆 ∈ (0, 1) of compute allocated to science. All production and scientific616bottlenecks are automated and output converges to617618 Õ619 𝑌𝑡 = 𝑍𝑡 𝐴 (1 − 𝜆) 𝑄 𝑡 + CEU(𝑠) 𝐻(𝑠) ,620 621 (3)622 𝑠∈𝒮𝑅623624625where626627 Õ 1/𝛽628 𝑍𝑡 ∝ 𝐵 𝜆 𝑄𝑡 + CEU(𝑠) 𝐻(𝑠) . (4)629 𝑠∈𝒮𝑃630631632 As before, aggregate production possibilities in the economy become additive in compute and633labor, with labor expressed in units of compute.634 More novel, the stock of productive knowledge 𝑍(𝑡) expands over time and scales with com-635puting resources. In standard semi-endogenous growth models, 𝑍(𝑡) scales with population, since636scientific work requires human brains and these are limited by the size of the population. Here, by637contrast, 𝑍(𝑡) scales with the economy’s computational resources 𝑄 𝑡 because these can be used to638automate all scientific bottlenecks.639 The fact that AGI makes it feasible to complete scientific work with compute generates a com-640pounded growth effect.641642Proposition 10. The growth rate of output and technology converge to643644 1 1645 𝑔𝑌 = 𝑔𝑄 𝑔 𝑍 = 𝑔𝑄646 647 +1648 𝛽 𝛽649650 The main difference with Proposition 3 is that, in an economy without other forms of scientific651progress, output scales with compute and 𝑔𝑌 = 𝑔𝑄 . The possibility of carrying out science with652compute implies that the growth rate is now higher by 𝑔𝑄 1653 654 𝛽 . This is because some compute is655used to expand productive knowledge, which then raises the productivity of all units of compute656employed in production, generating increasing-returns in compute 𝑄 𝑡 .657 The above results clarify that, even in an AGI world where compute can be used for science,658output scales with compute. There is not necessarily an intelligence or growth explosion, as this659660 15661 requires an exploding amount of compute.662 One interesting aspect has to do with the prioritization of compute across uses.663664Proposition 11. Suppose the economy discounts future resources at a rate 𝜌 ≥ 0. In order to maximize665welfare, in the long run, the economy allocates a fraction666667 1668 𝜆∗ =669 1+𝛽+𝜌670671of compute to science and the remaining to production. Along the transition, scientific bottlenecks are672prioritized and 𝜆(𝑡) starts above 𝜆∗ , eventually converging to 𝜆∗ .673674 This result provides some guidance on how an efficient economy organized to maximize dis-675counted aggregate output would prioritize scarce compute.676 In the short run, some priority should be given to scientific work, as this allows the rapid677expansion of production knowledge, which makes the economy more productive at all future times.678Other considerations not accounted for here, such as minimizing labor-market disruptions along679the transition, call for even more prioritization of scientific work and a more gradual deployment680of AGI for production (see for example Lehr and Restrepo, 2022).681 In the long run, compute is allocated to both uses, and the economy eventually automates all682production and scientific bottlenecks. The reason is that this strategy achieves the maximum scaling683of output with compute. Using all compute for science yields a growth rate of684685 1686 𝑔𝑌 = 𝑔𝑄 .687 𝛽688689Using all compute for production yields a growth rate of690691692 𝑔𝑌 = 𝑔𝑄 .693694695Using compute for both types of bottlenecks yields a higher growth rate because it exploits the696increasing-returns to scale synergies between production and research introduced by scientific697progress.698699700701702 16703 3 Discussion704705This chapter analyzed the long-run implications of AGI for production, growth, and labor markets.706The findings challenge both overly optimistic and overly pessimistic views about the future of707human labor. On one hand, human work does not become obsolete. Because compute remains708scarce, labor retains value by saving on computational resources. In addition, some accessory work709may be left untouched by AGI, offering a stable, persistent set of roles for humans. On the other710hand, while humans can still contribute economically, their role shrinks dramatically. The work711that remains for people—whether bottleneck or accessory—commands a fixed value, bounded by712the compute required to replicate it. In a growing economy, this means wages stagnate and the713labor share collapses. Even skilled workers performing essential tasks will earn only what they714save in compute—no more. In sum: humans can still work in an AGI economy—so long as they715are endowed with time and skill—but their contribution becomes economically negligible. The716economy keeps growing; we stay in place.717 An interesting implication is the potential persistence of unautomated accessory work. For718many socially intensive tasks—such as care work, hospitality, or therapy—the compute required719to emulate human warmth or social intuition may be enormous. Even if such work is technically720automatable, it may remain economically impractical to do so. As a result, these domains could721continue to offer meaningful work for people. Yet these roles, while important, are not a source of722growing income. The wages they offer are capped, possibly traded at a discount, and unresponsive723to aggregate growth. Accessory work may provide continuity and stability in a transforming724economy, but are not a source of growing riches providing human labor a central economic role.725 The transition to AGI also presents distinct challenges. If the bottleneck is technological—i.e., if726the key constraint is developing AGI systems themselves—then inequality may rise sharply during727the transition, as certain types of labor temporarily gain extraordinary value. From the viewpoint728of workers, the transition will feel uncertain and even unfair. People with some skills may see729their wages fall rapidly to their compute-equivalent levels while others, by a sheer act of luck, may730see their wages rise above it, earning a large premium just because the work they do turned out731to be among the last we figured how to automate. This feels harder to navigate than a transition732where compute binds, and wage growth is punctuated by gradual declines and the gradual need733to reallocate labor. One crucial question here is how to help workers navigate a transition where734735736 17737 technology binds. How can we help them share risks? Would other policies such as deploying the738technology gradually help here (see also Lehr and Restrepo, 2022)?739 As labor ceases to be the primary driver of value, economic policy must confront a basic question:740how can we share the income generated by compute? In a world where AGI performs all bottleneck741work, income generated by production flows to those who own or control computational resources.742One approach is to redistribute the gains from compute through universal dividends. An alternative743is to reimagine compute as a public or semi-public resource, akin to land or natural capital, with744returns broadly shared.745 Looking beyond wages and output, AGI also raises deeper questions about meaning and pur-746pose. Historically, labor supplied the bottlenecks: it powered economic growth and gave individuals747a central role in production. That connection is severed in the world of AGI. Human work no longer748drives progress; it is no longer needed to improve living standards. If tomorrow half the population749stopped working, no one would notice. In the AGI economy, we won’t be missed.750 Would people still choose to work, even when it makes no difference? Would work remain751meaningful when it ceases to be economically essential? Or would we stop working altogether,752not because we were replaced, but because we chose to step aside—seeking fulfillment in a world753where our skills no longer matter?754755756757758 18759 References760761Acemoglu, Daron, and David H. Autor. 2011. “Skills, Tasks and Technologies: Implications for762 Employment and Earnings.” In Handbook of Labor Economics. Vol. 4, , ed. Orley Ashenfelter and763 David Card, 1043–1171. Amsterdam:Elsevier.764765Acemoglu, Daron, and Pascual Restrepo. 2018. “The Race between Man and Machine: Implica-766 tions of Technology for Growth, Factor Shares, and Employment.” American Economic Review,767 108(6): 1488–1542.768769Acemoglu, Daron, and Pascual Restrepo. 2019. “Artificial Intelligence, Automation, and Work.”770 In The Economics of Artificial Intelligence: An Agenda. , ed. Ajay Agrawal, Joshua Gans and Avi771 Goldfarb, 197–236. Chicago, IL:University of Chicago Press.772773Aghion, Philippe, Benjamin F. Jones, and Charles I. Jones. 2019. “Artificial Intelligence and774 Economic Growth.” In The Economics of Artificial Intelligence: An Agenda. , ed. Ajay Agrawal,775 Joshua Gans and Avi Goldfarb, 237–282. Chicago, IL:University of Chicago Press.776777AI Impacts. 2023. “How much computing capacity exists in GPUs and TPUs in778 Q1 2023?” https://wiki.aiimpacts.org/doku.php?id=ai_timelines:hardware_and_ai_779 timelines:computing_capacity_of_all_gpus_and_tpus, Accessed: 2025-07-01.780781Amodei, Dario. 2024. “Machines of Loving Grace.” Essay on Anthropic CEO’s website, https://www.782 darioamodei.com/essay/machines-of-loving-grace.783784Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of Recent Techno-785 logical Change: An Empirical Exploration.” Quarterly Journal of Economics, 118(4): 1279–1333.786787Bostrom, Nick. 2003. “Are You Living in a Computer Simulation?” https://www.788 simulation-argument.com/, Accessed: 2025-07-01.789790Bostrom, Nick. 2014. Superintelligence: Paths, Dangers, Strategies. Oxford University Press.791792Drago, Luke, and Rudolf Laine. 2025. “The Intelligence Curse: A Map to Maintaining Human793 Economic Relevance in the Age of AGI.” Essay series at intelligence-curse.ai, Accessed July794 2025.795796 19797 Feenstra, Robert C., and Alan M. Taylor. 2017. International Trade. . 4th ed., New York:Macmillan798 Learning.799800Hilbert, Martin, and Priscila López. 2011. “The World’s Technological Capacity to Store, Commu-801 nicate, and Compute Information.” Science, 332(6025): 60–65.802803Jones, Charles I. 1995. “R&D-Based Models of Economic Growth.” Journal of Political Economy,804 103(4): 759–784.805806Kokotajlo, Daniel, Scott Alexander, Thomas Larsen, et al. 2025. “AI 2027: A Detailed, Falsifiable807 Forecast of the Next 2–3 Years.” AI Futures Project blog, Accessed 2025-07-04.808809Korinek, Anton, and Donghyun Suh. 2024. “Scenarios for the Transition to AGI.” arXiv preprint,810 2403.12107.811812Kortum, Samuel. 1997. “Research, Patenting, and Technological Change.” Econometrica, 65(6): 1389–813 1419.814815Lehr, Nils H., and Pascual Restrepo. 2022. “Optimal Gradualism.” National Bureau of Economic816 Research NBER Working Paper 30755.817818Open Philantropy. 2020. “How Much Computational Power Does It Take to819 Match the Human Brain?” https://www.openphilanthropy.org/research/820 how-much-computational-power-does-it-take-to-match-the-human-brain/, Accessed:821 2025-07-01.822823Sandberg, Anders, and Nick Bostrom. 2008. “Whole Brain Emulation: A Roadmap.” Future of824 Humanity Institute, University of Oxford Technical Report 2008-3.825826Segerstrom, Paul S. 1998. “Endogenous Growth without Scale Effects.” American Economic Review,827 88(5): 1290–1310.828829Trammell, Philip, and Anton Korinek. 2023. “Economic Growth under Transformative AI.” Na-830 tional Bureau of Economic Research NBER Working Paper 31815.831832833834835 20836