Why would compute-replacement cost determine wages instead of productivity?
This explores the claim that once AI can do a job, what an employer pays a human stops tracking the value that human creates and starts tracking what it would cost to run a machine doing the same thing.
This explores the argument that in an AI-saturated economy, your wage is set not by how much value your work produces but by the price of the cheapest thing that could replace you — and that cheapest thing is compute. The core of this idea comes from What happens to human wages in an AGI economy?, which argues that as AI automates the highest-value bottleneck tasks first, human pay drifts away from reflecting economic contribution and toward reflecting the cost of buying enough GPU time to replicate that contribution. Labor's share of the economy shrinks toward zero not because humans become worthless, but because their wage gets anchored to a falling external benchmark.
The reason this substitution holds is that in a competitive market, a price floats down to the marginal cost of the next-cheapest supplier. If a task can be done by either a person or a model, an employer won't pay the person more than it costs to run the model — regardless of how productive the person is. Productivity used to set the price because there was no substitute; once compute becomes a substitute, productivity sets the ceiling on value created, but compute cost sets the wage. The two decouple.
What makes compute a genuine substitute (rather than a fixed alternative) is that its cost is fluid and falling. Several notes show compute is not one fixed resource but a budget that can be reshuffled to get more for less: smaller models with extra inference time match bigger ones on hard problems Can inference compute replace scaling up model size?, and spending that inference budget adaptively — little on easy tasks, lots on hard ones — beats spending it uniformly Can we allocate inference compute based on prompt difficulty?. There are even cheaper tricks lurking, like recomputing weights instead of moving them on constrained hardware Does recomputing weights cost less than moving them on mobile?, and persistent agents that amortize cost across reused context so the real denominator becomes finished work, not tokens Do persistent agents really cost less per token?. Each of these pushes the replacement price down — and the wage benchmark down with it.
There's a deeper framing worth pulling in: this is what it looks like when intelligence stops behaving like a commodity. Two notes argue AI output isn't a fixed, possessable object but a contextual flow — a token valued by what it does for the receiver, not what it is Does AI actually commodify expertise or tokenize it?, Is AI fundamentally changing how value gets produced?. When the thing you're selling is a flow generated on demand, its price collapses toward the marginal cost of generating the next unit — which is compute. The wage-equals-compute story is the labor-market version of that same collapse.
The corpus also pushes back on treating this as inevitable. Whether AI displaces or merely reshuffles labor depends on how exposure is distributed: concentrated exposure lets workers shift to non-displaced tasks and offsets job losses Does concentrated AI exposure enable workers to adapt and reallocate?, and the inequality outcomes hinge on access and incentives, not the technology itself Does generative AI inevitably worsen or reduce inequality?. There's also a productivity wrinkle: AI's gains show up when people apply skills they already have, and vanish when they try to learn new ones When does AI actually boost worker productivity? — which hints that the human work resisting compute-replacement is exactly the work where value and substitutability haven't yet decoupled.
Sources 10 notes
As AGI automates bottleneck work first, human wages shift from reflecting economic value to reflecting compute costs. Labor's share of GDP approaches zero even as some accessory work remains human, driven by compute-allocation efficiency rather than irreplaceability.
Snell et al. (2024) showed that inference-time compute trades off against model parameter scaling, especially on difficult prompts. This reveals pretraining and inference compute are not independent resources.
Research shows inference effectiveness varies dramatically by prompt difficulty. Reallocating the same total compute adaptively—giving easy prompts less and hard ones more—substantially outperforms larger models under uniform budgets.
MobileLLM shows that on memory-bound mobile hardware, sharing weights between adjacent transformer blocks by recomputing one block twice uses less latency than fetching separate weights, gaining accuracy with no parameter increase.
A 115-day case study found 82.9% of tokens were cache reads. When context persists and reuses, the meaningful cost denominator becomes completed artifacts, not individual tokens.
AI output lacks the fixed, identical, possessable properties of commodities. Instead it functions like tokens—mutable mediums of exchange valued by what they do for receivers, not what they are.
AI production is organized around contextual token-flows generated at point of use, not identical mass-produced objects. This creates different effects than commodification: inflationary devaluation, contextual variation, and skill transformation from production to validation.
Analysis of task-level AI exposure across firms 2010-2023 shows that while higher mean exposure reduces labor demand, more concentrated exposure (affecting few tasks) enables workers to reallocate to non-displaced tasks, producing modest net employment effects.
An interdisciplinary review found that across information, work, education, and healthcare, generative AI can both exacerbate and reduce inequality. The direction is determined by access, integration, and incentive structures, not the capability itself.
Studies showing AI productivity gains measured tasks within workers' existing domains. When workers used AI to learn new skills, productivity gains disappeared and learning suffered, suggesting prior findings do not generalize to skill acquisition.