280-Fold AI Cost Crash: Why Wages Can Hold Their Floor

The cost of running a GPT-3.5-level AI system collapsed 280-fold between 2022 and 2024, according to the Stanford HAI 2025 AI Index Report. That single number pushes the AI debate into macro territory: if digital labour is nearly free, human wages collapse and the labour share of GDP evaporates. That is the naive read. A better market read requires a framework that David Ricardo formalised in 1817: comparative advantage. Even when one agent holds an absolute edge in every task, both can gain from trade if each specialises in the activity where its opportunity cost is lowest. For humans, the binding constraint is time. For AI, the constraint is compute. The division of labour that follows is far more nuanced, and the macro transmission through wages, consumption, and the Fed's reaction function depends on exactly where the boundary between cheap compute and expensive human judgement gets drawn.

The 280-Fold AI Cost Decline That Rewrites the Wage Equation

The Stanford HAI 2025 Data Point

The Stanford HAI 2025 AI Index Report documented a staggering re‑pricing of digital labour. A model that cost hundreds of thousands of dollars to run only two years ago now costs a few hundred. For economists, this is a supply‑side shock to the price of a specific input. When an input gets radically cheaper, it triggers substitution across the economy: activities that previously made no sense to automate suddenly become viable. That is why the AI‑labour debate has moved from factory floors to radiology suites and legal briefs.

The Ricardian Read: Absolute Advantage Without Comparative Advantage

The panic around AI routinely points to absolute advantages. Large language models write clearly, summarise documents quickly, and generate passable code in seconds. In these discrete tasks, AI is a direct competitor. If a job were nothing but a collection of such tasks, the human worker would be in trouble. The Ricardian challenge, however, is to identify where AI holds a comparative advantage – and whether that advantage operates at the job level or only inside narrower tasks. Comparative advantage is determined by opportunity costs. For humans the binding constraint is time; for AI the constraint is compute. Those are very different inputs, and the difference is large enough to keep humans firmly in the picture.

From Cheaper AI to a Lower Wage Floor: The Restrepo Convergence Model

How Wages Track Compute Costs

Restrepo (2025) formalised what happens when the price of compute that can replicate a skill keeps declining. In his model, wages converge to the cost of the compute needed to reproduce the worker's output. If a paralegal task that cost $40 in human time can be performed by an API call that costs $0.01, the wage for that task drifts toward $0.01, not $40. That is a direct transmission channel from AI hardware deflation to the labour share of national income. When the digital worker is nearly free, the bargaining power of the human worker evaporates, and the returns to capital capture an ever‑larger slice of GDP.

The Consumer Spending Channel and the Fed's Rate Path

The labour share of output has been under pressure for decades from globalisation and automation. AI adds a new layer because it attacks cognitive routine – the one area where many advanced‑economy workers still earned a premium. If the Restrepo convergence plays out, the consequences for consumer spending are large. Households that receive a declining share of income cannot sustain the same rate of consumption growth. That would act as a persistent drag on top‑line revenue growth for consumer cyclicals, potentially offsetting margin gains that firms capture from the same AI‑driven cost cuts.

This wage‑suppression channel is not yet visible in aggregate data; average hourly earnings in the US are still running above pre‑pandemic trend. The expansion of AI tools could change that. The 341,000‑job gap between payrolls and the household survey, first analysed by AlphaScala, underscores the uncertain state of the labour market. Strong headline payrolls can mask a substitution of quality for quantity, where full‑time roles give way to part‑time or gig work that AI tools make more efficient. If wage growth begins to bend toward the falling cost of digital substitutes, the rate‑cut thesis gains a powerful new ally. That would align with the view expressed by Kantrowitz, who sees rate cuts in three months despite hot inflation, as disinflationary forces from AI could eventually dominate.

The Radiologist Case: Where AI Competes, Where It Amplifies

The Boundaries of Cheap Compute: When the Transmission Breaks

Physical and Economic Limits to the 280-Fold Decline

The 280‑fold cost decline is spectacular. Extrapolating it forward indefinitely is dangerous. The Stanford HAI 2025 AI Index Report itself documents that the rate of improvement is beginning to encounter physical and economic barriers. Fabricating cutting‑edge chips requires immense capital expenditure. The energy consumption of the largest training runs is rising faster than the efficiency gains. If compute‑cost deflation slows – or stalls – the Restrepo convergence process loses its engine. The wage floor stops falling, and the labour share stabilises.

For investors, this is the key variable. The transmission from cheaper AI to lower wages is not a deterministic slide. It is a function of the pace at which compute hardware improves relative to the complexity of human judgement. When that pace slows, the comparative‑advantage boundary shifts back toward labour, and the disinflationary impulse from AI weakens. Traders should treat chip fab capacity reports, energy prices, and the next AI Index Report as first‑order macro inputs.

Practical Rule for Traders: Distinguishing AI‑Competitor Jobs from AI‑Tool Jobs

This distinction turns the comparative‑advantage lens into an investable framework. Sectors heavy in AI‑competitor tasks – back‑office administration, basic legal drafting, entry‑level programming – could see wage compression and margin pressure for labour‑intensive firms. Sectors built on AI‑tool tasks – healthcare delivery, skilled trades, negotiation‑heavy services – may benefit from productivity gains without the same wage erosion. The equity market will eventually price this divergence, and the starting point for that repricing is a clear understanding of which jobs fall on which side of the boundary.

Next Concrete Marker: The Stanford HAI AI Index Update

The macro transmission chain now rests on a single, observable trend: the future rate of decline in AI compute costs. If the next generation of hardware or algorithms delivers another order‑of‑magnitude reduction, the wage‑suppression channel intensifies. The case for structurally lower terminal rates strengthens, and the risk to consumer cyclicals grows. If, on the other hand, the physical limits of chip‑making and energy supply begin to bind, the labour share will prove more durable than the Restrepo model implies. That would support wage growth, consumer spending, and the aggregate demand component of inflation.

The narrative that AI will hollow out the workforce overnight is an absolute‑advantage story, and it is incomplete. Comparative advantage – a concept formalised in 1817 by a man studying English cloth and Portuguese wine – still tells traders more about how the macro picture will evolve than any single‑stock AI hype. The next concrete marker will be the update to the Stanford HAI AI Index, which will reveal whether the 280‑fold cost decline was a one‑time adjustment or the start of a secular collapse that genuinely rewrites the wage‑setting equation. Until then, the safe trade is to watch where the machine's opportunity cost is high – and bet that human judgement still earns a premium exactly there.