AI capability gains alone do not guarantee mass substitution of workers: the economics of substitution hinges on inference energy costs, integration amortisation and non-eliminable supervision. Energy markets, compute supply chains, regulatory regimes and capital-market patience — not model performance in isolation — determine where and when AI displaces tasks.

Main Finding

The paper argues that contemporary AI–labour debates exhibit "capability determinism": observers infer large-scale labour substitution directly from model demonstrations without doing the firm-level unit-cost calculation that actually determines substitution. Reframing substitution as a unit-cost test shows AI must clear three cost lines — (1) variable inference cost (dominated by electricity and compute supply chains), (2) amortised integration cost, and (3) residual supervision cost — and that energy markets, supply-chain geopolitics, integration economics, regulatory regimes, and capital-market patience largely decide whether and where substitution happens. As a result, substitution will be selective, configurational, and materially more modest and uneven than capability-driven narratives imply.

Key Points

• Capability determinism: The dominant discourse slides from demonstrations of model capability to claims about large-scale job substitution and then to distributional policy responses, skipping the economic test firms actually perform (fully loaded unit-cost per task).
• Substitution as accounting: Substitution is fundamentally a firm-level, per-unit cost comparison between human labour and machine-per-task costs (task ≠ job; tasks within jobs can be substituted unevenly).
• Three cost components on the AI side:
  • Variable inference cost: each model run (inference) consumes electricity and chip cycles; cost per token × tokens per task is the variable marginal cost that scales with use.
  • Amortised integration cost: fixed and semi-fixed costs (building data centres, chips, staff, software integration, insurance, grid connections) that must be spread across volume; high utilisation is required to drive these per-task down.
  • Residual supervision cost: non‑zero ongoing human oversight, legal/professional liabilities, and regulation that limit how low supervision costs can fall.
• Electricity is the binding constraint: (a) large-scale inference runs in megawatt to gigawatt facilities; (b) cooling and ancillary systems add to energy use (typical PUE ≈ 1.3); (c) electricity price sets a floor for long-run cost per inference; and (d) regional grid constraints and geopolitical energy shocks push electricity costs up.
• Current API prices are often subsidised: frontier providers are frequently running losses on inference (capital-funded), so published per-token/API prices understate the long-run economic price firms face if capital markets demand profitability.
• Technical improvements can increase per-task token usage: better reasoning and agentic/multi-step workflows (and retrieval augmentation) often require many more tokens per task, raising variable costs even as capability rises.
• External variables decide the clearing of substitution: energy regulation, compute supply-chain geopolitics (chips, rare materials, fabrication), professional regulation, integration costs, and capital-market patience determine when/where substitution is economically viable.
• Empirical implication: substitution will be uneven across locations and tasks; policy should focus on the conditions that enable substitution (energy, supply, integration, regulation), not only on downstream redistribution.

Data & Methods

• Methodology: conceptual and theoretical analysis combining:
  • Sociology/history of technology (technological determinism literature; Wyatt 2008; Heilbroner, Winner).
  • Operational accounting / unit-cost economics (fully loaded cost comparisons; Drury; Horngren).
  • Task-based labour economics framing (Acemoglu & Restrepo; Frey & Osborne).
• Evidence base: synthesis of industry reports, multilateral agency data, technical AI literature, and market analyses rather than novel primary microdata:
  • Energy and data-centre metrics (IEA reports; data-centre electricity growth ≈ 12%/year since 2017; PUE ~1.3).
  • Cost facts about training vs inference (training = large one-off spend; inference repeated per use).
  • Hardware/supply-chain dependencies (Nvidia H100/H200, TSMC fabrication; reliance on concentrated material sources).
  • Market analyses of provider economics (cited Epoch AI analysis indicating gross margins ≈ 30% on inference but overall operating losses).
  • Observations about token counts: modern reasoning/agentic workflows can increase tokens per task by an order of magnitude.
• Limitations: not an econometric or empirical causal study; no micro-level firm data on true per-task fully loaded costs is presented. The paper reframes and reinterprets existing quantitative facts to produce a prescriptive analytic lens.

Implications for AI Economics

• Research agenda shifts:
  • Move from capability-centric exposure scores to rigorous estimates of fully loaded per-task costs (include electricity, cooling, PUE, amortised capex, integration and supervision).
  • Empirically measure tokens-per-task for real-world workflows (including retrieval, agentic loops, prompt engineering) and map how token intensity changes with model capability.
  • Model geographic heterogeneity: grid capacity, electricity prices, local regulation, and compute supply-chain exposure should be central variables in predicting local substitution.
  • Study capital-market dynamics: the role of subsidised inference pricing and the conditions under which markets demand economically sustainable prices.
• Policy implications:
  • Energy policy and grid planning are labour-policy levers: constraints or subsidies in electricity supply materially affect AI substitution viability.
  • Supply-chain and geopolitics matter: compute resilience, chip fabrication policy, and trade in critical materials influence the cost floor for inference.
  • Regulation and professional liability set a non‑zero floor on supervision costs; rules around liability, auditing, and certification will change substitution incentives.
  • Integration costs and firm adoption dynamics: policies that lower integration friction (standards, interoperability, public data infrastructure) can accelerate substitution in some sectors; conversely, high integration costs slow it.
• Labour-market expectations:
  • Expect selective, configurational substitution rather than wholesale displacement. Tasks that are token-cheap, highly standardised, and deployable in regions with cheap reliable electricity and cheap integration will be replaced first.
  • Distributional policy should be targeted and geographically informed, not premised on uniform, rapid mass displacement.
• Practical takeaways for economists and policymakers:
  • Include energy-price and grid-capacity scenarios in models forecasting automation.
  • Require transparency on provider unit economics when using API-price-based exposure estimates.
  • Prioritise obtaining firm-level data on utilisation rates, amortisation schedules, supervision requirements, and the true marginal cost per task to improve forecasts and policy design.