Acemoglu’s <0.71% cumulative TFP ceiling
The data
Upper-bound 10-year US Total Factor Productivity gain from generative AI: 0.55–0.71% cumulative. Upper-bound 10-year US GDP gain: 0.9–1.6% cumulative.
Derived via Hulten’s theorem from a task-decomposition of the US economy. For comparison, optimistic forecasts (McKinsey, Goldman Sachs Economists) put 10-year US productivity gains near 9% cumulative — >10× higher.
Adjusted for the easy-to-learn / hard-to-learn task distinction (hard tasks resist current AI architectures), the realistic ceiling pulls down toward <0.55% cumulative.
Source
Daron Acemoglu, The Simple Macroeconomics of AI, NBER Working Paper 32487, May 2024. Surfaced through: acemoglu-simple-macroeconomics and synthesized in bear-case-synthesis.
What it shows
A formal lower-bound on the macro impact of generative AI under standard labor-economics assumptions. Even if every task-level AI demonstration is real, the aggregate productivity gain is bounded by (fraction of tasks impacted) × (average task-level cost saving) — which the paper estimates as ≈ 4.6% × 15.4%.
It is the single bear-case number most in direct tension with wef-7-6t-projection. Both cannot simultaneously be right.
How to cite
Acemoglu (NBER WP 32487, 2024) bounds the 10-year US TFP gain from generative AI at 0.55–0.71% cumulative — about an order of magnitude below McKinsey / Goldman Sachs forecasts.
Connections
- task-based-framework — the concept page where the derivation lives; companion table at ai-task-exposure-decomposition.
- ai-task-exposure-decomposition — the underlying 19.9% / 23% / 15.4% / 4.6% decomposition.
- capital-labor-divergence — same Acemoglu paper supplies the distributional story.
- bear-case-synthesis — pillar 1 of the four-pillar bear case.
- wef-7-6t-projection — the direct quantitative tension.
- H2_u-curve-of-value — if the macro ceiling is this low, value-capture cannot be widely distributed across the stack.
- scaling-wall — the architectural argument for why hard-to-learn tasks stay hard.