The AI Productivity Paradox: Too Early to Tell

Capital expenditure on AI infrastructure — data centres, GPUs, power capacity — has grown at a pace that would ordinarily be expected to show up in aggregate productivity statistics within a few years. So far, the productivity data in most advanced economies has not moved decisively. This gap has revived comparisons to Robert Solow's famous 1987 observation: "You can see the computer age everywhere but in the productivity statistics."

The historical pattern of general-purpose technologies

Economic historians who study general-purpose technologies — electricity, the internal combustion engine, computing — have documented a consistent lag between when a technology is invented and when it materially lifts productivity statistics. Electrification of US factories, for example, took several decades to show up meaningfully in output-per-worker data, because the gains required not just installing the new technology but redesigning entire workflows and organisational structures around it. Simply replacing a steam engine with an electric motor in the same factory layout produced almost no efficiency gain — the productivity boost came only once factories were rebuilt from scratch around distributed electric power.

Why AI may follow a similar path

• Complementary investment — training data pipelines, workflow redesign, and organisational change — often lags the initial capital spending by years.
• Measured productivity statistics are notoriously bad at capturing quality improvements and time savings in knowledge work, which is where a large share of near-term AI gains are concentrated.
• Diffusion across the broader economy — beyond the technology sector itself — historically takes longer than the initial adoption curve within tech-native firms.

None of this proves that AI will eventually deliver the productivity gains its proponents expect — that remains an open empirical question. But it does mean that the absence of a productivity inflection in the data so far is weak evidence against the AI investment thesis, given how consistently prior general-purpose technologies have shown the same multi-year lag. The more informative signal to watch is not aggregate GDP-per-hour statistics in the near term, but firm-level case studies of workflow redesign — which tend to lead the aggregate data by several years.