Does the Digital Economy Have a New Business Cycle? The Evidence Remains Tentative

The digital economy may need different economic indicators. That does not mean researchers have already identified a new economic rhythm. This distinction runs through The Digital-Energy Recomposition of the Kitchin Cycle: Exploratory Evidence from U.S. IT and Irish Data-Centre Series (1995–2025), an exploratory arXiv preprint.

The researchers ask whether a short business cycle traditionally associated with inventories is being reorganised around computing capacity, semiconductors and energy. Their evidence identifies changes in statistical relationships, but does not establish that electricity consumption can forecast growth or that a coherent digital-energy cycle has replaced the old one.

A change that predates the designated AI era

The Kitchin cycle describes fluctuations over roughly three to five years, traditionally linked to the accumulation and liquidation of inventories. The paper’s proposed update treats computing capacity and energy commitments as possible counterparts to physical stockpiles.

To explore this idea, the researchers analyse quarterly U.S. data from FRED and Irish data from the Central Statistics Office. Their headline U.S. sample covers 1995–2025; Irish metered data-centre electricity observations cover 2015–2024. They examine quarter-to-quarter logarithmic changes rather than simply comparing rising levels.

One important result concerns timing. A statistical search for a change in the U.S. IT investment–GDP correlation selects the third quarter of 2014. The reported correlation falls from about +0.45 before that boundary to −0.12 afterwards. The structural-break statistic exceeds the threshold the researchers use for significance.

A separate comparison of the correlations before and after this selected date also survives the paper’s Bonferroni multiple-testing adjustment, with an adjusted p-value of 0.006. However, the researchers caution that this is not independent confirmation: the boundary was chosen from the same data.

By contrast, comparing their imposed Cloud Expansion and AI Era windows produces an adjusted p-value of 0.164 for IT investment, which does not meet the conventional 5% threshold. Their AI Era begins in 2021. The defensible conclusion is therefore narrower than an AI-driven rupture: the reported weakening of the IT–GDP relationship began years earlier. The paper calls this “Strategic Decoupling”; the correlation change alone does not establish firms’ motives.

Ireland’s result depends on what output means

Ireland supplies a more direct physical measure: electricity consumed by data centres. In the 2021–2024 window, its quarterly growth has a correlation of approximately +0.41 with real GDP growth. But this estimate uses just 16 paired observations and is not statistically significant at the conventional 5% level.

More importantly, replacing real GDP with real GNP reduces the correlation to approximately zero. Using real gross value added instead produces a result similar to GDP.

The researchers interpret this contrast as evidence that multinational activity helps explain the apparent electricity–output connection. GNP accounts for net factor income flows with the rest of the world, making it a different lens on an economy whose headline GDP is heavily influenced by multinational operations. The result cautions against reading the GDP correlation as evidence of a domestic economic cycle; it does not establish a causal tax mechanism. The preferred modified national-income measure, GNI*, is available only annually and cannot supply the quarterly comparison used here.

The timing analysis also fails to establish electricity demand as an advance signal. Across the full Irish sample, the largest positive correlation occurs when GDP precedes data-centre electricity consumption by four quarters. None of the nine ordinary cross-lag correlations is statistically significant, and the paper’s predictive tests do not survive their multiple-testing adjustment. A coincident or lagging interpretation fits the point estimates better than a forecasting claim, but remains tentative.

A new cycle still needs a rhythm

Changing correlations are not enough to demonstrate a recurring three-to-five-year cycle. The paper therefore also examines which frequencies contribute most strongly to fluctuations.

Neither U.S. nor Irish GDP has its three largest reported spectral peaks in the Kitchin range. U.S. semiconductor production does have a four-year peak, but it is the second-largest, not the dominant one. Irish data-centre electricity instead shows peaks around annual periods, which the researchers associate with seasonality.

These findings temper the proposed digital-energy framework. Semiconductor results also face unresolved statistical concerns: stationarity tests give conflicting signals even after transformation. Short samples, approximate inference assumptions and pandemic-related disruptions further limit interpretation.

The useful contribution is consequently a measurement warning, not a forecasting dashboard ready for use. Longer samples, comparisons across data-centre hubs and economic measures less distorted by multinational accounting are needed before contracted electricity or computing capacity can credibly be treated as replacements for traditional inventory indicators.