Faster obsolescence of computing capital can explain the 1990s productivity surge and the long-run slowdown: a calibrated growth model shows a temporary productivity boom driven by skill accumulation, then a permanent lower growth path and falling labor share as short-lived designs make capital scarcer and more profitable. BEA-based decomposition reveals an increase in inferred obsolescence beginning in the mid-1990s consistent with the model's timing, though the paper relies on calibration and pattern-matching rather than causal empirical identification.

Main Finding

An exogenous, sustained increase in the obsolescence rate of capital designs (driven by the spread of short-lived computing equipment and software) can generate a mid-1990s–style productivity boom followed by a persistent slowdown, together with a declining labor income share. The mechanism: faster obsolescence reduces the discounted value of future design payoffs, eliciting a short-run reallocation toward skill accumulation and production (a temporary productivity surge), then a persistent shift of effort toward producing short‑lived capital designs that makes efficient capital relatively scarce, raises capital’s flow profits, lowers the labor share, and slows long‑run growth.

Key Points

• Empirical motivation

  • BEA asset-level investment and service‑life data imply an increase in aggregate economic depreciation (interpreted largely as technological obsolescence) beginning in the mid‑1990s. Excluding computing equipment and software removes the trend.
  • The timing coincides with a U.S. productivity boom in the mid‑1990s followed by a prolonged productivity slowdown and a decline in the labor share.

• Theoretical mechanism (model)

  • Framework: Acemoglu (2003)–style endogenous growth model with directed technical change. Final output uses two complementary aggregates: efficient labor (raw labor × skills) and efficient capital (raw capital × designs).
  • Only capital designs are subject to exogenous obsolescence at rate δ_obs; new skills and designs are produced by innovation using a common pool of labor/effort.
  • Free entry and HJB valuation determine the present value of a new skill or design; obsolescence raises the effective discounting of design profits.
  • Short-run response to a permanent rise in δ_obs:
    • Value of future design profits drops → immediate weakness in capital innovation incentives.
    • Resources shift toward production and skill creation → temporary productivity boom (faster skill accumulation).
  • Transition / medium/long run:
    • Efficient capital becomes scarcer relative to efficient labor; under complementarity this raises capital’s flow profits vs. skills → labor share falls.
    • Equilibrium reallocates more effort to (now more immediately profitable) capital design creation despite shorter lives, leaving fewer durable designs and slower long‑run labor‑augmenting growth.
  • Balanced growth path remains labor‑augmenting growth (design stock asymptotically offsets obsolescence rather than expanding).

• Quantitative results and fit

  • The calibrated model, with a mid‑1990s permanent increase in obsolescence, reproduces qualitatively (and roughly quantitatively) the observed pattern: a roughly decade‑long productivity boom followed by slower long‑run growth, a persistent decline in the labor share beginning in the mid‑1990s, and a decline in measured capital efficiency.
  • The model also produces the feature often noted in data: rising innovative effort but declining productivity returns to that effort.

• Robustness and caveats

  • Results are robust across a broad set of parameter variations (elasticities, markups, etc.) reported in sensitivity analyses.
  • Key modeling abstractions: obsolescence treated as exogenous and uniform across designs; symmetric firms; physical depreciation set to zero for simplicity; vintage/embodied capital structures abstracted away.
  • Empirical measurements of obsolescence involve uncertainty (decomposition of economic depreciation into physical vs. technological components, the quality of BEA service‑life estimates).

Data & Methods

• Empirical construction of an implied obsolescence rate

  • Data source: BEA asset‑level investment and official service‑life (depreciable life) tables for equipment and software (covering 1970–2023).
  • Method: decompose observed economic depreciation into physical depreciation and technological obsolescence (following capital‑embodied technical‑change literature, e.g., Cummins & Violante 2002; Sakellaris & Wilson 2004). Because physical depreciation changes slowly, measured rises in economic depreciation are interpreted as accelerated obsolescence.
  • Finding: a clear upward trend in the implied obsolescence rate from the mid‑1990s onward; the trend disappears when computing equipment and software are excluded.

• Theoretical model and quantitative implementation

  • Model: continuous‑time endogenous growth with CES aggregation for final output combining efficient labor and efficient capital; Dixit–Stiglitz aggregation across skill varieties and design varieties; monopolistic markups; free entry into innovation; HJB equations to value an innovation’s future profit stream; homogeneous labor supply allocated between production, skill innovation, and design innovation.
  • Calibration: parameterized to U.S. macro aggregates (productivity, labor share, innovation effort) and the measured obsolescence series; a permanent obsolescence shock is introduced around the mid‑1990s and transitional dynamics are examined.
  • Comparative statics and multiple sensitivity checks reported (elasticities of substitution, markups, productivity of innovation, etc.).

Implications for AI Economics

• Relevance to AI-driven technology cycles

  • AI hardware, software, and platform ecosystems exhibit rapid iterative deployment, frequent incompatibilities, and short product cycles — consistent with the paper’s notion of accelerated obsolescence in computing‑based capital.
  • If AI adoption raises effective obsolescence of capital designs, the model predicts:
    • A possible near‑term productivity surge as firms and workers adapt (skill accumulation and reallocation toward production), followed by a medium/long‑run slowdown in productivity growth.
    • A persistent decline in the labor income share as efficient capital becomes relatively scarcer and capital‑design profits rise.
    • Greater aggregate innovative effort directed toward short‑lived AI designs, producing diminishing returns to innovation (more R&D / deployment but less long‑run productivity per unit effort).

• Policy and measurement takeaways for AI economists and policymakers

  • Monitor obsolescence, not just invention counts: track depreciation/obsolescence metrics of AI‑related capital (service lives, compatibility attrition, abandonment rates) to understand macro impacts.
  • Complement investment with durable complementarities: policies that improve backward compatibility, interoperability, and standardization may lengthen effective design lives and mitigate negative long‑run effects.
  • Focus on skill accumulation and diffusion: sustaining long‑run growth requires investing in broadly transferable skills and institutions that complement rapidly changing AI designs.
  • Rethink incentives: shorter design lifespans bias private returns toward quickly monetizable design innovations; consider R&D subsidies, tax treatment, or IP rules that encourage investments with longer social returns (e.g., interoperability, open standards, durable training/education).
  • Measurement improvements: official statistics should better capture quality changes, intangible durability, and the role of obsolescence in productivity accounting when evaluating the macro effects of rapid AI-driven change.

• Research directions

  • Endogenize obsolescence: model interactions between technology (Moore‑style improvements), compatibility standards, and strategic firm behavior in determining obsolescence.
  • Heterogeneous firms and platforms: evaluate how market structure (platform dominance, lock‑in) interacts with obsolescence to affect innovation incentives and distributional outcomes.
  • Empirical micro evidence: link firm‑level AI deployment, turnover of capital/software vintages, and firm‑level productivity/wage dynamics to validate the obsolescence channel.
  • Policy experiments in models: simulate standards, subsidy, or tax interventions to quantify mitigation of the long‑run slowdown and labor‑share decline.

Limitations to keep in mind: the obsolescence shock is treated as exogenous and inferred from aggregate BEA data (subject to measurement error); the model abstracts from vintage capital structures and heterogeneous firm/market power effects that may also matter for AI‑era dynamics.