Technology overload? macroeconomic implications of accelerated obsolescence

The paper opens from the observation that technology has been moving fast enough that each innovation renders its predecessor obsolete well before the end of its useful life, yet economic growth in many advanced economies has slowed — contradicting the conventional wisdom that faster technological change should drive higher long-run productivity growth. The author notes this contradiction may be less enigmatic than it first appears, citing studies that have previously shown a link between periods of increased technological dynamism and slowdowns in aggregate productivity growth. Rapid technological progress often reveals labour-hoarding effects as a greater share of resources is initially directed to capital to develop new technologies, and the adoption of new technologies temporarily reduces capital efficiency because of the learning curve in integrating them into production. The paper’s distinctive move is to ask what happens if increased technological dynamism becomes a permanent condition rather than a transitory episode followed by a rebound in productivity growth as workers acquire new skills.

Following a productivity boom from the mid-1990s to the mid-2000s, the US non-farm business sector has experienced a notable slowdown that appears structural, accompanied by a drastic decline in the labour share and a significant decline in capital efficiency — and the author reports that these regularities are not limited to the US but extend to advanced economies in general. The reported US labour productivity growth rates by period are 1.70% for 1970–1995, 2.95% for 1996–2005 and 1.47% for 2006–2019. The labour income share is measured as the share of total value added going to labour as labour compensation plus proprietors’ income; capital efficiency for equipment and software is measured as the inverse of the incremental capital–output ratio. The author emphasises that the emergence of these patterns coincides with the mid-1990s, widely considered the true beginning of the computing revolution.

Three reasons are given. First, by vintage capital theory there is a direct correspondence between the rate and direction of technical change and obsolescence; under labour–capital complementarity, where long-run growth is labour-augmenting, a faster pace of change that outpaces skill accumulation can reduce the productive efficiency of capital despite technological improvements. Second, recent research has documented a significant and broad-based increase in the economic depreciation rate of capital in the US, and the literature suggests that increased value losses in existing capital are more likely driven by faster obsolescence than by physical depreciation. Third, the digital era has distinctive features: adoption of information processing equipment and software has expanded significantly alongside integration of computing-based designs into traditional capital assets, and these technologies are uniquely prone to fast obsolescence because of exponential advances in processing power and limited compatibility across successive design generations — a characteristic the author says sets computing apart from earlier technology types.

The author decomposes BEA economic depreciation into physical and obsolescence components, constructing an implied physical depreciation rate using a Federal Reserve Board approach that derives depreciation schedules explicitly excluding obsolescence, and then backing out obsolescence as the residual time-related value loss. The BEA calculates economic depreciation by the Hulten–Wykoff declining-balance method based on resale prices, so its aggregate rate reflects both age effects (physical decay) and time effects (technological obsolescence), which cannot be separately identified without an independent estimate. Using BEA investment data and the FRB perpetual inventory method, the author constructs counterfactual capital stock series adjusted for age effects only, for 42 non-residential private equipment and software assets, relying on the Mohr–Gilbert hyperbolic decay function with curvature parameter b = 0.75 and time-invariant service lives from the BEA Handbook for Fixed Assets. Investment data are deflated using a constant-quality consumer price index — deliberately, because the aim is to isolate physical decay rather than obsolescence due to technological change — and a mid-year convention is applied per BEA practice. An aggregate obsolescence index is then built recursively and the implied obsolescence rate defined as its log change.

Both implied obsolescence series are relatively stable through the 1970s–1990s and then trend upward, with the constant-dollar measure rising from an average of 4.2% over 1970–1995 to 6.6% by the late 2010s, and the current-dollar measure from 5.6% to 6.4%. In period averages for equipment and software, constant-dollar economic depreciation d rises from 11.8% (1970–1995) to 16.2% (2015–2023) and the current-dollar measure d$ from 13.1% to 16.0%, while the physical rate rises from 8.1% to 10.5%. Before 1995 both economic and physical rates trend upward but the gap between them is relatively stable — about 3.7 percentage points on the constant-dollar measure and 5.0 points on the current-dollar measure — whereas after the mid-1990s the physical and economic rates begin to diverge. The author reports both constant- and current-dollar rates because each has limitations: the constant-dollar rate is the conventional choice and appropriate for the calibration here but is subject to base-year effects in aggregation that may bias trends at the endpoints of long series, while the current-dollar rate avoids base-year distortions but reflects expenditures rather than actual productive capacity. The stated key takeaway is the common shift in timing and direction, pointing to a rise in capital obsolescence regardless of the measurement approach.

The paper uses the endogenous growth framework of Acemoglu (2003) with directed technical change, adding three tailored assumptions: work efforts allocated between production and creative activities are of uniform quality; efficient labour and efficient capital are complementary factors in final output; and technology designs are capital-embodied and replaced at a constant rate defined as the obsolescence rate. Final output is a CES combination of efficient labour and efficient capital, each a Dixit–Stiglitz aggregation of intermediate unit services; the representative household has CRRA preferences and a constant endowment of work effort allocated between goods production and factor augmentation, and physical depreciation of capital is assumed to be nil for simplicity. The assumption that each new arrival immediately renders the previous one economically outdated gives a one-to-one mapping that keeps capital efficiency constant along a balanced growth path; given a steady-state level of capital efficiency, labour continuously acquires new skills to complement the latest designs, driving overall productivity growth. In this economy, productivity gains from technological advances eventually accrue to labour as it develops the skills to use new technologies; “skills” collectively refer to knowledge, training and experience.

Five parameters are calibrated to match five long-run US conditions observed before the mid-1990s: productivity growth of 1.70%, a labour income share of 62.6%, a rate of interest of 7.7%, a 75% share of work effort in production, and a 12.5% share in creative activities. The calibrated parameters are the substitutability between labour inputs (α = 0.7119) and between capital inputs (β = 0.5631), the transformation parameters for creative effort (b_h = 0.3360, b_z = 0.4000) and the rate of time preference (ρ = 0.0430). Three parameters are set externally: the CES parameter ψ = −1, corresponding to an elasticity of substitution of 0.5, which the author sets on the basis of own estimates for the non-farm business sector falling between 0.3 and 0.8 across specifications and estimators — implying gross complementarity — and consistent with previous evidence typically in the 0.5–0.7 range; the elasticity of intertemporal substitution parameter θ = 2, from Hall (1988); and the initial obsolescence rate δ_z = 5%. The interest rate target is matched to the pre-tax real return on capital in the non-farm business sector excluding financials, calculated as total capital income divided by capital stock at replacement cost. The work-effort targets have no direct real-world counterpart, so the author uses the Acemoglu–Autor (2011) occupational categorisation — treating non-routine cognitive occupations as creative activity — and assumes an equal split of creative effort between capital and labour augmentation initially.

The economy, initially on a balanced growth path, is hit by a rise in the obsolescence rate from 5% to 7.5%, dated right after 1995 in line with the empirical evidence and the widespread adoption of computing technologies. The initial 5% is set agnostically as an average of the two constructed series over the pre-1995 period; 7.5% is the post-1995 target where both measures approximately rise. The author notes the initial value is largely consistent with previous estimates derived using alternative methodologies — 4% in Hornstein and Krusell (1996), 4.8% for equipment and software in Cummins and Violante (2002), 2.5% for equipment and structures in Hobijn (2000), and 8% in Sakellaris and Wilson (2004) applying FRB-PIMS to firm-level manufacturing data. Although the obsolescence rate may have increased gradually, it is treated as a one-time shift for tractability, which the author describes as providing a clear approximation of how the economy adjusts to a new balanced growth path. Only the obsolescence rate is targeted; the transitional behaviour of all other model variables is left untargeted.

On impact, labour productivity growth jumps up momentarily as skill accumulation temporarily accelerates to keep pace with the rapid introduction of new capital, but increased obsolescence simultaneously starts to drive down the efficiency of capital because embodied technologies are now short-lived, and the interest rate begins to fall. Accelerated replacement reduces efficient capital per efficient labour, which raises demand for efficient capital relative to efficient labour and so raises the relative compensation per efficiency unit of capital. That price effect increases the profitability of creating new technology designs for capital, redirecting resources away from developing new labour skills and decelerating skill accumulation over time, so the productivity boom gradually fades out within a decade. Because demand for efficient capital remains higher in the new balanced growth path, investment incentives for capital technologies stay strong and rapid replacement is reinforced — the higher obsolescence rate, in the author’s phrase, now feeds on itself. As more resources are allocated to capital, the share of income going to labour begins to decline, and under gross complementarity the capital-biased reallocation causes the labour share to settle at a lower level while the economy converges to a slower long-run productivity growth path.

In the new balanced growth path the model predicts equilibrium growth falling from 1.70% to 1.51% and the labour income share falling from 62.6% to 56.4%, against observed data of 1.47% and 56.9% — a model-implied 0.19 percentage-point reduction in productivity growth and 6.2 percentage-point decline in the labour share. The comparison periods for the untargeted moments are 2006–2019 for productivity and 2010–2023 for labour share, evaluated after a model transition period of about ten years defined as the half-life of convergence to the new balanced growth path. The author states these model-derived outcomes closely follow the observed declines.

Despite the transitional burst in labour productivity, the model does not predict an increase in the labour share during that period, and the author notes the data do not show a significant upward change over the same period either. The proposed mechanism is a difference in the transition dynamics of wage growth and productivity growth: wage growth in the model is driven by the accumulation of labour skills and by changes in the efficient capital-to-efficient labour ratio, and while that ratio is constant along a balanced growth path, the rise in obsolescence reduces it outside one, causing wage growth to lag productivity growth and redirecting financial benefits to capital rather than labour during the transition. This transitional divergence creates a wedge between wage and productivity levels in the long run: US data show a decoupling of payment per hour from labour productivity with a gap of about 20 log points by 2019, most of which developed between the mid-1990s and mid-2000s before stabilising, and the model predicts a gap of similar magnitude to that observed after the mid-1990s. The author also notes that one reason for the pattern, at least in the model, is that the mid-1990s boom is transitory.

Faster obsolescence results in the economy holding a larger stock of capital in a raw sense, but a stock that contains fewer efficiency units in the long run, because embodied designs become obsolete quickly before their productive gains are fully realised — so efficient capital used in production becomes scarcer relative to the initial balanced growth path. In the data the author measures capital efficiency as the inverse incremental capital–output ratio using real equipment and software investment and real GDP, arguing this metric is attractive because it relies on widely available series and directly measures how much additional capital is required to produce one more unit of output. Empirically the inverse ICOR falls by about 0.68 log points relative to its pre-mid-1990s average; the model’s transition reproduces this downward shift, albeit with a slightly smaller magnitude. Similar declines appear when the metric is computed using equipment investment only and in BEA–BLS capital productivity series.

The qualitative outcomes remain robust across four sensitivity scenarios, but two quantitative sensitivities are flagged: the magnitude of the model-implied productivity slowdown is sensitive to the initial allocation of work effort, and the extent of the labour-share decline is shaped by the starting obsolescence rate. In Scenario I, varying the initial share of work effort in production over {0.70, 0.75, 0.80}, a high production share makes long-run productivity growth less responsive to an increase in obsolescence and a lower production share makes the economy more responsive — which the author reads as implying that economies starting from a more innovation-oriented baseline are more sensitive to obsolescence shocks — while the labour-share decline is quantitatively unaffected. In Scenario II a lower risk aversion of θ = 1.5 makes equilibrium productivity growth slightly more responsive but has a negligible effect on the labour share. In Scenario III an alternative elasticity of 0.6 (ψ = −0.67) leaves both productivity growth and labour share insensitive — expected, since along a balanced growth path both are independent of ψ given a steady-state efficient capital-to-labour ratio — though a higher degree of complementarity does lead to faster convergence. In Scenario IV a lower initial obsolescence rate of 2.5%, which the author notes is likely more appropriate for the full capital stock including long-lived structures, leaves the productivity slowdown unresponsive but implies a smaller decline in the labour share than the baseline.

The author assembles a panel of 16 advanced economies — large Western European economies, the Western offshoots (Canada, Australia, New Zealand, the US) and the larger East Asian economies (Japan, South Korea) — over the 1980s to 2019, and reports that labour productivity growth boomed from the mid-1990s to about 2005 then slowed sharply, while the labour income share declined gradually until the late 1990s and considerably more steeply thereafter. All country-level measures are GDP-weighted averages with country fixed effects removed to eliminate mechanical variation from country entry or exit. Productivity data come from the panel of Fernald, Inklaar and Ruzic (2025), reaching back to 1985; labour income shares are constructed from OECD Annual National Accounts as compensation of employees plus two-thirds of mixed income divided by value added at factor cost, restricted to the business sector and excluding real estate, with a uniform mixed-income share of 10% of value added assumed across countries. Capital efficiency is again the inverse ICOR, using Penn World Table 10.01 investment with price indexes anchored to BEA deflators as a quality-adjustment benchmark; it shows a pronounced and persistent decline that began in the 1990s, intensified in the early 2000s and stabilised at a lower level after the mid-2000s. Capital productivity, measured as real GDP over real capital services, had fallen by approximately 0.4 log points by 2019 relative to the 1985–1995 average. The author notes that measuring obsolescence rates across countries is hardly possible because international depreciation datasets are constructed differently from the BEA’s concept, which is why the cross-country section examines capital efficiency instead.

The author states that the paper’s qualitative outputs neither contradict nor dismiss prevailing explanations of slowing growth and the declining labour share, and frames the obsolescence channel as a contribution to the debate rather than a replacement. On the productivity side, the paper notes that despite increasing innovation rates and R&D intensity the slowdown appears to be a genuine change rather than a measurement artifact, and that existing accounts attribute it to declining knowledge diffusion or to diminishing returns to inventive activity; the paper’s addition is that the increasing pace of technological change may itself be a contributing factor, since newer technologies rapidly displacing recently developed ones reduce the returns to innovation and hinder knowledge diffusion. On the labour-share side, it positions the mechanism alongside explanations based on task automation and rising capital intensity, acknowledging automation as a possible consequence of rising obsolescence rates given labour’s limited adaptability to frequent technology changes rather than treating it as a rival explanation. The author also notes that the two literatures — on growth deceleration and on the labour share — have largely been examined separately, and presents the paper as offering a unified framework connecting productivity, innovation and income distribution.

The paper’s central claim is framed throughout as identifying one possible driver and a potential contribution, not a demonstrated sole cause. The empirical obsolescence series is an implied rate, inferred by decomposition rather than observed, and depends on the FRB-based physical depreciation construction, the choice of a constant-quality deflator, and BEA service-life assumptions; the author acknowledges that early observations in the disaggregated series are unavoidably noisy and applies mild smoothing at the asset level. The measured decomposition covers equipment and software only, which is why the alternative lower initial obsolescence rate is explored for the full capital stock including structures. The shock is modelled as a one-time shift for tractability rather than the gradual increase the data may reflect. And the cross-country evidence is on capital efficiency and productivity rather than on obsolescence itself, which the author states cannot be measured comparably across countries.