Academic research: technology adoption S-curves

Generated by /academic-research on 2026-06-16. Synthesized across 3 rounds from 17 peer-reviewed papers (see Provenance). Treat as raw material — review before promoting into a project or thread. Context: vault/threads/technology-adoption-s-curves Note: this topic mixes academic and practitioner knowledge. A complementary /autoresearch pass (Tony Seba / RethinkX framing + recent industry case studies) is recommended; the two clippings can be promoted side by side. One key paper here (Gogerty 2026) is a 0-citation SSRN preprint — its specific learning-rate figures should be treated as provisional.

Summary

The S-shaped (logistic) adoption curve is one of the most robust stylized facts in the study of technological change — the dominant pattern across dozens of technologies (Geroski 2000) — but it is generated by several distinct mechanisms (information epidemics, firm heterogeneity, density dependence), which matters because they imply different drivers of steepness. Two complementary quantitative traditions describe it: the Bass diffusion model (adoption driven by coefficients of innovation p and imitation q) on the demand side, and Wright's Law / experience curves (unit cost falls a roughly constant % per doubling of cumulative production) on the cost side. The headline finding for an investor: technological progress is genuinely forecastable — Wright's Law produced the best hindcasts among six candidate laws across 62 technologies, with forecast error growing predictably (~2.5%/yr in the log) (Nagy et al. 2012), and the error distribution is regular enough to assign a probability that one technology will beat another by a future date (Farmer & Lafond 2015). But forecastability has sharp limits: the ceiling/market-potential M and the Bass p,q are notoriously unstable and sensitive (Massiani et al. 2015), learning rates themselves change over time so "past learning ≠ future learning" (Carlino et al. 2025), and real curves are often asymmetric around the inflection (Easingwood et al. 1981). The error has a systematic direction: authoritative forecasters have repeatedly and severely underestimated the long-run deployment of cost-declining technologies like solar PV and wind (Lopez et al. 2025; Way et al. 2022) — while near-term expectations are often simultaneously over-hyped (the hype-cycle pattern). For equities, adoption-lag dynamics generate predictable cycles in valuations and risk premia as "growth options" convert into "assets in place" (Gârleanu et al. 2012), and abnormal returns vary by the technology's life-cycle stage (Lee 2018).

Findings

1. The S-curve is the dominant pattern — but it has several distinct generating mechanisms

The literature on diffusion is vast and crosses many disciplines, but it organizes around one stylized fact: usage of a new technology over time typically traces an S-curve (Geroski 2000). Geroski's survey is important because it shows the same S-curve can arise from at least three different mechanisms, each with different implications:

• Epidemic model — adoption is limited by the spread of information about the technology (how to use it, what it does); diffusion is slow until knowledge propagates. This is the mechanism most analogous to the Bass "imitation" term.
• Probit / firm-heterogeneity model — different firms have different goals and capabilities, so they cross the adoption threshold at different times; diffusion is the cumulative distribution of those thresholds.
• Density-dependence model (from population ecology) — legitimation and competition jointly establish then ultimately limit take-up.

Rogers' Diffusion of Innovations supplies the canonical vocabulary that sits on top of these: the adopter categories (innovators → early adopters → early majority → late majority → laggards), the five-stage individual decision process (knowledge, persuasion, decision, implementation, confirmation), and the four diffusion elements (innovation, communication channels, time, social system) (Guo et al. 2024). The practical takeaway: "the curve is an S" is not a single causal claim — which mechanism dominates a given case (information friction vs. firm heterogeneity vs. cost) determines what would accelerate or stall it.

2. Quantifying the curve: the Bass model and logistic/Gompertz fits — and why the parameters are slippery

The Bass diffusion model decomposes adoption into two forces: innovators who adopt independently (coefficient p) and imitators who adopt as a function of how many have already adopted (coefficient q), scaled by a market potential M. Maximum-likelihood estimation of these parameters outperforms OLS on goodness-of-fit and one-step-ahead forecasts, and yields standard errors and sample-size guidance (Schmittlein & Mahajan 1982).

The crucial caveat — directly relevant to "are we capturing enough growth" — is that the parameters are hard to pin down. In an empirical investigation of new automotive technologies (focused on EVs), literature values for p and q vary dramatically and are "discussible," while ad-hoc estimates from market data are "poorly conclusive"; worse, the estimated p is highly sensitive to the assumed market potential M, so for plausible ranges of M, p swings wildly. The authors issue an explicit warning to users of these forecasts and to anyone basing decisions on them (Massiani et al. 2015). This is the quantitative root of the "where's the ceiling?" problem.

Two refinements matter for fitting real data:

• Real diffusion is frequently asymmetric — the curve need not be symmetric around its inflection point. The Non-Symmetric Responding Logistic (NSRL) model lets the imitation effect vary over time so the inflection responds to the substitution process (Easingwood et al. 1981). Assuming a symmetric curve can therefore misplace "how much is left."
• A simple logistic does not always fit the S well; piecewise / fuzzy-logistic models that allow the curve to bend can outperform both the technology-substitution model and the Norton–Bass model, especially with short or noisy data series (Tseng et al. 2014). The classic substitution framing (one technology replacing another along a logistic) dates to Stern et al. 1975, which also flagged early that such forecasts are vulnerable to political and price shocks (their plastic-for-glass forecast was upset by the oil-price spike).

3. Cost drives the knee: Wright's Law and experience curves

If the Bass model describes adoption, Wright's Law describes the cost decline that often triggers and sustains it. Wright's Law (1936, from aircraft manufacturing) holds that unit cost falls as a power law of cumulative production — a constant % reduction per doubling, the "learning rate." Across an expanded dataset of 150 technologies (97 with Wright's-Law fits), the median learning rate is ~21% per doubling, consistent with earlier findings; GPU compute shows the highest learning rate ever measured (reported as 89.1% cost reduction per doubling, R²≈0.995), while nuclear power exhibits persistent anti-learning (costs rising with cumulative build) (Gogerty 2026 — 0-citation SSRN preprint; treat specific figures as provisional). The nuclear "anti-learning" result independently echoes the energy-transition literature below.

The constant-learning-rate assumption is itself imperfect: analysing 87 technologies in the Performance Curve Database, stepwise (changing) learning rates fit better for 58 of them, and — critically — "the observed learning rate is not a good predictor of future learning"; learning rates change for ~66% of technologies (Carlino et al. 2025). Their recommendation has a direct investing analogue: favour early-stage technologies nearing cost-competitiveness, combined with decision-making-under-uncertainty methods, rather than betting on a single extrapolated rate.

4. How forecastable is adoption — and what does the error look like?

This is the most decision-relevant body of evidence. Using a database of 62 technologies, Nagy et al. 2012 rigorously hindcast six proposed "laws" of technological improvement:

• Wright's Law (cost ∝ power law of cumulative production) produced the best forecasts, with Moore's Law (exponential improvement in time) close behind.
• Because production itself tends to grow exponentially, Wright's and Moore's laws become nearly indistinguishable in practice (Sahal's point).
• Most importantly: technological progress is forecastable, with the (root-mean-square) logarithmic error growing roughly linearly with horizon at ~2.5% per year. That gives a quantitative handle on how much to trust a multi-year extrapolation.

Farmer & Lafond 2015 formalize this as a correlated geometric random walk with drift across 53 technologies, deriving a closed-form forecast-error distribution that collapses many technologies onto one universal distribution. The practical payoff is explicitly investment-shaped: their method can estimate the probability that a given technology will outperform another technology at a future date — i.e. it turns "which cost curve wins, and when" into a probability, not a point guess. (Popular write-up: Gwynne 2013, which also catalogues how badly expert point-predictions have historically erred.)

5. The systematic direction of error — long-run underestimation of cost-declining tech, near-term over-hype

Two error patterns coexist and must be held simultaneously:

(a) Long-run underestimation of disruptive, cost-declining technologies. An ex-post analysis of the IEA's World Energy Outlook (1993–2022) finds it has "severely underestimated the growth in key renewable technologies including solar PV and wind" across scenarios, while overestimating fossil and nuclear generation; it imposed an effectively "artificial cap" on renewables growth (Lopez et al. 2025). Examining 26 global energy scenarios, Carrington et al. 2018 find all of them failed to account for plausible upper levels of solar PV growth, and argue these conservative projections become self-fulfilling by deterring investment. Building on the Nagy/Farmer probabilistic methods, Way et al. 2022 (Joule, 376 citations) note "most energy-economy models have historically underestimated deployment rates for renewable energy technologies and overestimated their costs," and conclude that simply extrapolating current exponential deployment trends implies a near-net-zero system within ~25 years at multi-trillion-dollar net savings. This is the rigorous, peer-reviewed core of the Tony Seba thesis: the consensus forecasting apparatus is structurally biased against technologies on steep cost curves.

(b) Near-term over-hype. The mirror-image error is the Gartner "hype cycle" — a peak of inflated expectations followed by a trough of disillusionment before a productivity plateau. The hype cycle is influential in real corporate technology-investment decisions, but its empirical validity is weak: testing it against news-article counts and search interest for energy technologies, Steinert & Leifer 2010 "strongly criticize" the approach for lacking an operational mathematical model and conflating two different underlying theories (expectation-hype and the technology S-curve). The investing-relevant synthesis of (a) and (b) is essentially Amara's Law — we overestimate a technology's effect in the short run and underestimate it in the long run — which maps cleanly onto being too early (paying for hype near the peak of expectations) versus being well-positioned for the structurally underestimated long-run S-curve.

6. Disruption framing (Christensen) — influential, but contested as a predictive tool

Christensen's disruptive-innovation theory is one of the most influential business theories of the century, but its scientific status is genuinely contested, and the contest centres on exactly the property an investor would want: prediction. Critiques (catalysed by Jill Lepore's New Yorker essay) identify three root problems — an inadequately constrained definition of "disruptive innovation," an inconsistent unit of analysis, and a failure to account for managerial agency (Weeks 2015). A 2024 synthesis catalogues the live controversies — definitional ambiguity, the case-study/generalizability problem, outcome bias, and "the exploration of its predictive and prescriptive potential" — concluding the theory's value may be more performative (shaping how managers act) than predictive (Lile et al. 2024). Where it has been tested closely (nine cases across six Indian industries), the core concepts find "broad support" but break down in developing-economy contexts, prompting calls for a more generalized definition (Madhusudan et al. 2022). Implication for the thread: use Christensen as a qualitative lens for direction (low-end / new-market footholds that improve up-market), not as a quantitative forecast — the quantitative work lives in the Bass/Wright/logistic literature above.

7. What makes a curve steep: network effects, increasing returns, and lock-in

"Some S-curves are steeper than others" has a literature. Steepness and the eventual ceiling are governed substantially by increasing-returns / network effects (Arthur's framework): when adoption raises the value or lowers the cost of further adoption, you get positive feedback, tipping, and lock-in. A simulation extension shows there is "no lock-in without further stabilizing returns" — i.e. the steep self-reinforcing phase requires the returns mechanism to persist (Roedenbeck & Schulz 2007). Under network effects, vendor strategy (de facto vs. de jure standards, first-mover proprietary choices) shapes who wins and how the market splits, and early adopters get locked in by switching costs (Lee & Mendelson 2007). A comparative case study of energy/road transport enumerates the concrete lock-in mechanisms that set curve steepness and the height of incumbent resistance: learning effects, economies of scale and scope, network externalities, informational increasing returns, technological interrelatedness, and institutional effects (Klitkou et al. 2015). For evaluation: the presence/strength of these mechanisms is the leading indicator of a steep curve and a defensible eventual share.

8. The investing translation — adoption stage → valuations, risk premia, and returns

The bridge from adoption curves to equity outcomes has direct peer-reviewed support:

• Adoption lags create valuation and risk-premium cycles. In an asset-pricing model with general-purpose technology, large innovations are embodied in new capital only with a lag (firms must adopt via investment). This adoption process generates cycles in asset valuations and risk premia as firms convert "growth options" into "assets in place," and helps explain valuation patterns around major innovations, the lead-lag between the stock market and output, and rising consumption-return correlations at long horizons (Gârleanu et al. 2012). The "growth option → asset in place" conversion is the financial expression of moving up the S-curve — and it implies the risk premium (hence expected return) is highest before the conversion, i.e. earlier on the curve.
• Abnormal returns vary by life-cycle stage. An event study of augmented-reality firms on Korea's KOSDAQ finds portfolios earned returns above benchmark during the "Peak of Inflated Expectations" stage, but showed no consistent abnormal-return pattern during the "Trough of Disillusionment" (Lee 2018). This is a caution as much as a confirmation: the easy excess return clustered in the hype phase, not uniformly across the curve.
• Portfolio construction changes when assets follow experience curves. When the "assets" are competing technologies obeying Wright's Law, the cost–investment positive feedback distorts standard mean-variance portfolio theory: it creates multiple local optima and a genuine tension between concentrating investment to drive one technology rapidly down its cost curve vs. diversifying to hedge failure; the discount rate plays a decisive role (Way et al. 2017). This is a formal warning against naively diversifying across technologies on different curves.
• General IT investment correlates positively with firm profitability and stock returns, mediated by process innovation and supply-chain optimization, though effect sizes on ROA/ROE were not all strongly significant in one cross-industry study (Hadi et al. 2023) — supportive but weaker evidence, included for completeness.

Contradictions and open questions

• Forecastable vs. unstable. Wright's-Law cost curves are impressively forecastable in aggregate (Nagy et al. 2012; Farmer & Lafond 2015), yet the learning rate is unstable per-technology (Carlino et al. 2025) and the Bass p,q,M are badly identified (Massiani et al. 2015). Reconciliation: cost trajectories are more forecastable than adoption timing and ceiling. An evaluation lens should lean on the cost curve for direction and treat the ceiling/timing as a distribution, not a point.
• Where exactly is the inflection, in real time? Curves are asymmetric (Easingwood et al. 1981) and often only well-estimated after the fact. Detecting "we are at the knee, growth is accelerating" prospectively — the single most valuable signal for capturing remaining runway — remains hard and is the central open question for the investing application.
• Underestimation vs. over-hype direction. The energy-transition evidence shows authoritative long-run underestimation (Lopez et al. 2025; Way et al. 2022) while the hype-cycle tradition warns of near-term over-expectation. Both can be true (Amara's Law), but knowing which regime a given name is in at a given moment is unresolved.
• Does disruption theory predict, or only describe? Open and contested (Weeks 2015; Lile et al. 2024).
• Selection/optimism bias in the energy literature. Several of the strongest "underestimation" papers (Carrington et al. 2018; Way et al. 2022) are themselves written by transition advocates; their forecasts also assume current exponential trends continue, which is the very assumption a saturating S-curve eventually breaks. Treat the "trillions in savings / near-net-zero in 25 years" as a directional argument, not a calibrated base case.

Priors check

No priors were captured for this run (research was initiated autonomously to seed the technology-adoption-s-curves thread). If you want to log how this evidence updated any prior beliefs — e.g. about how forecastable adoption is, or whether the "underestimation" bias is exploitable — use /calibrate.

Provenance

Rounds run: 3 (full)

Sub-questions by round:

Round 1 (broad survey):

1. Diffusion of innovations and the logistic S-curve (Rogers; mechanisms behind the S-curve)
2. The Bass diffusion model — coefficients of innovation/imitation, parameter estimation
3. Wright's Law / experience curves — cost decline per cumulative doubling, learning rates
4. Statistical forecastability of technological progress — Wright vs. Moore, error structure

Round 2 (drill-down):

1. Christensen disruptive-innovation theory — empirical validity and predictive accuracy — targeting the "disruption" leg of the toolkit
2. Systematic forecast bias in energy tech (IEA/solar PV) — targeting the direction of error
3. Logistic-substitution / market-potential & inflection estimation — targeting the "where's the ceiling, how steep" problem

Round 3 (resolve remaining uncertainty):

1. Technology life cycle and equity returns / asset pricing — targeting the investing translation
2. Gartner hype cycle / Amara's Law — targeting near-term-overestimate vs. long-run-underestimate
3. Network effects, increasing returns, lock-in — targeting the determinants of curve steepness

Papers reviewed (17 total; R1: 6, R2: 6, R3: 5):

Round 1:

• Models of technology diffusion — Geroski (2000). Research Policy — 1554 citations, literature review. Canonical survey: the S-curve arises from epidemic, probit, and density-dependence mechanisms.
• Analyzing the Diffusion of Innovations Theory — Guo et al. (2024). Scientific and Social Research — 13 citations. Rogers' S-curve, adopter categories, 5-stage decision process.
• The choice of Bass model coefficients to forecast diffusion for innovative products (EVs) — Massiani & Gohs (2015). Research in Transportation Economics — 151 citations. p,q highly variable and sensitive to assumed market potential M; warning to forecast users.
• Maximum Likelihood Estimation for an Innovation Diffusion Model — Schmittlein & Mahajan (1982). Marketing Science — 302 citations. MLE for Bass; standard errors and sample-size for forecast accuracy.
• Statistical Basis for Predicting Technological Progress — Nagy, Farmer, Bui, Trancik (2012). PLoS ONE — 281 citations. Wright's Law best of six laws; log error grows ~2.5%/yr; progress is forecastable.
• Wright's Law Extended: Experience Curves Across 150 Technologies — Gogerty (2026). SSRN preprint — 0 citations. Median learning rate ~21%; GPU compute highest measured; nuclear anti-learning. Provisional — preprint.

Round 2:

• How predictable is technological progress? — Farmer & Lafond (2015). Research Policy — 247 citations. Closed-form forecast-error distribution; probability one technology beats another by a date.
• Is disruption theory wearing new clothes or just naked? — Weeks (2015). Innovation — 69 citations. Christensen critique: definition, unit of analysis, managerial agency.
• Rethinking disruptive innovation: theoretical controversies — Lile et al. (2024). Innovation — 12 citations. DIT's predictive-vs-performative status; outcome bias.
• An ex-post analysis of the IEA's World Energy Outlook — Lopez et al. (2025). Renewable and Sustainable Energy Reviews — 48 citations. IEA severely underestimated solar PV/wind, overestimated fossil/nuclear.
• The politics of energy scenarios (IEA conservative solar PV) — Carrington & Stephenson (2018). Energy Research & Social Science — 72 citations. All 26 scenarios under-projected solar PV; conservative forecasts deter investment.
• A nonsymmetric responding logistic model for forecasting technological substitution — Easingwood, Mahajan, Muller (1981). Technological Forecasting and Social Change — 141 citations. Real S-curves are often asymmetric around the inflection.

Round 3:

• Empirically grounded technology forecasts and the energy transition — Way, Ives, Mealy, Farmer (2022). Joule — 376 citations. Models historically underestimate renewable deployment / overestimate cost; probabilistic forecasts imply large savings from a fast transition.
• Technological Growth and Asset Pricing — Gârleanu, Kogan, Panageas (2012). Journal of Finance — 172 citations. Adoption lags → valuation/risk-premium cycles; growth options → assets in place.
• The Relationship between Technology Life Cycle and Korean Stock Market Performance — Lee (2018). International Journal of Financial Studies — 2 citations. Abnormal returns concentrated in the "Peak of Inflated Expectations" stage.
• Wright meets Markowitz: portfolio theory with experience curves — Way, Lafond, Farmer et al. (2017). Journal of Economic Dynamics and Control — 43 citations. Cost–investment feedback distorts mean-variance optimization (concentrate vs. diversify).
• The role of lock-in mechanisms in transition processes — Klitkou et al. (2015). Environmental Innovation and Societal Transitions — 308 citations. Enumerates lock-in mechanisms (learning, scale, scope, network externalities) that set curve steepness.

Also surfaced (cited inline, not in the retained count): Stern, Ayres & Shapanka 1975 (logistic substitution; shock vulnerability); Tseng et al. 2014 (piecewise logistic); Madhusudan et al. 2022 (Christensen in developing economies); Steinert & Leifer 2010 (hype-cycle critique); Roedenbeck & Schulz 2007 and Lee & Mendelson 2007 (network effects/lock-in); Hadi et al. 2023 (IT investment & returns); Gwynne 2013 (popular write-up of Nagy et al.).

Tools used: mcp__consensus__search (Consensus — Semantic Scholar, PubMed, Scopus, ArXiv). Filters applied: none. Generated: 2026-06-16 13:30 UTC